UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Characterizing single neuron activity patterns and dynamics using multi-scale spontaneous neuronal activity… Mitelut, Catalin C. 2017

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


24-ubc_2018_february_mitelut_catalin.pdf [ 33.77MB ]
JSON: 24-1.0357482.json
JSON-LD: 24-1.0357482-ld.json
RDF/XML (Pretty): 24-1.0357482-rdf.xml
RDF/JSON: 24-1.0357482-rdf.json
Turtle: 24-1.0357482-turtle.txt
N-Triples: 24-1.0357482-rdf-ntriples.txt
Original Record: 24-1.0357482-source.json
Full Text

Full Text

Characterizing Single Neuron Activity Patterns and DynamicsUsing Multi-Scale Spontaneous Neuronal Activity Recordings ofCat and Mouse CortexbyCatalin C. MitelutBSc-Dual Honours, Physics and Philosophy, University of Windsor, 1999MSc, Physics, University of Windsor, 2001JD, The University of British Columbia, 2004MA, Philosophy, University of Windsor, 2012a thesis submitted in partial fulfillmentof the requirements for the degree ofDoctor of Philosophyinthe faculty of graduate and postdoctoral studies(Neuroscience)The University of British Columbia(Vancouver)November 2017c© Catalin C. Mitelut, 2017AbstractThroughout most of the 20th century the brain has been studied as a reflexive system with everimproving recording methods being applied within a variety of sensory and behavioural paradigms.Yet the brains of most animals (and all mammals) are spontaneously active with incoming sensorystimuli modulating rather than driving neural activity.The aim of this thesis is to characterize spontaneous neural activity across multiple temporaland spatial scales relying on biophysical simulations, experiments and analysis of recordings fromthe visual cortex of cats and dorsal cortex and thalamus of mouse.Biophysically detailed simulations yielded novel datasets for testing spike sorting algorithmswhich are critical for isolating single neuron activity. Sorting algorithms tested provided low errorrates with operator skill being as important as sorting suite. Simulated datasets have similarcharacteristics to in vivo acquired data and ongoing larger-scope efforts are proposed for developingthe next generation of spike sorting algorithms and extracellular probes.Single neuron spontaneous activity was correlated to dorsal cortex neural activity in mice.Spike-triggered-maps revealed that spontaneously firing cortical neurons were co-activated withhomotopic and mono-synaptically connected cortical areas, whereas thalamic neurons co-activatedwith more diversely connected areas. Both bursting and tonic firing modes yielded similar mapsand the time courses of spike-triggered-maps revealed distinct patterns suggesting such dynamicsmay constitute intrinsic single neuron properties. The mapping technique extends previous work tofurther link spontaneous neural activity across temporal and spatial scales and suggests additionalavenues of investigation.Synchronized state cat visual and mouse sensory cortex electrophysiological recordings revealedthat spontaneously occurring activity UP-state transitions fall into stereotyped classes of events thatcan be grouped. Single visual cortex neurons active during UP-state transitions fire in a partiallypreserved order extending previous findings on high firing rate neurons in rat somatosensory andauditory cortex. The firing order for many neurons changes over periods longer than 30-minutessuggesting a complex non-stationary temporal neural code may underly spontaneous and stimulusevoked neural activity.This thesis shows that ongoing spontaneous brain activity contains substantial structure thatcan be used to further our understanding of brain function.iiLay SummaryThe brains of all animals including mammals are spontaneously active yet much of neuroscienceresearch has focused on studying the brain’s response to specific stimuli such as specific odors,pictures or sounds.The work presented in this thesis is aimed at characterizing brain activity that occurs spon-taneously. The focus is on recording the spontaneous activity of single neurons and relating it toother nearby neurons and to other more distant areas of the brain.It is shown that existing single neuron detection methods are adequate for analysis. Singleneurons fire simultaneously with other neurons to which they are anatomically directly connectedwith neurons from deeper areas of the brain firing in more unique patterns. Spontaneously activesingle neurons in visual areas of the brain are also shown to fire in specific orders that appear tochange over time.iiiPrefaceThe author’s PhD researched has been published or is in review in 6 journals and has beenpresented in 9 conference posters.The author’s work relating to biophysical modeling described in Chapter 3 has been published inone journal Hawrylycz et al., 2016, one pre-print journal (Jun et al., 2017a - currently in preparationfor peer-reviewed journal submission), one article in review (Jun et al., 2017b) and was presentedat multiple conferences (Mitelut et al., 2014, 2015; Gratiy et al., 2015, 2016; Vyas et al., 2016). Anadditional publication on the Allen Institute network model is currently in preparation (Gratiy etal., 2017).The optical imaging work described in Chapter 4 has been published (Xiao et al., 2017) andwas presented at two conferences (Mitelut et al., 2016; Xiao et al., 2016).The work described in Chapter 5 has been presented at one conference (Mitelut et al., 2015)and is being prepared for publication.Additional publications resulting out the author’s PhD research not discussed in this thesisinclude: a python-based toolbox used for analysis of widefield calcium activity recordings (Hauptet al., 2017); a publication on the use of SpikeSorter (Swindale et al., 2017); and a fourth researchproject on the neural correlates of spontaneous mouse behaviour initiation which was presented attwo conferences (Mitelut et al., 2017a,b) and is being prepared for publication.Experiments on cats were carried out by Nicholas V. Swindale and Martin Spacek (Swindaleand Spacek, 2014). Experimental work in cats was covered by UBC Ethics Certificates A04-0098and A11-0280. Experiments on mice were carried out by the author. Experimental work in micewas covered by UBC animal application certificates A14-0266 and A13-0336.The author declares no conflicts of interest.ivTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiNote to Reader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xivAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvIntroduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2Thesis summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6Experimental and Analytical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 9Electrophysiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9Optical imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14Analysis and computational methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17Spike Sorting Algorithm Testing Using In Vitro and In Silico Datasets . . . . . . 27Electrophysiology and single unit spike sorting . . . . . . . . . . . . . . . . . . . . . . . . 28Using a novel in vitro extracellular-intracellular dataset for spike sorting validation . . . . 30Existing spike sorting ”ground-truth” datasets . . . . . . . . . . . . . . . . . . . . . 30Dataset acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31Results of sorting in vitro datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39Using novel in silico datasets for spike sorting validation . . . . . . . . . . . . . . . . . . . 40Background - computational models of neurons and networks . . . . . . . . . . . . . 41Results of sorting in silico datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45vOngoing efforts for spikes sorting suite development . . . . . . . . . . . . . . . . . . . . . 54Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55Optical Mapping of Spontaneous Single Neuron Activity . . . . . . . . . . . . . . . 58GCaMP6 mapping of spontaneous activity of barrel cortex and thalamic neurons . . . . . 60Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94GCaMP6 mapping of spontaneous activity of auditory and visual cortex neurons . . . . . 98Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101VSD mapping of spontaneous activity of auditory and visual cortex neurons . . . . . . . . 101Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107Multiple Classes of UP-state Transitions and Non-Stationary Single Neuron Fir-ing Order Revealed by Local-Field-Potential Event Clustering During SlowOscillations in Mouse and Cat Cortex . . . . . . . . . . . . . . . . . . . . . . . . 109Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115SWS and UP/DOWN-states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123Clustering multi-laminar LFP events reveals distinct UP-state transition classes . . . 123LECs have discrete laminar CSDs patterns common across tracks and animals . . . 125Most neurons lock to LEC-defined UP-states . . . . . . . . . . . . . . . . . . . . . . 128UP-state transitions have mesoscale correlates in mouse dorsal cortex . . . . . . . . 134Average UP-state transition latencies change over time . . . . . . . . . . . . . . . . . 136Triplet histograms during synchronized state recordings . . . . . . . . . . . . . . . . 141Peak latencies change systematically over time . . . . . . . . . . . . . . . . . . . . . 143Spiking order is present outside UP-state transitions . . . . . . . . . . . . . . . . . . 152Measuring spiking order using adaptive coincidence detection . . . . . . . . . . . . . 158Measuring spiking order using pair-wise cross-correlogram order and the principle ofadditivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173viBibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176Appendix Bionet, Biovis and cluster simulations . . . . . . . . . . . . . . . . . . . . 202Appendix Spikesortingtest.com . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206Appendix Openneuron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208Appendix Raspberry-Pi camera - custom software . . . . . . . . . . . . . . . . . . . 210viiList of TablesTable 1 Cat experiment details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9Table 2 Cat recordings and LEC cluster details . . . . . . . . . . . . . . . . . . . . . . . . 125Table 3 Mouse recordings and clustering details . . . . . . . . . . . . . . . . . . . . . . . . 126Table 4 Cat visual cortex - LEC groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128Table 5 Mouse sensory cortex - LEC groups . . . . . . . . . . . . . . . . . . . . . . . . . . 130viiiList of FiguresFigure 1 Steps involved in spike-sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29Figure 2 Simultaneous extracellular-intracellular in vitro experiments . . . . . . . . . . . . 32Figure 3 Characterizing extracellular/intracellular datasets . . . . . . . . . . . . . . . . . 33Figure 4 Detection threshold vs. false positive rates . . . . . . . . . . . . . . . . . . . . . 35Figure 5 Spatial separation test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38Figure 6 In vitro datasets multi-operator sorting results . . . . . . . . . . . . . . . . . . . 40Figure 7 Single neuron morphologies and synapse locations . . . . . . . . . . . . . . . . . 42Figure 8 Extracellular potentials of multi-compartment neuron models . . . . . . . . . . . 43Figure 9 Allen Institute network model of mouse V1 . . . . . . . . . . . . . . . . . . . . . 44Figure 10 Bionet network examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46Figure 11 Simulated neuron morphologies and physiology . . . . . . . . . . . . . . . . . . . 47Figure 12 8-channel probe simulated traces and sorting results . . . . . . . . . . . . . . . . 48Figure 13 In silico datasets multi-operator sorting results . . . . . . . . . . . . . . . . . . . 49Figure 14 High-density electrode simulation layout . . . . . . . . . . . . . . . . . . . . . . . 50Figure 15 High-density simulated extracellular recordings . . . . . . . . . . . . . . . . . . . 51Figure 16 High-density simulated extracellular recordings - sorting results . . . . . . . . . . 52Figure 17 Sorting results - 2-column vs 1-column electrode layouts . . . . . . . . . . . . . . 53Figure 18 Sorting results are similar across in vitro, in vivo and in silico datasets . . . . . 54Figure 19 Allen Institute active conductance neurons have small spatial propagation . . . . 56Figure 20 Experiment setup: multichannel electrode recordings and spike-triggered-averaging 62Figure 21 Spectral distribution of spontaneous activity . . . . . . . . . . . . . . . . . . . . 63Figure 22 Deconvolution does not substantially change STMs . . . . . . . . . . . . . . . . . 65Figure 23 Control recordings and convergence of STMs with # of spikes . . . . . . . . . . . 67Figure 24 Topographic properties of neurons and neighbouring similarity metric . . . . . . 69Figure 25 LFP and MUA triggered STMs: cortical neuron examples . . . . . . . . . . . . . 71Figure 26 LFP and MUA triggered STMs: subcortical neuron examples . . . . . . . . . . . 73Figure 27 STMs vs SPMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75Figure 28 STM comparisons with Allen Brain Atlas projection maps . . . . . . . . . . . . . 76Figure 29 Cortical and subcortical motifs and STMTDs . . . . . . . . . . . . . . . . . . . . 78Figure 30 Firing modality-defined STMs and STMTDs . . . . . . . . . . . . . . . . . . . . 80Figure 31 Heuristically defined bursting reveals similar STMTDs for cortical neurons . . . 81Figure 32 Heuristically defined bursting confirms similar STMTDs for subcortical neurons . 82ixFigure 33 Single neurons participate in stereotyped network activity patterns during allspontaneous activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83Figure 34 Cortical and subcortical examples of STM-space grouping and decomposition . . 85Figure 35 Example of STM vs. Variance STM . . . . . . . . . . . . . . . . . . . . . . . . . 86Figure 36 STMTD clustering suggests STMTD dynamics represent discrete single neuronphysiological properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89Figure 37 STM hemodynamic corrections are small compared to SNR of calcium activity . 91Figure 38 Body movement contributes minimally to STMs . . . . . . . . . . . . . . . . . . 93Figure 39 ROI-specific dynamics reveal cortical stereotypy and subcortical diversity . . . . 95Figure 40 Single neuron STMs - auditory and visual cortex . . . . . . . . . . . . . . . . . . 99Figure 41 Auditory cortex neurons - firing rates and STM-space distributions . . . . . . . . 100Figure 42 Visual cortex neurons - firing rates and STM-space distributions . . . . . . . . . 102Figure 43 Single neuron VSD STMs - visual cortex . . . . . . . . . . . . . . . . . . . . . . . 103Figure 44 Single neuron VSD STMs - auditory cortex . . . . . . . . . . . . . . . . . . . . . 105Figure 45 Single Neuron STMs - Subcortical Neurons . . . . . . . . . . . . . . . . . . . . . 106Figure 46 Example (#1) subcortical neuron VSD STM - complete motif . . . . . . . . . . . 107Figure 47 Example (#2) subcortical neuron VSD STM - complete motif . . . . . . . . . . . 108Figure 48 Power spectrograms and stimulus annotations - Cat C3 . . . . . . . . . . . . . . 111Figure 49 Power spectrograms and stimulus annotations - Cats C1 and C2 . . . . . . . . . 112Figure 50 Power spectrograms and stimulus annotations - Cat C4 . . . . . . . . . . . . . . 114Figure 51 Power spectrograms and stimulus annotations - Cat C5 . . . . . . . . . . . . . . 116Figure 52 LFP negativities mark single neuron UP-state transitions . . . . . . . . . . . . . 120Figure 53 Multi-channel LFP is more stereotyped than single neuron potentials during UP-state transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121Figure 54 Latency order results from other studies . . . . . . . . . . . . . . . . . . . . . . . 122Figure 55 Clustering LFP events during synchronized cortical states . . . . . . . . . . . . . 124Figure 56 CSD-based LEC grouping and temporal precision . . . . . . . . . . . . . . . . . . 127Figure 57 LEC groupings - cat visual cortex . . . . . . . . . . . . . . . . . . . . . . . . . . 129Figure 58 LEC groupings - mouse visual cortex . . . . . . . . . . . . . . . . . . . . . . . . . 130Figure 59 LEC groupings - mouse barrel cortex . . . . . . . . . . . . . . . . . . . . . . . . . 131Figure 60 LEC groupings - mouse auditory cortex . . . . . . . . . . . . . . . . . . . . . . . 131Figure 61 Single neuron UP-state latencies in cat and mouse cortex . . . . . . . . . . . . . 132Figure 62 UP-state latencies: 1-Sec interval . . . . . . . . . . . . . . . . . . . . . . . . . . . 133Figure 63 Mesoscale VSD correlates of UP-state transitions . . . . . . . . . . . . . . . . . . 135Figure 64 Average UP-state latencies change over time . . . . . . . . . . . . . . . . . . . . 137Figure 65 UP-state latency drift - visual cortex neurons . . . . . . . . . . . . . . . . . . . . 140Figure 66 Triplet histograms - examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141xFigure 67 Examples of triplet-histogram drift . . . . . . . . . . . . . . . . . . . . . . . . . . 142Figure 68 Peak latency-space analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145Figure 69 Additional methods for tracking latency drift . . . . . . . . . . . . . . . . . . . . 147Figure 70 PL similarity across spontaneous and stimulus recordings . . . . . . . . . . . . . 148Figure 71 Trends in UP-state transition distributions across spontaneous and stimulusrecordings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150Figure 72 Latency drift vs. firing rate and cell type - cat visual cortex . . . . . . . . . . . . 151Figure 73 Method for detection of ordered synchrony . . . . . . . . . . . . . . . . . . . . . 153Figure 74 Ordered synchrony matrix - awake mouse recording . . . . . . . . . . . . . . . . 154Figure 75 Ordered synchrony is preserved between spontaneous and stimulus evoked record-ings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156Figure 76 Similarity of ordered synchrony across spontaneous and stimulus evoked periods 157Figure 77 Local time-window metrics of neural synchrony . . . . . . . . . . . . . . . . . . . 160Figure 78 Computing firing order using CCHs and the principle of additivity . . . . . . . . 163Figure 79 Local time-window metrics of neural synchrony . . . . . . . . . . . . . . . . . . . 164Figure 80 BioVis - visualization of cortical patches . . . . . . . . . . . . . . . . . . . . . . . 203Figure 81 BioVis - somas of multi-layer mouse V1 patch . . . . . . . . . . . . . . . . . . . . 204Figure 82 Spikesortingtest.com - workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207Figure 83 Openneuron - data pre-processing functions . . . . . . . . . . . . . . . . . . . . . 209Figure 84 Openneuron - UP-state latency toolbox . . . . . . . . . . . . . . . . . . . . . . . 209Figure 85 Raspberry-Pi camera - dropped frame histograms . . . . . . . . . . . . . . . . . . 211Figure 86 Computing STMs using frame tracking on the Raspberry-Pi . . . . . . . . . . . . 212xiList of Abbreviationsα-BTX α-bungarotoxinAP action potentialCCH cross correlation histogramCOM centre of mass - referring to histogram distributionsCSD current source densityEEG electroencephalogramEMD earth mover’s distancefMRI functional magnetic resonance imagingFP fluorescent proteins (or false positive)fps frames per secondFWHM full width half maximumGECI genetically encoded calcium indicatorGFP green fluorescent proteinGUI graphical user interfaceIM intra muscularISI inter-spike intervalIV intra-venousLEC LFP event classLGN lateral geniculate nucleusLFP local field potentialL2/3 layer 2/3 neuronsL4 layer 4 neuronsL5 layer 5 neuronsxiiMUA multiunit activityNSG Neuroscience Gateway, UCSDNRT thalamic reticular nucleusPC principal componentPCA principal component analysisPETH peri-event time histogramPL peak latencyPTP peak-to-peak amplitude of extracelluar eventREM rapid eye movementRF receptive fieldRMS root mean squareSTA spike-triggered averageSTD standard deviationSTM spike-triggered-mapSVM single value metricSWS slow wave sleepV1 primary visual cortexVSD voltage sensitive dyeTC thalamo-cortical neuronWFOM wide field optical mappingxiiiNote to ReaderIn Chapter 5, the term LFP-event-class, i.e. LFP event class (LEC), is used interchangeablywith UP-state transition. The basis for this equivalence is explained in detail in the chapter. Briefly,several publications have linked large amplitude multi-channel LFP deflections to single-neuron UP-state transitions (e.g. Volgushev et al., 2006; Chauvette et al., 2010). The equivalence is also madebased on findings made within the chapter, e.g. LECs have large amplitudes and occur at a ratepreviously associated with UP-state transitions and single neurons spike substantially during LECevents. As the author did not carry out single neuron patch pipette recordings to confirm thetraditional definition of UP-state against the LECs, the equivalence between LECs and UP-statetransitions is provided as putative at this time. However, Chapter 5 contains an analysis of theshortcomings of a single-neuron patching approach for defining global UP-state transitions and thesuperiority of the LEC defined method.In Chapter 5 the term stable neuron is used to refer to spike sorted neurons that had voltage PTPamplitudes >75µV and that fired consistently throughout the entire recording periods considered(see also Methods). These stable neurons were unlikely to have substantial spike-sorting errorsand were chosen based on qualitative attributes. In the absence of ground truth data, they weredetermined to be the best neurons available (in those specific datasets) for analysis.xivAcknowledgementsMy first supervisor, Nicholas Swindale, who 4 years ago gave me an opportunity to work in neu-robiology despite my limited knowledge and years of being out of academic pursuits. Nick neverfailed to answer my never-ending stream of questions (many of them by email) and always gaveme wise advice that I came to appreciate days, months or years later. His scientific integrity andinsistence on clear evidence for conclusions will guide my research for the rest of my career.My second supervisor, Timothy Murphy, who welcomed me into his lab 2 years ago and gaveme encouragement and support while learning new experimental methods. Even a few minutes ofadvice from Tim on research and career directions was always significant in shaping my interestsand thinking.My external supervisor, Costas Anastassiou, who hired me at the Allen Institute for Brain Sci-ence over 3 years ago based on our shared enthusiasm for modeling despite my lack of neurobiologyexperience. Costas provided me with research and financial support for years and I will foreverbe indebted to the opportunities provided by him, Christof Koch, Michael Hawrylycz and manyothers at the Allen Institute who have worked hard to built a unique place for doing neuroscience.My colleague Martin Spacek whose practical advice, shared code and home made beer eased thedifficulties of navigating the interdisciplinary research projects tackled in this thesis.My many other colleagues who helped me along the way: Allen Chan, for his help, advice andsupport on working with mice and optical imaging. Jamie Boyd for his patience and for sharing hiswisdom of both hardware and experimental approaches. Alex McGirr for our many discussions onspontaneous neural activity in healthy and diseased brains. Sergey Gratiy for his friendship, ourmany insightful discussions and providing me with a powerful modeling tool for carrying out myresearch. Lastly, my friend and colleague Bruno Herculano for listening to my compaints duringexpected, but still challenging, graduate school experiences.Amitava Majumdar and the staff at the Neuroscience Gateway at the University of California,San Diego, which provided me with more than 1-million core-hours (and counting) of free clustertime and with prompt and effective support with the many challenges involved in running complexnetwork simulations on remote servers thousands of kilometers away.The thousands of software developers who made my research possible by providing free softwarexvtools: Linux, Python, Geany, Git, LATEX, BibTEX, GIMP.The work presented here was supported by funding from Canadian Institutes of Health Research(CIHR), the Natural Sciences and Engineering Research Council of Canada (NSERC), the Univer-sity of British Columbia neuroscience bursary program and the Allen Institute for Brain Science.xviIntroduction“Traditionally studies of brain function have focused on task-evoked responses. Bytheir very nature, such experiments tacitly encourage a reflexive view of brain function.Although such an approach has been remarkably productive, it ignores the alternativepossibility that brain functions are mainly intrinsic, involving information processingfor interpreting, responding to and predicting environmental demands. Here I arguethat the latter view best captures the essence of brain function, a position that accordswell with the allocation of the brains energy resources.“Raichle, 2010The brain of the Portia (genus) spider (Salticidae family) contains 600,000 neurons and can fit”comfortably on the head of a pin” (Prete, 2004). Yet the spiders are capable of complex huntingbehaviours including: trial and error behaviour against novel prey or novel situations and remem-bering the new approaches (Wanless, 1975); solving problems and exhibiting sequences of behaviourmimicking prey-spider web signals (Harland and Jackson, 2000); and making hunting plans that cantake up to one-hour and involve detouring past an incorrect route while loosing sight of their prey(Wilcox and Jackson, 2002). How can such a small animal exhibit complex behaviour requiringplanning and delay of automatic hunting instincts? The answer - for the most part - is we do notknow, and only very recently have electrophysiologists successfully made neurophysiology recordingin the brains of such small organisms (Menda et al., 2014).The human brain contains almost 100 billion neurons and can also support complex, uniqueand high-level abilities such as executive-functions, long-term memory and social behaviours. Inboth spiders and humans the most interesting and cherished behaviours are not related to theprocessing of immediate stimuli in the environment, e.g. recognizing an odor or a colour. Yet,largely due a limited understanding of nervous systems and limitations in experimental tools, mostof 20th century systems neuroscience studies have focused on characterizing sensory processing, forexample, how sensory systems instantiate fast circuits (i.e. on the scale of 1ms-1sec) to processstimuli (e.g. visual, olfactory or auditory inputs; Marr and Poggio, 1976). Recently, however,it is becoming more accepted that brains do much - if not most - of their work during periodsof spontaneous activity and that incoming sensory stimuli largely modulate such activity (Foxand Raichle, 2007; Raichle, 2010). If what drives brain evolution is the development of systemsthat avoid purely reflexive behaviours with ever increasing delays between sensory experiences andbehaviour, then understanding spontaneous (i.e. not stimulus evoked) neural activity is the keyto understanding higher brain function. This thesis is aimed at investigating spontaneous neuralactivity across multiple spatial and temporal scales. It relies on cat visual cortex and mouse cortexand thalamus recordings and focuses on characterizing what single, spontaneously firing neuronscan tell us about neural activity during both spontaneous and stimulus evoked periods.1BackgroundMaking sense of the meaninglessly large (Barlow, 1961) neural activity patterns from brain record-ings remains a significant - if not the most central - challenge of neuroscience. Theoretical andanalytical approaches have generally referred to this problem as that of identifying the neuralcode that is implemented by nervous systems. In seeking to answer these and other foundationalquestions, modern experimental neuroscience methods have improved substantially and allow forthe study of neural activity with unprecedented temporal and spatial scale and precision. Sensorysystems (e.g. visual cortex of mouse) can now be studied with millisecond precise recordings fromdozens to 100 or more single neurons using electrophysiology tools (with hundreds to thousandsof neurons soon being possible Jun et al., 2017b) and with lower temporal resolution but acrossa few millimeters using imaging methods to capture the activity of as many as 10,000 neuronssimultaneously (Pachitariu et al., 2017).Yet, it is unclear whether new technologies alone employed in sensory areas using existingand future reflexive experimental paradigms will significantly advance our understanding of brainfunction. In fact, it has been known for many years that brains are spontaneously active consuming10 times as much energy as expected by weight alone (Clarke and Sokoloff, 1999) with intrinsicbrain activity consuming merely 5% less energy than used during intense mental effort (Sokoloff etal., 1955).The debate about whether the primary task of nervous systems is to represent and processsensory stimuli versus being intrinsically active to predict environmental and internal demands ismore than 100 years old (Sherrington, 1906; Brown, 1914). Some theoretical work suggested thateven sensory systems (which make up only a part of cortical tissue) are not merely tasked withrepresentation of stimulus or relay of their content but also at ”reducing redundancy” (Barlow,1961) in sensory information and passing only errors to internal systems that perpetually generateand evaluate models of the world. Thus, sensory (and most other neural) systems presumably makeunconscious inferences (Clark, 2013): they compare inferences with incoming stimulus informationand only errors between the expectation and the stimulus are propagated to higher brain regions.Other similar theories followed, such as predictive coding theories - which claims that all brain areas(not just low level sensory systems) are tasked with producing and testing hypotheses about theexternal world against low level sensory input (Rao and Ballard, 1999). More recently, discussionshave also focused on the intrinsic properties of cortical circuits, e.g. as central pattern generators(CPG) (Yuste et al., 2005) and viewing even the psychological ego as a result of systems thatpredict behaviour and environment demands (Llinas, 2001).Thus, focusing exclusively on reflexive neuroscience paradigms (i.e. recording neuronal activityin response to stimuli) may be insufficient for a complete understanding of brain function. Thereason is that there is mounting evidence that sensory systems are driven by other non-sensorycortical areas as well as subcortical structures to be spontaneously active and are only partially2modulated by incoming sensory stimuli (Raichle, 2010).However, reflexive neuroscience should not be abandoned. Many, if not most findings on corticalfunction have been based on studying the response - often the firing rates - of neurons during sensorystimulation (for example, Adrian, 1926; Hubel, 1959; Shadlen and Newsome, 1994). This traditionalapproach has lead to the firing-rate neural coding theory which states that neurons representinformation (presumably for other, downstream neurons) through their firing rates. There is ampleevidence for firing rate based coding, for example: visual cortex neurons (Hubel, 1959), LGN andV1 neurons in awake and attending monkeys (Oram et al., 1999), single patched neurons in thebarrel cortex of anesthetized rats (Latham et al., 2006; London et al., 2010), modeling (Softky, 1995;Shadlen and Newsome, 1998; Oram et al., 1999) and many other studies. And decision makingstudies over the past two decades have shown neurons in parietal cortex (a higher-level associationarea in monkeys) increase their firing rates to reflect accumulation of optimal stimuli over multiplestimulus presentations spanning many seconds (Gold et al., 2007).But over the past couple of decades evidence of precise spike timing relationships being presentin cortex, especially low sensory cortical areas, has been slowly accruing. At the single neuron level,the evidence that precise input times are important has been available for almost two decades inthe form of spike-timing dependent plasticity (STDP) (Bi and Poo, 1998). In recordings of multipleneurons there has also been evidence of spike timing is important in many studies: auditory systems(Jeffress, 1948), visual psychophysics experiments (Burr and Ross, 1979), cat visual cortex (Grayet al., 1989), chains of firing neurons (”synfire chains”) in monkey and other cortical recordings(Abeles and Gerstein, 1988; Abeles, 1991), theoretical approaches (Thorpe, 1990), anesthetizedcat visual cortex slices (Mainen and Sejnowski, 1995), monkey frontal cortex during behaviouraltasks (Vaadia et al., 1995), extrastriate (visual) cortex of behaving monkeys (Bair and Koch, 1996),retinogeniculocortical pathway in cat (Usrey et al., 2000), somatosensory thalamocortical pathwayin rat (Ahissar and Arieli, 2001), hyppocampus (Harris et al., 2002, 2003), thalamocortical neurons(Salami et al., 2003), rat auditory cortex (Wehr and Zador, 2003), visual cortex and simulations(Azouz and Gray, 2003), mouse visual cortex (Ikegaya et al., 2004), human somatosensory afferents(Johansson and Birznieks, 2004), cat visual cortex (Fregnac and Zador, 2005; Havenith et al., 2011),at visual cortex (Meliza and Dan, 2006), retina (Hausselt et al., 2007; Gollisch and Meister, 2008),rat somatosensory and auditory cortex (Luczak et al., 2007, 2009, 2013; Bermudez-Contreras et al.,2013; Luczak et al., 2015), tactile perception (Mackevicius et al., 2012), gerbil inferior colliculus (au-ditory hub) (Garcia-Lazaro et al., 2013). Some of the evidence for spiking timing being importantsuggests that neurons that prefer the presented stimulus respond earlier during the presentation:theory (Thorpe, 1990; VanRullen et al., 2005), auditory cortex (Heil, 2004); somatosensory cortex(Johansson and Birznieks, 2004); retina (Gollisch and Meister, 2008).However, combining the study of spike timing with spontaneous activity paradigms is very chal-lenging. The reason is that there are no precise triggers as in the case of stimulus or behaviourparadigms (e.g. stimulus ON-time or behaviour onset) for triggering single neuron activity analysis.3A good starting point, however is to note that spontaneous neural activity can have non-randomstructure. In fact, an early hypothesis that the ”machinery” underlying spontaneous cortical ac-tivity during slow-wave-sleep K-complexes (i.e. spontaneously occuring transitions in neurons fromhyperpolarized non-spiking to active, spiking states) was used by cortex to process sensory stimuli(Amzica and Steriade, 1998a) lead to findings of preserved temporal structure in neuron firingorder. In particular, several studies over the past decade have found that the firing order of highfiring rate neurons is similar during spontaneous (including anesthetized) states as during pro-cessing (Luczak et al., 2007, 2009, 2013; Bermudez-Contreras et al., 2013; Luczak et al., 2015).Ironically, spontaneous activity transitions occuring during sleep (in particular, slow-wave-sleep)when sensory stimulus does not seem to propagate to higher cortical areas (e.g. executive functionareas) may provide an intrinsic trigger which can enable the study of neural activity similar toreflexive methods which use the onset of sensory stimuli or behaviour as triggers. Other findingsover the past two decades using two-photon imaging and single neuron patch clamping methodsalso suggest that co-activated neuronal ensembles during spontaneous activity can also be similarto patterns observed during stimulus evoked periods (Cossart et al., 2003; Ikegaya et al., 2004;Miller et al., 2014; Carrillo-Reid et al., 2015) with some suggesting a move away from the singleneuron to ensembles as the basic units of cortical processing (Yuste, 2015).While there are interesting hypotheses and theoretical concepts about cortical function, arguablymost modern neuroscience discoveries continue to be driven by novel experimental paradigms withelectrophysiology and more recently optical imaging, being the principal methods for studyingneural systems with high temporal and spatial precision.Electrophysiolgy is unarguably the most common and successful method for investigating neuralactivity over the past 150 years and has been the main method of recording for most of the studiesdiscussed so far. While it has been known since the 17th century (Jan Swammerdam) that frogmuscles respond to electrical stimulation, recordings from single nerve fibers were only made in 1928(Adrian and Bronk, 1928) and were followed by glass micro-electrode recordings of single neuronsin cat hippocampus in 1940 (Renshaw et al., 1940). By that point it had been known for some timethat single neurons have periodic large ”negative” electrical defelctions (now called somatic ”actionpotentials”, APs) following early observations by Emli du Bois Reymond in 1943 (see Schuetze,1983 for a review). In addition to direct single neuron recordings, electrical activity in the brain hasalso been measured by inserting a wire or glass electrode into neuronal tissue and recording the total”extracellular” electrical potential (relative to a reference) from all (nearby) sources. The earliestextracellular recordings were made by Steve Kuffler beginning in the 1940s in muscle (Kuffler,1946) and cat retina (Kuffler, 1953). Tungsteen microelectrodes were developed (Hubel, 1957;Green, 1958; Hubel, 1959) and lead to significant findings in cat visual cortex (Hubel and Wiesel,1962). Many other innovations followed, including tetrodes that rely on groups of 4 wires bundledtogether (Gray et al., 1995), to silicon polytrodes that can simultaneously record from dozensor hundreds of sites (Drake et al., 1988). Extracellular recordings capture the total Local Field4Potential (LFP) of both single neuron and collective neuron electrical activity from all nearby cellsincluding synaptic activity, sodium and calcium spiking and even much slower glial electrical activity(Buzsa´ki et al., 2012). While some have found that more than 95% (i.e. of the total amplitude)of the LFP originates from current sources <250µm from the extracellular contact (Katzner et al.,2009) some studies have suggested LFP can represent neuronal electrical activity from up to ∼1cmaway (Yoshinao and Charles, 2011). Such differences in estimates likely arise from differences intissue across different areas, animals or neuronal activity states (e.g. cortical state). There havebeen multiple approaches for developing high-density silicon multisite electrodes (polytrodes) overthe past two decades (Drake et al., 1988; Henze et al., 2000; Harris et al., 2000; Csicsvari et al.,2003; Blanche et al., 2005a; Bere´nyi et al., 2014) with the most recent electrodes containing ≈700(Jun et al., 2017b) and ≈1000 (Lopez et al., 2017) channels (up to 384 simultaneously accessible).While single neuron spikes remain central to understanding neuronal function, one limitationto using high-density extracellular electrodes to detect single neuron activity is that single unitisolation - i.e. ”spike sorting” - can become very challenging and computationally expensive whenrecording many neurons on hundreds of electrode channels. In fact, even for older electrode record-ings containing only a few simultaneously acquired voltage channels, sorting algorithm developmentrequires ”ground-truth” datasets - i.e. datasets containing both extracellular voltage records forsorting and the actual spiking patterns of neurons (i.e. the rasters) in order to verify and quan-tify the sorting results. There are very few in vivo and in vitro ground-truth datasets. However,with modern computational modeling tools, datasets can also be generated using biophysically de-tailed in silico (i.e. simulated) cortical networks that can generate large numbers of ground-truthspiking rasters and arbitrary electrode layout configurations. Computational modeling of singleneurons and simulations of network activity started in with single compartment models over 70years ago (McCulloch and Pitts, 1943). Dendritic processing was incorporated into models using”cable theory” (Rall, 1964) and with the advent of modern computers, new simulation programswere developed (e.g. NEURON, Hines, 1986; Hines and Carnevale, 1997; Migliore et al., 2006).Most recently, multi-purpose, highly biophysically detailed, cortical network simulations of rat so-matosensory (Markram et al., 2015) and mouse visual cortex (Hawrylycz et al., 2016; Gratiy etal., 2017) are becoming available. Building highly detailed multi-scale models of cortex will becentral to integrating the very large - and growing - experimental datasets and they can also playimportant roles in improving experimental and analytical approaches in addition to spike sorting.In addition to electrophysiological recordings and modeling research, over the past three decadesnew optical imaging methods have enabled the recording of neuronal activity as a function ofintracellular calcium concentrations (i.e. calcium imaging; Mank et al., 2008; Peterka et al., 2010;Broussard et al., 2014; Lin and Schnitzer, 2016) or as a function of membrane voltage (i.e. voltagesensitive dyes - VSDs; Peterka et al., 2010). The most common calcium imaging methods relyon fluorescent proteins (FPs) that report the intracellular concentrations of calcium in neurons.Much work in the past 15 years has focused on improving a particular class of calcium reporters5that green FPs (GFP) called GCaMP, (Nakai et al., 2001), with GCaMP6f (Chen et al., 2013b),in particular, showing a 28 fold fluorescence change from 0-1M calcium concentrations providinga substantially high signal-to-noise-ratio (SNR). There are some problems such as photobleaching(fluorophore loosing its fluorescence) and phototoxicity (excitation light damaging neurons) andsingle action potential detection is still not possible as calcium signals are biased to reporting singleneuron spiking bursts (than tonic spiking modes; Theis et al., 2016). In contrast to intracellularcalcium probes, membrane voltage reporters (VSDs) are ideal for recording the excitation state of aneuron as they have fast responses and even report sub-threshold as well as supra-threshold neuralactivity (Cohen et al., 1974; Blunck et al., 2004; Peterka et al., 2010). Over the past few decadesdyes developed for invertebrates have been quite successful in reporting single neuron activityyet mammalian dyes are still limited to providing signals of ensemble averages of post-synapticresponses (Kenet et al., 2003). Lastly, genetically encoded calcium indicators (GECIs) in micehave greatly simplified calcium imaging (e.g. Madisen et al., 2010, 2015) and many available mouselines targeting specific neuronal populations are now available (Taniguchi et al., 2011; Madisen etal., 2015), though only a few genetically encoded voltage indicators are currently available (e.g.Siegel and Isacoff, 1997).In sum, over one hundred years of neuroscience investigations have identified several importantquestions in systems neuroscience, with arguably the most central goal being to understand theneural code employed by the spiking of neurons during both stimulus processing and spontaneousactivity. Many experimental tools have been developed to increase neuron yield and targeting speci-ficity with more complex computational modeling tools becoming available for use in neuroscience.While our understanding of spontaneous activity lags behind our knowledge of stimulus and taskevoked neural responses, a number of recent novel approaches are paving the way for closing thisgap.Thesis summaryThis thesis is aimed at investigating spontaneous neural activity in electrophysiology and imagingrecordings. It leverages existing (e.g. cat visual cortex recordings) and new (e.g. mouse cortex)recordings and computational tools (e.g. Allen Institute for Brain Science network models) andconnects novel findings to existing studies on both spontaneous and stimulus evoked activity.Chapter describes experiments and computational methods. For this thesis a number of animalrecordings were made and several computational toolboxes were developed and employed. First,cat visual cortex electrophysiology recordings acquired previously (Swindale and Spacek, 2014) arebriefly described. Mouse sensory cortex and thalamus electrophysiology and imaging recordingsacquired specifically for this thesis are described next. Last, the computational methods employedin this thesis are described in detail including several unique approaches to analysis of neuronalactivity.6Chapter describes in vitro datasets and in silico modeling methods employed for spike-sortingalgorithm testing. The outstanding problem of testing spike-sorting algorithms is described alongwith a historical background of spikesorting suite development and biophysically detailed singleneuron and network models. A novel in vitro dataset acquired by previous researchers (Anastassiouet al., 2015) is reviewed and several spike sorting tests results are discussed. The remainder of thechapter focuses on simulations of extracellular activity and the results of spike sorting such datasets.First, there is a brief review of the Allen Institute biophysically detailed model of mouse V1 andthe Blue Brain Project (BBP) single neuron models. Next a number of sorting results based onsimulations of hundreds to several thousand neurons are presented and reviewed (Mitelut et al.,2014, 2015; Jun et al., 2017a). The chapter ends on a discussion of the realism of simulated datasetsfor spike-sorting testing and ongoing efforts to develop largely automated spikesorting methods (e.g.Jun et al., 2017b).Chapter describes simultaneous widefield imaging and electrophysiology recordings experimentsand analysis for generating dorsal cortex neuronal activity maps from single neuron spontaneousactivity. There are three main sections. The first section has been previously published (Xiao etal., 2017) and contains simultaneous electrophysiology and calcium imaging recordings and analysisobtained from transgenic GCaMP6 mouse recordings in barrel cortex and sensory thalamus. Theaim was to characterize how the spontaneous activity of cortical and subcortical neurons (recordedextracellularly) correlates with widefield bilateral dorsal cortex activity (recorded via calcium imag-ing). The findings are that barrel cortex neurons are co-activated (during awake and anesthetizedstates) with expected functional networks involving bi-hemispheric barrel and motor cortex areas(i.e. ”consensus” maps). In contrast, thalamic neurons are shown to have greater diversity ofco-activation and also reveal more complex temporal dynamics in relation to cortex (for example,some thalamic neurons prefer to fire during large scale cortical depression phases). The remainingtwo sections of the chapter briefly examine additional datasets of recordings using GCaMP6s mice(section 2) and VSD recordings (section 3) in visual and auditory cortex. The findings of theselast two sections largely confirm the approach of the first section, but show that both auditoryand visual cortex neurons may have more diverse maps and spatio-temporal dynamics than barrelcortex neurons.Chapter describes a novel method for clustering spontaneously occuring UP-state transitionsduring synchronized cortical states and investigates single neuron firing order during such tran-sitions. This chapter relies on previously acquired cat visual cortex electrophysiology recordingscoupled with newly acquired mouse sensory cortex electrophysiology and imaging datasets. A novelmethod is developed that is similar to spike sorting which can identify and classify UP-state tran-sitions in cat V1 and mouse sensory cortex using local field potential (LFP) from extracellularrecordings. The LFP events are clustered into distinct classes (i.e. LECs) based on their waveformsimilarity. It is also shown that LECs can be similar within and across animals and that they can belocalized in time with high tempral precision (≈5-15ms). Additionally, it is shown that >90% of all7neurons exhibit spiking latency peaks during LEC-defined UP-state transitions even in visual cor-tex - similar to previous findings in rat (non-visual) cortex. Perhaps most importantly, it is shownusing multiple methods that many neurons can change their latency peaks or distributions duringsynchronized states (i.e. relative to UP-state transitions) and can also change their firing orderoutside of UP-states and during desynchronized cortical states. The firing order changes occur overperiods of many minutes (e.g. 30-120 minutes) and rates of change vary across neurons suggestingthat any underlying spike timing dependent neural code must employ a transient coding/decodingscheme.In sum, the work in this thesis is aimed at improving single unit isolation (via spike sortingalgorithm testing) and characterizing spontaneous neural activity in cat visual cortex and mousesensory cortex across multiple spatial and temporal scales. The findings of unique cortical mapsand drifting firring order are novel and suggest further avenues for research into spontaneous neuralactivity.8Experimental and Analytical MethodsElectrophysiologyCat acute electrophysiological recordingsExperimental procedures were carried out in accordance with guidelines established by the Cana-dian Council on Animal Care and institutional protocols approved by the Animal Care Committeeof the University of British Columbia. The experiments were carried out by Nicholas V. Swindaleand Martin Spacek and the methods have been described in full elsewhere (Swindale and Spacek,2014). Five cats were used in total: three were normal domestic cats, while two were heterozygouslipoprotein lipase deficient, left over from an unrelated study (Table 1). Cat ID’s (used throughoutthis thesis) are: C1, C2, C3, C4 and C5. Initial stages of each animal experiment were performedwith the supervision of a veterinarian. For 3 of the 5 cats (C3, C4, C5), initial sedation was byintra muscular (IM) injection of dexmedetomidine (25 µg/kg) and initial analgesia by IM injectionof butorphanol (0.3 mg/kg). An intra-venous (IV) catheter was inserted, and initial anesthesia wasinduced by IV injection of sodium thiopental or propofol. An endotracheal tube was then insertedand a catheter placed in the urethra. The animal was placed in a stereotaxic frame and its headfixed in place with ear bars coated in topical anesthetic (5% lidocaine). The stereotaxic frame wasmounted on an air table which was floated prior to polytrode insertion to minimize vibrations.Following sedation and surgery, electrophysiological recordings were made in the visual cortexof cats with either 0.5 - 1.5% isoflurane and 70% N2O + 30% O2 (animal IDs C1, C2 and C3)or with continuously infused propofol (6-9 mg/kg/hr) and fentanyl (4-6 110 g/kg/hr) (C4 andC5). Heart rate and blood oxygenation were monitored with a pulse-111 oximeter (Nonin 8600V).Animal Sex Age Weight Source DrugsID (y) (kg) set Anesthestic Paralytic OtherC1 M 7 7.0 HLLD iso + N2O α-BTX buprC2 M 5 5.1 HLLD iso + N2O α-BTX bupr, atrop, xylaC3 F 1 3.2 UCD iso + N2O PB bupr, dobutC4 F 1 3.4 UCD prop + fent PB bupivicaineC5 F 1 3.5 UCD prop + fent PBTable 1: Cat experiment details. UV: Unique Ventures, Balmoral, MB. HLLD: heterozygouslipoprotein lipase deficient cats, UC Davis. UCD: UC Davis. iso: isoflurane. prop: propofol.fent: fentanyl. PB: pancuronium bromide. α-BTX: α-bungarotoxin retrobulbar injection. bupr:buprenorphine. atrop: atropine. xyla: xylazine. dobut: dobutamine. Adapted from Spacek,2015 with permission.9Mean arterial blood pressure was monitored with a Doppler blood pressure monitor (Parks Medical811-B) placed on a shaved section of hind leg. Body temperature was maintained at 37C with ahomeothermic blanket (Harvard Apparatus). Eye movements were prevented either by continuousinfusion of pancuronium bromide (C3, C4 and C5) or by retrobulbar injections of -bungarotoxin(Tocris) (C1 and C2). Pupils were dilated with tropicamide drops (0.5%). The eyes were refractedto a viewing distance of 57 cms with rigid gas-permeable contact lenses. After the craniotomy wasmade, the dura removed and a nick made in the pia with an ophthalmic slit knife. The electrodewas then inserted perpendicularly into the cortex under visual control so that the upper recordingsites lay ≈100-200µm below the surface. The craniotomy was filled with agarose gel (2.5%, TypeIII-A, Sigma-Aldrich, St. Louis, MO) in artificial CSF at 38 40C. Recordings were made with54-channel polytrodes (University of Michigan Center for Neural Communication Technology andNeuroNexus).Polytrodes were either of 2-column (C2, C3 and C5) or 3-column (C1 and C4) design. Voltagesignals were analogue bandpass filtered between 0.5 and 6 kHz, sampled at a rate of 25 kHz anddigitized with 12-bit resolution (Blanche et al., 2005b). LFP recordings were obtained from 10 ofthe 54 channels, and fed in parallel to a separate set of amplifiers. These channels were numberedsequentially from 1 10, starting at the top of the electrode. On the 3-column electrodes thechannels were 130µm apart, with the exception of channels 9 and 10, which were 65µm (C1) or97µm (C3) apart. On the 2-column electrodes the channels were 150µm (C2 and C4) or 195µm (C5)apart, with channels 9 and 10 being 100µm (C2 and C4) or 195µm (C5) apart. LFP voltages wereanalogue bandpass filtered between 0.1 and 150 Hz, sampled at a rate of 1 kHz and digitized with12-bit resolution. Experiments lasted up to 3 days. Data reported here were obtained from a totalof 14 electrode penetration sites in 5 adult cats. Recording sites were in area 17 and receptive fields(not reported here) were typically within 10 degrees of the area centralis. Individual periods ofspontaneous activity lasting from 5-60 minutes (average = 27 minutes) were recorded for each trackwith most tracks having multiple spontaneous activity recording periods. During these periods thecat viewed, either binocularly or monocularly through the dominant eye, through 3 mm artificialpupils, a field of uniform luminance (58 cd/m2) on a CRT monitor (Iiyama HM903DTB) with arefresh rate of 200 Hz, positioned 57 cms from the eyes.Cat visual stimuliVarious sets of stimuli were used in cat visual cortex recordings and has been described elsewhere(Spacek, 2015; Swindale and Spacek, 2014). The approach is briefly described here (and largelyadapted from previous publications). The receptive field (RF) was first identified in real-time usinga mouse-controlled oriented bar. Drifting bars stimuli consisted of white and/or black bars oneither a grey background with bars 10 or 6◦ long and 0.5 or 0.3◦ wide, drifting at 2.5 or 5◦/s for 4s,for a total of 10 or 20◦. Flashed gratings were composed of stationary, spatially sinusoidal gratings(of various sizes) displayed for 40ms each. Drifting gratings were similar to gratings and consisted10in various temporal (usually ¡ 5Hz) and spatial frequencies. M-sequence stimuli (Shapley et al.,1991; Reid et al., 1997) were composed of 32 x 32 pixel movies with 65535-frames each presented for20 or 40ms (summing all frames gave a grey image). Spontaneous activity was generally acquiredwhile the cat viewed a blank grey screen.Two types of natural scene movie stimuli were used: a (64×64 pixel) video provided by PeterKo¨nig lab using recordings from a camera mounted on a cat’s head as it behaved naturally (Kayseret al., 2003); and newer movies filmed by Martin Spacek using a digital camera (Canon PowerShotSD200) at 320×240 pixel resolution and 60 frames per second (fps) of dense grass and foliageand other environments while also emulating sudden saccade-like movements. Movies were 1-5minutes in duration and were presented either in entirety, up to 8 times in a single recording, orin short 4.5-5s clips that were repeated 200-400 times. Most stimuli were presented at 200Hz, withselected recordings at 66Hz. All stimuli were shown monocularly depending on which eye revealedthe most neural activity (except full-screen flashes and some blank screen stimuli). During eachtrack, stimulus recordings took several hours with varied type of stimuli were presented usually inrandomly chosen order by the experimentalists (Martin Spacek and Nicholas Swindale).Mouse selectionMice experiments used in Chapter 4 were carried out by Dongsheng Xiao, Mattheiu Vanni andthe author. Mice experiments used in Chapter 5 were carried out by the author. Animal proto-cols (A13-0336 and A14-0266) were approved by the University of British Columbia Animal CareCommittee and conformed to the Canadian Council on Animal Care and Use guidelines. Animalswere housed in a vivarium on a 12 h day light cycle (7 AM lights on). Experiments (Chapter 4 andChapter 5) were performed towards the end of the light cycle (i.e. late afternoon/early evening).For recordings in Chapter 4, transgenic GCaMP6f mice (males, 24 months of age, weighing 2030g; n=16), were produced by crossing Emx1-cre (B6.129S2-Emx1tm1(cre)Krj/J, Jax # 005628),CaMK2-tTA (B6.Cg-Tg(Camk2a-tTA)1Mmay/DboJ, Jax # 007004) and TITL-GCaMP6f (Ai93;B6;129S6-Igs7tm93.1(tetO-GCaMP6f)Hze/J, Jax # 024103) strains (Madisen et al., 2015). Forrecordings in Chapters 4 and 5 transgenic GCaMP6s mice (n=3: Chapter 4; n=13: Chapter5) were produced by crossing Emx1-cre (B6.129S2-Emx1tm1(cre)Krj/J, Jax # 005628), CaMK2-tTA (B6.Cg-Tg(Camk2a-tTA)1Mmay/DboJ, Jax # 007004) and TITL-GCaMP6s (Ai94;B6.Cg-Igs7tm94.1(tetO-GCaMP6s)Hze/J, Jax # 024104) strains. The presence of GCaMP expressionwas determined by genotyping each animal before each surgical procedure with PCR amplification.These crossings are expected to produce a stable expression of the 3 calcium indicator variants(GCaMP3, GCaMP6s and GCaMP6f (Chen et al., 2013b) specifically within all excitatory neuronsacross all layers of the cortex (Vanni and Murphy, 2014). Control experiments (Chapter 4), as-sessing the specificity of spike triggered average maps, were performed in Thy1-GFP-M mice (n=6;Jax # 007788). No method of randomization was used since all mice belonged to the same samplegroup. Samples sizes were chosen based on previous studies using similar approaches (Mohajerani11et al., 2013; Vanni and Murphy, 2014; Chan et al., 2015).Mouse surgeryMice were anesthetized with isoflurane (1.5-2%) for induction and during surgery and a reducedmaintenance concentration of isoflurane (0.5-1.0%) or urethane was used later during anesthetizeddata collection. In some cases, animals were allowed to wake up following isoflurane anesthesia forawake imaging (see Chapter 4, section Multimodal recording in awake mice). Throughout surgeryand imaging, body temperature was maintained at 37oC using a heating pad with a feedbackthermistor. For cortical recording experiments, mice were placed on a metal plate that was mountedon a macroscope. In order to minimize movement artifact (due to breathing and heartbeat),the exposed skull was fastened to a stainless steel head-plate with cyanoacrylate glue and dentalcement. A 9 x 9 mm bilateral craniotomy (bregma 3.5 to -5.5 mm, lateral -4.5 to 4.5 mm) coveringmultiple cortical areas was made as described previously (Mohajerani et al., 2013). For sub-corticalexperiments, mice were placed in a stereotaxic apparatus and an incision was made in the midlineto expose the skull as in cortical experiments. A burr hole was then unilaterally drilled (usuallyin the right hemisphere) above the thalamic area (stereotaxic coordinates considering a 45 degreeangle: between 1.7±0.3 mm posterior to bregma and 1.6±0.4 mm lateral to midline). The angulartilt relative to a perpendicular penetration to the cortical surface was estimated to be of less than5o (Hunnicutt et al., 2014). In cases where the laminar probe was inserted (as opposed to a glasselectrode) a craniotomy was only made for the probe insertion site and cortical GCaMP imagingwas performed through intact bone.Mouse acute electrophysiological recordingsFor recordings in Chapter 5, GCaMP6s mice (13 males, 2-8 months of age, weight 20-35g) were usedalong with wildtype mice (C57/BL6; 3 males, 2-8 months of age, weight 20-35) for chronic record-ings (1 mouse) and voltage-sensitive-dye recordings (2 mice). Transgenic GCaMP6s mice were pro-duced by crossing Emx1-cre (B6.129S2-Emx1tm1(cre)Krj/J, Jax #005628), CaMK2-tTA (B6.Cg-Tg(Camk2a-tTA)1Mmay/DboJ, Jax #007004) and TITL-GCaMP6s (Ai94;B6.Cg-Igs7tm94.1(tetO-GCaMP6s)Hze/J, Jax #024104) strain (Madisen et al., 2015). They expressed calcium withinexcitatory neurons across all layers (Vanni and Murphy, 2014). GCaMP6s (Chen et al., 2013b)expression was validated by genotyping with PCR amplification. Surgical procedures for acuteexperiments are described in full elsewhere (Xiao et al., 2017). Briefly, mice were anesthetizedwith isoflurane (1.5-2%) for induction and during surgery with subsequent recording periods underreduced concentration of isoflurane (0.8-1.2%) or urethane (≈1mg/gram body weight). For acuteexperiments the skull was fixed to a head-plate to stabilize recordings and facilitate imaging. Extra-cellular recordings in mice were made with 64-channel polytrodes (A1x64-Poly2-6mm-23s-160-A64;NeuroNexus, Ann Arbor, MI) which have a 2-column (32 channels per column) staggered-format12with vertical and horizontal (inter-column-distance) of 46µm covering an ≈1450µm length of theprobe. Voltage signals were acquired using a headstage amplifier (RHD2164, IntanTech, Los An-geles, CA) and USB interface board (RHD2000, IntanTech) at a sample rate of 25 kHz (16bit).Electrodes were inserted perpendicular to the surface of the cortex using a micro-manipulator (MP-225, Sutter Instrument Company) with some exceptions (noted in main text) required for simulta-neous optical imaging of dorsal cortex. Cortical penetration was tracked using micro-manipulatorcoordinates and varied between ≈900µm to 1450µm (mean of 1256µm±157µm).Mouse chronic electrophysiological recordingsChronic tetrode implant recordings were performed in three mice (C57/BL6) with surgery protocolsas described above (and in Xiao et al., 2017) with changes specific to tetrode implants. Specifically,tetrodes consisted of bundles of 16 wires (each 15µm in diameter) grouped in sets of 4, cut at an angle(to capture neural activity at different depths) and attached to an electronic interface board (EIB;Neurotek, Toronto, Ontario). The EIB size was 5.3mm x 9.8mm and used an industry standardOmnetics connector that protruded vertically ≈5mm. The EIB was mounted on a makeshift driveconsisting of two screws attached to a 10mm x 3mm mounting board such that adjustment of thescrews allowed for partial advancement of the EIB (and tetrodes) into tissue even after the mountingboard was fixed in place. The total weight of the constructed tetrode was ≈1.2g. Following a small≈1.5mm craniotomy, the EIB (with tetrodes protruding ≈1.0mm past the mounting board lowersurface) was lowered into cortex using a micro-manipulator and the mounting board was cementedto the skull. Following surgery, the mouse was allowed to recover under a heat lamp and monitoreddaily for one week.Mouse awake acute electrophysiological recordings - external labOne additional awake visual cortex extracellular electrophysiology recording was obtained from aNtsr1-Cre mouse. The recording was carried out by Martin Spacek and Laura Busse, at Ludwig-Maximilians Universitat in Munich, using animal protocols and experimental paradigms previouslydescribed (Erisken et al., 2014). Briefly, mice were chronically implanted with a head-post whichallowed them to be placed on an air-suspended Styrofoam ball for habituation to head-fixation.Visual cortex recordings were subsequently made following habituation using a 32 channel linearsilicon probe (Neuronexus, A1x32-5mm-25-177-A32) and while the mouse viewed natural scenemovie stimuli identical to those presented to cats described above.Mouse sensory stimulationDifferent types of stimuli were used to confirm the correct insertion of the polytrode into sensorycortical areas for recordings presented in Chapter 5 (for recordings in Chapter 4, see Methodsbelow). To identify barrel cortex penetrations, a whisker was attached to a piezoelectric device13(Q220-A4-203YB, Piezo Systems, Inc., Woburn, MA) and stimulated using single 1ms (or 10ms)square pulse. To identify visual cortex, a 1ms LED pulse of green (or blue) light was used. Toidentify auditory cortex, a 1ms cross-frequency noise chirp was used (note: in some experiments,single tones and frequency sweeps were also used). Stimulus trials varied between 30 to 100 perrecording with inter-stimulus intervals of 3s to 10s. The location of retrosplenial cortex (RS)was inferred using Allen Mouse Brain Atlas (Lein et al., 2007) reported coordinates and relativecoordinates for auditory, visual and barrel cortex ROIs where available. For all analyses reportedin Chapters 4 and 5 the electrode insertion was confirmed using sensory stimuli as revealed in theelectrophysiological responses (LFP average, CSD average, and/or single unit response) were used.In vitro slice recordingsSlice recording (i.e. in vitro) data reported and analyzed in Chapter 2 came from in vitro ex-periments carried out previously and reported elsewhere Anastassiou et al., 2015. The data forthat study was collected by two authors: C. Anastassiou and R. Perin. All figures and analysis inChapter 2 were created by the author. Slice recordings were made at a lower temperature (∼20oC)and as a result little to no spontaneous spiking was observed (see Anastassiou et al., 2015). The invitro recordings thus contained lower noise, i.e. ∼6µV - 9µV (post 200Hz high-pass filtering) thanthe noise observed in some in vivo recordings, i.e. ∼9µV - 12µV reported herein (see Chapters 4and 5).Optical imagingCalcium imagingMouse optical imaging recordings used in Chapter 4 were carried out by Dongsheng Xiao andMatthieu Vanni. Mouse optical imaging recordings used in Chapter 5 were carried out by theauthor. Images of the cortical surface were recorded through a pair of front-to-front video lenses (50mm, 1.4 f:30 mm, 2 f) coupled to a 1M60 Pantera CCD camera (Dalsa) (Vanni and Murphy, 2014).To visualize the cortex, the surface of the brain was illuminated with green light (but not duringimage acquisition). Calcium indicators were excited with blue-light-emitting diodes (Luxeon, 470nm) with bandpass filters (467-499 nm). Green emission fluorescence was filtered using a 510-550 nmbandpass filter or collected in a multi-band mode as described below. For single wavelength greenepifluorescence 12-bit images at varying time resolution (20-100ms; i.e., 10-50 Hz) were collectedusing XCAP imaging software. In order to reduce file size and minimize the power of excitationlight used, camera pixels were binned (8 x 8) producing a resolution of ≈68µm/pixel. These imagingparameters have been used previously for voltage sensitive dye imaging (Mohajerani et al., 2013) aswell as anesthetized GCaMP3 imaging of spontaneous activity in mouse cortex (Vanni and Murphy,2014) and awake GCaMP6 imaging in mouse cortex with chronic window (Silasi et al., 2016).14RGB calcium imagingIn some experiments (see main text Chapter 4, Chapter 5), a multi-wavelength strategy was em-ployed to correct for potential green epifluorescence signals that were associated with non-calciumdependent events. The method was a variant of the spectral correction strategy described byothers (Ma et al., 2016; Wekselblatt et al., 2016) where changes in green reflected light near theisobestic point of hemoglobin are monitored and the calcium activity signal is accordingly corrected.The stratgey employed here was related to previous work using blue excitation and reflected light(Sirotin and Das, 2010). Assuming that hemoglobin is the primary absorber in brain tissue invivo, changes in blood volume or oxygenation affect both excitation and emission of light used forwide-field imaging (Ma et al., 2016). The strategy makes use of short blue wavelength referencelight that is also near a hemoglobin isosbestic point. While others have used a strobed LED pre-sentation with a subset of frames providing a green reflected light reference image (Ma et al., 2016;Wekselblatt et al., 2016), the approach was to take advantage of an RGB camera sensor to allowsimultaneous acquisition of a shorter wavelength blue ≈447nm signal that correlates strongly withgreen reflected light signals. This strategy provides a short blue light reference without the needfor strobing which can limit time resolution and potentially entrain some neuronal rhythms (Iac-carino et al., 2016) and is more technically demanding from a hardware synchronization standpoint.The strategy relies on the Raspberry Picams RGB sensor (Waveshare Electronics RPi Camera F)to independently resolve signals attributed to blood volume changes as blue reflected light, whilesimultaneously collecting green epi fluorescence (GCaMP6). A Chroma 69013m multi-band filter(10mm diameter) was mounted over the image sensor allowing blue, green, and red signals to besimultaneously obtained in separate channels of the camera’s RGB sensor with less than 10% crosstalk between channels. Two Luxeon LEDs were used: 1) Royal-Blue (447.5nm) LUXEON Rebel ESLED with added Brightline Semrock 438/24 nm filter to provide a short blue wavelength reflectedlight signal that is expected to report blood volume changes; and 2) a blue 473 nm Luxeon RebelES LED for excitation of GCaMP6 with a Chroma 480 nm/30 nm excitation filter. In prelimi-nary analyses the short wavelength blue signal correlated positively with apparent blood volumeartifacts that were revealed by parallel experiments using green reflected light imaging (r=0.93,see Fig 4.18). Given that the short-blue reflected light signals provided a surrogate indicator ofgreen reflected light (i.e. they are highly positively correlated) this signal was used in a ratiometriccorrection strategy. While the shorter blue wavelength light will scatter more than a green strobedreflected light signal used by others (Ma et al., 2016; Wekselblatt et al., 2016), analysis of greenreflected light and short blue reflected light indicated that two were highly correlated, suggestingthat the major artifacts observed were associated with large blood volume changes in superficialcortical layers.15Mouse VSD imagingVSD imaging was carried out as described previously (Mohajerani et al., 2010, Mohajerani et al.,2013; Vanni and Murphy, 2014). Briefly, either a unilateral craniotomy (1 wildtype C57/BL6 mouse;bregma 2.5mm to -4.5 mm, lateral 0mm to -6mm) or bilateral craniotomy (1 wildtype C57/BL6mouse; bregma 3.5mm to -5.5mm, lateral 4.5mm to -4.5mm) was made with the underlying duraremoved and RH1692 dye (Optical Imaging, New York, NY) (Shoham et al., 1999) dissolved inHEPES-buffered saline (0.621mg/ml) was added to cortex for 60-90min. VSD imaging began≈30min following washing of unbound VSD. VSD data (12 bit monochrome) was captured with6.67ms (150Hz) temporal resolution using a CCD camera (1M60 Pantera, Dalsa, Waterloo, ON)and EPIX E4DB frame grabber with XCAP 3.1 imaging software (EPIX, Inc., Buffalo Grove IL).Mouse sensory stimulation and ROI mappingFor optical imaging recordings used in Chapter 4 sensory stimuli were used to confirm sensorycortical and sub-cortical areas using forelimb, hindlimb, whisker and visual stimulation (for stimuliused in Chapter 5 see above). To stimulate the forelimbs and hindlimbs, thin acupuncture needles(0.14 mm) were inserted into the paws, and a 0.2-1 mA, 1 ms electrical pulse was delivered. Tostimulate a single whisker (C2), the whisker was attached to a piezoelectric device (Q220-A4-203YB,Piezo Systems, Inc., Woburn, MA) and given a single 1ms tap using a square pulse. The whisker wasmoved at most 90 µm in an anterior-to-posterior direction, which corresponds to a 2.6o deflection.A 1 ms pulse of combined green and blue light was delivered as visual stimulation. Averages ofsensory stimulation were calculated from 20-40 trials of stimulation with an inter-stimulus intervalof 10s.Multimodal recording in awake miceTo initiate wakefulness isoflurane and oxygen were stopped and the anesthesia mask was removed.Calcium imaging data were obtained over the following 1 hour. The body temperature of mice wasmaintained with a heating pad. Awake calcium imaging of spontaneous activity was performed inthe absence of visual and auditory stimulation. A behavioral monitoring camera was used to confirmthat mice were awake and relatively unstressed with grooming and whisking occasionally observed.An analgesic, buprenorphine, was injected (0.075 mg per kg body weight intraperitoneally) 2-4hbefore awake calcium recordings. A second Dalsa 1M60 camera (150 Hz) or Raspberry PicamsRGB sensor (60Hz) was used to capture body and whisker movements under infrared illumination.Awake mice movement correctionsWhile relatively few large body movements were observed during awake imaging sessions, theirimpact on mapping was evaluated by generating [Ca] activity maps from periods of quiescence. Toidentify periods of movement or quiescence, the standard deviation of luminance fluctuation was16calculated for each pixel. This approach showed that most of the movements were localized on thefacial (whisker and jaw) and forepaw regions. A region of interest was manually drawn for eachframe and the sum of absolute value of the gradient was calculated by subtracting each frame fromthe previous one within this region (see Fig 4.19). This gradient profile within the region of interestwas temporally smoothed at 0.1Hz and the median and standard deviation were calculated (σ).Periods of quiescence were identified as having a gradient lower than [median+σ/10], while periodsof movement were higher than [median+σ]. To more selectively identify periods of quiescenceisolated from any movement, an exclusion window of 10s was applied at the beginning and the endof each period of quiescence and only periods of more than 10s were used for comparison analysis.[Ca] activity maps (i.e. STMs; see main text) were then generated using spikes for periods ofquiescence and compared with maps using all spikes with little overall differences (see main text).HistologyFor some recordings, pipettes (Chapter 4) were filled with dye (Texas red-dextran) or the rearof a laminar electrode shank (Chapter 4 and 5) was painted with fluorescent 1, 1-dioctadecyl-3,3,3,3-tetramethylindocarbocyanine perchlorate (DiI, ≈10% in dimethylfuran, Molecular Probes,Eugene, OR). Dye-labeled pipettes and electrodes were not used until the dimethylfuran solventhad evaporated. At the end of each experiment, animals were killed with an intraperitoneal in-jection of pentobarbital (24 mg). Mice were transcardially perfused with PBS followed by chilled4% PFA in PBS. Some mice brins were sectioned (coronal slices) (50µm thickness) on a vibratome(Leica VT1000S). Images of diI labeling with counter-stained DAPI were acquired using confo-cal microscopy (Zeiss LSM510) to reveal the electrode track and help identify the approximatesubcortical location of recorded single units.Raspberry Pi imagingRaspberry Pi (RPi) + Pi Camera Module (Picam) imaging was employed to facilitate off-linehemodynamic corrections via the simultaneous acquisition of RGB data (see Methods above). TheRPi and Picam required custom Python3.4 code for tracking single frame times. The cutom codeis described in the Appendix D.Analysis and computational methodsCoding toolboxesExcept as otherwise noted in the main text or here, all analysis was carried out using customFortran or Python2.7 code developed as part of an electrophysiology and optical physiology toolkitavailable online (www.github.com/catubc/openneuron) and described in Appendix C. Event trig-gered analysis tools have been referenced in previous publications (Xiao et al., 2017) and are also17available online (www.github.com/catubc/picam; www.github.com/catubc/sta maps).Single unit spike sortingThe program SpikeSorter (Swindale and Spacek, 2014, Swindale et al., 2017 was used for all spike-sorting results presented herein. Briefly, mouse electrophysiological recordings were high pass-filtered at 1kHz (cat V1 recordings were already high-pass filtered), and single spikes were detectedusing a threshold of 4 to 5 times the median of the standard deviation of the absolute voltagevalues divided by 0.6475 (Quian Quiroga et al., 2004) and a dynamic-multiphasic event detec-tion method previously described (Swindale and Spacek, 2015). Sorting was carried out by anautomated method and followed by manual inspection of units. Units sorted in Chapter 4 had aminimum (max channel) peak-to-peak amplitude of 40µV. Units used in Chapter 5 had a minimumpeak-to-peak threshold of 75µV.Single unit and imaging SNR exclusion criteriaSingle unit spikesorting results presented in Chapter 4 were subject to multiple exclusion criteria:only units with a peak-to-peak extracellular amplitude of at least 40µV, a minimum of 200 spikes,and a calcium cortical response (i.e. dF/F0 peak) of at least 1% were used. The reasons for thesecriteria were to improve single unit sorting isolation due to the lower-density of the extracellularpolytrode used (i.e. 100µm vertical spacing, single column).Single neuron properties used for STMsSTMs included in analyses were generated from a minimum of 200 spikes. Only STMTDs exhibit-ing fluorescence exceeding 1% dF/F0 (0.5% in some cases) were used for analysis. Sample sizeswere not pre-determined but are consistent with previous experiments using similar methodology(Mohajerani et al., 2013; Vanni and Murphy, 2014; Chan et al., 2015).LFP-clustering recordings and unit yieldFor data used in Chapter 5, the average recording length was: 8.8±3.0hrs in cat recordings,3.6±1.6hrs in acute mouse recordings, and 2.8hrs in the mouse chronic recording. Neuron yieldper electrode track was: 99±38 in cat V1 recordings, 85±29 in acute mouse recordings, and 31neurons in the chronic mouse recording. Average number of spikes per sorted unit was: 31807 incat recordings, 11974 in mouse acute recordings, and 11190 in the chronic mouse recording. Neuronfiring rates had log-normal distributions with long tails and medians of: 0.31Hz in cat recordings(86% of neurons fired <2Hz); 0.26Hz in acute mouse recordings (89% of neurons fired <2Hz); and0.68Hz in the chronic mouse recording (87% of neurons fired < 2Hz). Lastly, although continuousrecordings of up to 7hrs (mice) or 14hrs (cats) were made, most of the analysis focuses either on thefirst 2-3 hours of a recording (mice) or 2-3 hours of continuous recording (cat) to limit complexities18arising from spike sorting drift, polytrode damage to cortex and overall animal health deterioration.Clustering LFP events into LECsThe approach was to convert 10-channel LFP data to an extracellular recording resembling a high-pass (i.e. spike) record and use SpikeSorter for event detection, alignment, feature-extraction, clus-tering and cleanup. Other spikesorting suites or other clustering methods (e.g. template matching)were not explored but they should be equally valid if the broad steps described herein are followed.After selecting 10 approximately equally-spaced channels of LFP from the recordings, the deepestrecording channel was used identify synchronized states: i.e. recording periods with a synchronyindex (Li et al., 2009; Saleem et al., 2010) greater than 0.5 (i.e. periods where most of LFPpower lies in the 0.1-4Hz band). In cat V1 recordings, synchronized state periods accounted for:2.68±1.48hrs of total recording periods of 8.81±3.02hrs, ranging from 4% to 86% of total recordingperiods (Table 5.1). In mouse sensory cortex recordings, synchronized state periods accounted for:2.7±1.46hrs of total recordings periods of 3.17±1.40hrs, ranging from 56% to 100% of the totalrecording periods (Table 5.2).The synchronized state LFP record was next high-passed filtered (at 4Hz) to remove slowerLFP fluctuations. It should be noted that even under anesthetic preparations cortical activity canbecome desynchronized for shorter (or longer) periods, e.g. contain many lower amplitude high-frequency activity. One option for removing such desynchronized periods is to concatenate theremaining synchronized periods together and perform spike sorting and LFP clustering on the con-catenated records. A simpler approach, implemented here, was to manually mask desynchronizedstate periods on the basis of the synchrony index. This preserved the order of the LFP recordingsrelative the high-pass (i.e. single neuron spiking) record and made it easier to compare LFP eventtimes with neuron spikes at later stages of analysis.The LFP record was then treated as if it was a spike recording by setting the sampling rateto a higher value (usually 50Khz) than the original 1Khz rate of the LFP record. This simulateda speeding speeding up of the LFP record 50 times so that LFP events that previously spanned50-100ms now had a duration of 1-2ms and could be treated as single spike events by the spike-sorting software (albeit with very large spatial extent). Standard sorting methods were employedstarting with event-detection using a threshold of 5 times the standard deviation (STD) of theLFP signal divided by 0.6745 (Quiroga et al 2004) using a detection filter with a temporal window(Swindale and Spacek, 2015) of 1ms (equivalent to 50ms realtime), a temporal lockout of 3ms(150ms realtime) and a 2mm spatial lockout (i.e. in a 150ms realtime period only a single LFPevent could be identified). The events were next aligned to their mean using a least-squares criterion(described in Swindale and Spacek, 2015), principal components were calculated and events wereseparated using PCA into clusters called LECs (LFP event classes; see Chapter 5 text). LFPrecordings had typically between 1-4 LEC types in cat V1, 3-4 LEC types in mouse visual cortex,and 1-2 LEC types in mouse barrel and auditory cortex (see Tables 5.3, 5.4). The LECs were19aligned using a centre-of-gravity measure (COG; Swindale and Spacek, 2015) which reduces RMSerror between each event and the average template (rather than aligning to trough or peak) andall event times could then be exported with millisecond precision.LEC stability computationThe internal stability of all LFP events in each LEC was determined by measuring the standarddeviation of the FWHM of the first trough in each event (on the maximum amplitude channel ofthe LEC template). While not all LECs have a strong peak, all have troughs and single neuronspiking generally occurs shortly following the first trough. The FWHM of each event trough wascomputed (see G) and the distributions revealed that the vast majority of cat V1 LECs had astandard deviation <10ms.CSD computationCurrent-Source-Density (CSD) profiles were computed by taking the 2nd spatial derivative of theLEC templates (i.e. averages of all events in each LEC) across all LFP channels. For cat V1recordings the entire 10-channels were used (as the polytrodes were completely inserted into cortex).For mouse recordings one column of the polytrode was selected (i.e. 32-channels) and the CSDs werecomputed on the subset of electrodes which were inserted into tissue (see above). An exclusioncriterion was applied for very low amplitude CSDs: the maximum (absolute) values for currentsources or sinks had to be 200A/m3 resulting in rejection of 2 of 36 LECs in cat V1 recordings andnone in mouse.Calculation of peri-LEC-event-triggered-histograms (PETH)LEC-triggered single neuron histograms (PETHs) were computed using previously described meth-ods (Luczak et al., 2007, Luczak et al., 2009, Luczak et al., 2013; Bermudez-Contreras et al., 2013;Luczak and Bartho, 2012). Each spike time was replaced by a Gaussian to enable a more bin-width-independent computation of histograms. The gaussian width was 10ms instead of 20ms usedin previous work as UP-state transition detection methods used herein were more precise than pre-viously reported. Histograms were then computed from all events in each LEC. Peak latency (PL)for each neuron was chosen as the peak time of each neuron’s LEC triggered PETH (as opposed tothe centre-of-mass used by previous methods which were less precise Luczak et al., 2007).Computation of dF/F0Raw imaging data were saved in a binary format and processed using Python2.7 code as previouslydescribed (Xiao et al., 2017). To compute dF/F0 values, F0 was first computed for each eventby averaging the value of every pixel during the 3s period preceding the window of interest forthe event (i.e. -6s to -3s; Chapter 5) or 3s period preceding the event (i.e. -3s to 0s; Chapter 6)20with both approaches yielding similar results. dF/F0 was then computed by subtracting F0 fromeach raw frame and dividing by F0. As previously described (Xiao et al., 2017), other methodsfor computing dF/F0 were tested including calculating F0 as the average of the entire recording orby first band pass filtering the data (0.1Hz to 6.0Hz) with results largely similar across methods.ROI time courses (i.e. STMTDs in Chapter 4; ROI traces in Chapter 5) were defined as eitherthe dF/F0 at a pixel (Chapter 4) or the average activity in an ROI area (see Chapter 5). Spike-triggered-average Maps (STM) were defined as the maximum response calculated for each pixelwithin a time window of ±1 second of t=0ms.Computing VSD STMsVSD STMs and motifs (i.e. spatio-temporal dynamics; see Chapter 4) were computed as describedabove for GCaMP6. To be noted is that even during strong sensory stimuli, VSD signals for theRH1692 voltage sensitive dye peak at ≈0.5% dF/F0. Thus, averaging over many VSD image framesresulted in STMs with dF/F0 peaks between ≈0.1% to ≈0.25%. Such STMs were nonetheless asaveraging random frames (i.e. activity not related to the spiking of a single neuron) results in STMswith peaks <=0.01%).Computation of epoch correlation coefficientsChanges in PLs over time were tracked by converting the PL order in each 30 minute epoch inton-dimensional (n = # of neurons) and computing the pairwise correlation coefficient against vectorsin subsequent epochs.Sliding window PETH computation and controlsSliding window PETHs were computed using rasters from 30-minute sliding epochs in 1 minuteincrements. In keeping with methods previously described (Luczak et al., 2007, 2009; Bermudez-Contreras et al., 2013) a control was computed by dividing rasters into different groups and recom-puting the PLs. However, in contrast to previous methods which divided all spikes (in an entirerecording) into first vs. second half, spikes from each recording were divided into even vs oddevents. Thus, single neuron spikes triggered by each LEC event (i.e. occuring within ±100ms ofeach LEC) were split into two groups taking each alternating event. The approach used here encom-passes previous approaches as by comparing all possible epochs (spaced in 1 minute increments) itcaptures first vs second half and many other possible combinations. However, it additionally tracksthe difference within each epoch between odd and even spikes. Given that PLs drift over time,it is not appropriate to implement the method applied in other studies. Because cortical neuronsrecorded spiked in tonic as well as burst patterns during UP-state transitions (not shown) othercontrol metrics were tested, for example, by dividing PLs (as opposed to individual spikes) for eachLECs into even/odd groups and recomputing the PLs using the window approach. The results of21these tests were similar to those presented using even/odd spike splitting.Computing precisely repeating spike triplets - triplet histogramsTriplet histograms were computed as previously described (Abeles, 1982a, Luczak et al., 2007).Briefly, using 10ms-wide bins the inter-spike-interval between three cells’ spikes were binned anddisplayed in a 2D graph where the x-axis represents the inter-spike-interval histograms of all spikesof the first and second neuron and the y-axis represents the inter-spike-interval histograms of allspikes of the second and third neuron (see also main text).Stable neuron heuristicsFor several sections in Chapter 5, recordings were chosen and only stable units were selected forfurther analysis. Stable neurons were qualitatively selected as less likely to contain spike sortingerrors. In particular, the selected neurons had extracellular templates with a peak-to-peak (maxchannel) amplitude 100µV. Additionally, the spiking rate for each neuron was qualitatively similaracross the entire recording period to ensure that analysis based on spikes from different parts ofthe recording would not be biased due to substantially different firing rates (or a lack of spikingaltogether).Removing UP-state-locked spikingFor the computation carried out in Chapter 5 Figure 75 the goal was to determine whether co-occurring spikes during natural scene stimuli were evoked by stimulus - as opposed by UP-statetransitions. Accordingly spiking during/near UP-state transitions was removed: for each LEC eventacross all LEC classes in the recording all single unit spikes that fell within a ±100ms window ofthe LEC time were removed. During the recording reported (see main text) the neurons consideredhad on average 7935 spikes and the procedure removed an average of 922 of their spikes.Spike-triggered-map (STM) temporal dynamics (STMTDs)For the analysis in Chapter 4, STMTDs were defined as the time course of activity of the maximally(or minimally) activated pixel in a region of interest. The ROI chosen was the left-hemispherebarrel cortex and tracking the activity of the maximally (or minimally) activated pixel resulted ina 1-dimensional trace of usually 180 data points (i.e. 6 sec x 30Hz). The STMTD traces for allneurons were analyzed together using PCA and the resulting clusters were separated using k-meansclustering algorithm (n=3). Putative cell classification into inhibitory and excitatory cell types wasbased on the full-width-half-max of each unit’s positive and negative phases (Connors and Gutnick,1990; Pape and McCormick, 1995).22Computation of seed pixel correlation maps (SPM)SPMs were computed by subtracting the contribution of global and illumination fluctuations fromthe signal of each pixel (Vanni and Murphy, 2014), also known as global-signal-regression. Thespontaneous activity recording sequences were then temporally band-pass filtered (0.3-3Hz). Cross-correlation coefficient r values between the temporal profiles of one selected pixel and all the otherswere calculated (White et al., 2011; Mohajerani et al., 2013; Vanni and Murphy, 2014). Similaritiesbetween STM and SPM maps were performed by measuring the r-value Pearson correlation coeffi-cient between each pair of pixels. To compare the STM with anatomical database, brain imagingstacks of 140 slices were downloaded from the Allen Mouse Brain Connectivity Atlas providingAAV-virus tracing database (http://connectivity.brain-map.org/, (Oh et al., 2014). For each slice,the first dorsal 300 µm of brain fluorescence in the z-axis (i.e. depth) were summed to generatepartial maximum z-projection maps similarly to previous studies (Mohajerani et al., 2013).DeconvolutionPixel-wise calcium imaging deconvolution was done using a method described (Pnevmatikakis etal., 2016) and the code provided by the authors on Github (https://github.com/epnev/ca sourceextraction). Briefly, the method uses an autoregressive approach to estimate the calcium transientas an impulse response from the data itself (Pnevmatikakis et al., 2016). Using this approachthe time course of calcium transients is deconvolved via efficient non-negative, sparse, constraineddeconvolution. As explained in the main text (Fig 4.3) deconvolution did not substantially changethe STMs otherwise obtained.Cell spiking mode determinationGiven that single neurons can fire in both burst and tonic modes, it was determined whetherthese different modes could yield substantially different STMs. The methodology employed waspreviously described for defining main spiking modes of a thalamic neuron (Sherman and Guillery,2006, Fig 6.5, pg 236). Briefly, the method requires determination of the distribution of each spike’sinter-spike-interval (ISI) between the previous spikes (always positive; x-axis) and following spike(always negative; y-axis). The distribution is then plotted using logarithmic scales and clustered(Fig 4.11). Spike groups occurring in approximately each quadrant indicate different spiking modes:first spikes in a burst (bottom right), spiking occurring during a burst (bottom left), last spikesin a burst (top left), and tonic spikes (top right). The vast majority of cortical cells recorded inbarrel cortex did not exhibit multiple classes of spiking modes and only a few of thalamic cellsrecorded showed clear bursting modes while also passing minimum thresholds (see Methods: Singleunit activity analysis).23Single-spike STM-space analysisAs the STM method averages over the STMs for all spikes from a single neuron, it was importantto determine whether single spike STMs could be naturally clustered by similarity. The overall aimof the method was thus to group single spike STMs by similarity in a high-dimensional space and todetermine whether natural clusters occur or whether spontaneous activity could be separated frommanually partitioned sub-grouping of STMs. First, each 256 x 256 STM was subsampled to 64 x64 pixels and converted to 4096-Dimensional vector. The next step was to compute distributions ofthe STMs of all spikes (which have high variability) and evaluate them in a high dimensional space(or a reduced space using PCA). The lack of obvious clusters (in 2D) indicated that there wereno sub-groups of STMs present in the data which are removed by the STM averaging procedure.Next, STMs were grouped by similarity into 4 (or more) partitions to reveal active sub-networkspresent during single spiking. While inter-spike-interval (ISI) distribution were largely similar forthe sub-grouped networks, the resulting 4 sub-networks had substantial diversity indicating that (onaverage) spiking occurred during different types of active cortical networks with only one of thesesub-networks resembling the all-spike average STM (see Chapter 6, Fig 4.14D, 4 sub-network STMsand sum at the bottom). Next, spontaneous STMs - i.e. STMs obtained from all spontaneous datawithout consideration of spiking - were converted to an STM-space representation. The spontaneousSTMs were grouped into 4 sub-networks (similar to spike triggered STM partition) by re-using thespike generated sub-network centres (Fig 4.14B,F). This ensures that the spontaneous sub-networksare similar to the spike-triggered sub-networks. The resulting spontaneous sub-networks are similar- but not identical - to the cell spike triggered sub-networks (Fig 4.14H; note that the sum ismostly noise as expected when summing over all activity). Importantly, when subtracting thespontaneously active sub-network STMs (Fig 4.14H) from the spike-triggered sub-network STMs(Fig 4.14D), the results yield STMs which represented mainly single cell spiking components andare very similar to the overall average STM.Purity and completeness spike-sorting metricsIn addition to reporting false positives (FP - incorrectly assigned spikes in a single unit cluster)and false negatives (FN - missed spikes in a single unit cluster) for spike sorting data, two metrics- purity and completeness - were also developed which better describe spike-sorting error givenground-truth data from multiple neurons. There are very few experimental studies capturingintracellular and extracellular data simultaneously and almost none where two or more neuronsare recorded from simultaneously. One in vivo anesthetized rat hippocampus study successfullypatched 3 hippocampal cells simultaneously with extracellular waveforms (Henze et al., 2000; Note:3 cells were obtained from ≈ 30 animals). Because it only considered single neurons at a time, thisstudy quantified sorting errors using the traditional confusion matrix approach, i.e. characterizingfalse positives (e.g. spikes in a unit not belonging to the intracellularly recorded neuron) and false24negatives (e.g. spikes from a cell that were missed by the sorting algorithm). When consideringmore than one cell, FP and FN rates are not as useful because errors don’t include just missedspikes or sorted noise, but more complex over-splitting and under-splitting errors which traditionalconfusion matrices do not capture.The purity of a sorted unit Ui is defined as the largest number of spikes in the unit that comefrom a single ground truth neuron divided by the total number of spikes in that unit. For n neurons,the purity of a sorted unit Ui is thus computed by searching over all cells Cj for the maximummatch (i.e. intersection) of spikes:purityi =max|Ui ∩ Cj ||Ui|, 1 ≤ j ≤ n (1)The purity metric properly takes into account errors from under-splitting units, i.e. presence ofspikes from multiple cells in a unit, while also providing a measure of overall error relative tothe cell where most of the spikes come from. However, this is not sufficient to capture all theimportant information given ground truth spiking rasters. For example, a unit with 100 spikes,90 of which come from a unique neuron that fires 900 spikes has 90% purity, but only captures10% of the overall activity of that neuron. The completeness of a sorted unit is thus introducedto additionally capture how much of the overall spiking of the principally identified neuron wascaptured by the sorted unit. The completeness of a sorted unit Ui is defined as the largest numberof spikes in the unit that come from a single ground truth neuron divided by the total number ofspikes in the cell.completenessi =max|Ui ∩ Cj ||Cj |, 1 ≤ j ≤ n (2)The completeness metric thus accounts for how ”complete” a sorted unit captures the spikingactivity of its best matching cell. For a perfectly sorted recording, both purity and completenessmetrics = 1 for all sorted units.Hyper-angle computationThe angle between two N-dimensional vectors u and v is:θ = cos−1(~u · ~v|~u||~v|)(3)False discovery rate correction - Benjamini-HochbergThe false discovery rate (FDR) arises as a type of error from consideration of multiple compar-isons. For example, for a significance p-value set at 0.05%, on average, one out of 20 comparisonswill yield a statistically significant result regardless of the data types and distributions (includ-25ing noise). This error was corrected using Benjamini-Hochberg (BH) implementation in Python(statsmodels.sandbox.stats.multicomp.multipletests using the fdr_bh method). In con-trast to the Bonferroni correction which controls the overall family-wise error rate by dividing thep values of all comparisons by the # of tests (and is very conservative), BH controls for the falsediscovery rate, i.e. the ratio of false positives among the data considered. For the example providedabove, BH controls for the one false positive value that would occur in 20 tests for an FDR (p value)set to 0.05%.26Spike Sorting Algorithm TestingUsing In Vitro and In Silico DatasetsElectrophysiology, i.e. measuring the electrical activity of a cell or population of cells, is theoldest and arguably still the most common method for investigating neural activity in the nervoussystem. Most of the results in this thesis rely on tracking single neuron spiking activity as recordedin extracellular cortical and subcortical areas. Isolating and grouping spiking activity of neuronsfrom extracellular recordings is done using spike-sorting: i.e. methods for assigning spikes or eventsto individual neurons (also known as units). Yet, while high-density electrodes and spike-sortingalgorithms have been available for many years, adequate experimental or simulated ground truthdatasets to test and validate spike sorting algorithms have been limited. The limitations ariselargely due to the difficulty in capturing true spiking activity intracellularly (e.g. using patchclamping methods) while also recording extracellularly from very close by (i.e. <50-100µm). Yet,it is important to quantify how good spike sorting methods are using independently validateddatasets.Accordingly, this chapter is dedicated to generating in vitro and in silico datasets for testingspikesorting algorithms and confirming that the methods (and in particular SpikeSorter used herein,Swindale and Spacek, 2014) are adequate for future research. Biophysically simulated datasets canprovide ground truth spiking activity for many neurons (1000s or more) while also generating extra-cellular voltage records that can be spike sorted. While simulations have some inherent limitations(e.g. cannot yet fully emulate extracellular space or electrode damage to tissue), single neuron andnetwork models have become sufficiently accurate to provide a very good approximation of in vivoeextracellular recordings (see text below).Accordingly, novel in vitro (i.e. cortical slice) and in silico (i.e. simulated) datasets are providedand analyzed to assess the general performance of sorting methods, as well as different sorting suitesand multiple operators, and to assess the role of multi-channel probe configurations on spike sortingquality for electrodes similar to those used in Chapters 4 and 5. Simulated datasets were sorted byup to 7 different operators using 4 different spike sorting suites including SpikeSorter, KlustaKwik(Kadir et al., 2014), SpikingCircus (Yger et al., 2016) and an earlier development version of Kilosort(Pachitariu et al., 2016). Lastly, an ongoing cloud-based effort is presented for providing ground-truth datasets to the spike sorting development community. The general findings of this chapterare that: (i) simulated datasets are statistically similar to in vivo and in vitro recorded data, i.e.they have similar extracellular voltage signal standard deviations and similar neuron yield to invivo recordings; (ii) operator skill may be more important than spike sorting suite algorithms andfeatures; (iii) higher density probes provide much lower error rates; and (iv) with respect to single27neuron sorting quality, channel spacings <20µm may not provide better unit yield or sorting qualityfor neuron soma diameters of 20-30µm (or larger).Perhaps most relevant for this thesis, the spike sorting software used here, i.e. SpikeSorter(Swindale and Spacek, 2014), provides sorting results that are quantitatively as good as othersorting software suites.Results presented in this chapter from in vitro and in silico biophysically detailed simulationshave been published or have been presented at conferences: Mitelut et al., 2014; Mitelut et al.,2015; Hawrylycz et al., 2016; Vyas et al., 2016; Jun et al., 2017a; Jun et al., 2017b; Gratiy et al.,2017. Unless otherwise noted, all analysis, figures and simulations in this chapter were preparedby the author (the exception of Fig 3.18, which is adapted from Jun, Mitelut et al., 2017a).Electrophysiology and single unit spike sortingWhile extracellular recordings have been made for many years, it continues to be unclear what thetotal extent of the LFP is (see Introduction; also Katzner et al., 2009; Yoshinao and Charles, 2011;Buzsa´ki et al., 2012). What is better understood, however, is that extracellular recordings capturethe somatic action potential (AP) of nearby (<50µm) single neurons, i.e. the transmembranecurrent activity occuring in the soma, axon hillock and initial segment of a neuron during an actionpotential. More importantly, somatic action potentials of nearby neurons have sufficiently largevoltage amplitudes and can be separated and analyzed using a process called ”spike sorting” whichuses signal processing techniques to assign each spike to a set of unique cells (or units; Lewicki,1998). The term ”unit” is used to refer to neurons isolated from extracellular recordings as in theabsence of ground-truth confirmation, the identity of the isolated neuron is only putative. Thereare many spike-sorting approaches aimed at sorting recordings from high-density electrodes (foran older review see Lewicki, 1998; Lewicki, 1994, Gray et al., 1995, Fee et al., 1996, Zouridakisand Tam, 1997, Harris et al., 2000, Zouridakis and Tam, 2000, Hulata et al., 2002, Nguyen etal., 2003, Shoham2003, Quian Quiroga et al., 2004, Litke et al., 2004, Pouzat et al., 2004, Hazanet al., 2006, Bar-Hillel et al., 2006, Wood and Black, 2008, Wolf and Burdick, 2009, Gasthauset al., 2009, Calabrese and Paninski, 2011, Swindale and Spacek, 2014, Pachitariu et al., 2016,Yger et al., 2016, Jun et al., 2017a). As spike-sorting algorithm development can be a challengingendeavor with several different approaches employed, it is not reviewed at length here. An exampleis provided, however, showing the steps involved in one recently developed spike sorting approach(Swindale and Spacek, 2014). There are several steps including: identifying events, separatingevents from noise; clustering the events using principal-component-analysis (PCA); and mergingand cleaning of clusters using both automated and supervised methods (see Fig 3.1 as an exampleof steps involved).Remarkably, despite the central importance of spike-sorting to single neuron electrophysiology,adequate ground-truth datasets to test spike-sorting algorithms are very limited (see also below)28Figure 1: Steps involved in spike-sorting. Steps involved in one approach to spike sorting.(Adapted from Swindale and Spacek, 2014 with permission).with most sorting suites relying on repeated sorts from the same in vivo recording (e.g. units aresorted and added back into the extracellular record for resorting - leading to potentially biasedresults).There are several challenges to obtaining ground-truth dual-recording datasets in vivo whichrequire a single neuron to be patched (or otherwise recorded directly) while a nearby (<50µm)extracellular probe captures its activity. First, despite many advancements in single neuron patchclamping (Sakmann and Neher, 1984) or sharp electrode recording methods, it remains very chal-lenging to make even single neuron recordings in vivo (i.e. without a nearby extracellular electrode)for extended periods of time. Accordingly, the few simultaneous patch clamp (or juxtacellular Netoet al., 2016) and extracellular recordings datasets rely on cortical slices and usually capture onlya single neuron at a time for < 1 hour. Second, and most challenging, is targeting the single neu-ron patch clamp very near (i.e. <50-100µm) from an extracellular electrode. This is challengingbecause advancing the patch pipette is usually done without visual information (i.e. blindly) andsingle neurons are not easy to specifically target, even in slice, let alone when they must be veryclose to another extracellular probe. This results in at best in vitro dual-recording datasets thatcontain at most 1 neuron patched at one time (though most attempts fail, e.g. see discussion in29Anastassiou et al., 2015). Additionally, dual-recordings in vitro result in lower background activitythan observed in vivo and very little to no spontaneous activity in the patched neuron or othernearby cells (see Anastassiou et al., 2015).In this chapter the lack of ground-truth datasets is addressed through the use of novel in vitroand simulated datasets. The novel in vitro datasets come from single neuron recordings and areused for testing both single neuron and quasi-multi neuron recordings (i.e. multi-neuron recordingsgenerated by concatenating single neuron recordings). The simulated datasets are obtained usingthe Blue Brain Project (BBP) single neuron models and the Allen Institute for Brain Sciencenetwork models and provide extensive ground-truth spiking datasets are assessed against severalsorting suites and multiple operators. Additionally, the role of polytrode channel layouts (i.e.electrode spacing) on spike sorting quality is briefly investigated in simulated datasets.In the sections below 7 spike detection and sorting tests are applied to multiple ground truthdatasets starting with in vitro recordings where spike detection, minimum neuron separation dis-tance and multi-operator sorting tests. Next, biophysically detailed multi-compartment neuron andnetwork models are discussed and novel in silico, i.e. simulated, datasets are generated and sortedby multiple operators and across multiple electrode configurations. The chapter ends with a briefdiscussion of the realism of simulated datasets and a cloud-based extracellular sorting applicationcurrently being developed.Using a novel in vitro extracellular-intracellular dataset for spikesorting validationExisting spike sorting ”ground-truth” datasetsAs mentioned above, because of the importance of precisely identifying single neuron spikes to manyfields of neuroscience, there have been significant efforts to develop high-density electrodes and spikesorting solutions. Yet, there are very few ground truth datasets for evaluating the performance ofsorting algorithms or that test spike sorting on high-density electrode layouts (e.g. Jun et al.,2017b).An older study (Henze et al., 2000) recorded several rat hippocampus neurons using intracellularand bundled wire extracellular tetrodes. That study resulted in ∼3 neurons sufficiently isolated(out of 30 rat experiments) and the results were limited to analysis of spike sorting quality in singleneuron preparations. Another substantial limitation of the study was the use of wire tetrodeswhich, although still used by some labs, are being replaced by high-density electrodes which havedozens to hundreds of channels.A more recent in vivo study recorded single neurons using juxta-cellular electrode and high-density extracellular probes (Neto et al., 2016). While the probes were high-density, due to theexperimental challenges, out of the several dozen recorded neurons, only 1 neuron had sufficiently30large PTP amplitude (i.e. >100µV) to be useful for high density sorting algorithm testing (e.g. seecomments in Jun et al., 2017a).Another recent effort used simulated datasets consisting of up to 10 biophysically detailedneurons (from up to three unique neuron morphologies) connected to other single compartmentcells with LFP-like noise being added from in vivo recordings (Hagen et al., 2015). While thesesimulated recordings can generate a lot more data, they are limited in the variability of theirneuronal morphology and electrical behaviour.Additionally, only a few neurons are captured for each simulated polytrode at one time, syn-chronous spiking is largely absent (i.e. neuron connectivity was not implemented), and the methodcannot be expanded further (i.e. the number of neurons cannot be increased; private communicationwith E. Hagen).Dataset acquisitionData used in this section comes from in vitro experiments carried out previously for a previousstudy (Anastassiou et al., 2015) aimed at evaluating extracellular spike waveform variability. Thedata for that study was collected by C. Anastassiou and R. Perin. All figures and analysis in thissection were created by the author.Simultaneous extracellular and intracellular recordings of single neuron activity continues to bea very challenging endeavor. The findings reported herein use data from 10 individually whole-cellpatched excitatory neurons which were collected over ∼1 year of experimental efforts on a 12-patch pipette system (Perin and Markram, 2013). The experimental methods for obtaining theserecordings are described in detail elsewhere (Anastassiou et al., 2015). Briefly, recordings in ratsomatosensory cortex slices were carried out using whole cell patching with extracelluar electrodesinserted nearby (i.e. ∼50µm; Fig 3.2). Extracellular potentials were recorded an H32 extracellularprobe containing 4-shanks and 8-channel per shank (NeuroNexus, Ann Arbor, MI). Each shankhad eight recording sites (160 µm2 per site, 1-3 MΩ impedance) and inter-shank distance was 200µm. Recording sites were staggered to provide a two-dimensional arrangement (≈30µm verticalstaggered separation; Fig 3.2A). Both excitatory and inhibitory neurons were targeted and in onecase two neurons were simultaneously patched. Recordings lasted ≈1hour each.The patched cells were stimulated using constant and ramping direct-currents (Fig 3.3) as wellas alternating-currents (not analyzed here). Extracellular spike peak-to-peak (PTP) amplitudesranged from ∼40µV to >125µV. Under strong current stimulation some neurons fired repeatedspikes and the extracellular PTP amplitudes could decrease by up to ∼50%.Results of sorting in vitro datasetsSpike detection and sorting methods were applied to the original in vitro data in three tests: an eventdetection-only test (Test #1), a spike sorting test for minimum-distance spatial separation (Test3130μmFigure 2: Simultaneous extracellular-intracellular in vitro experiments. A. Schematicof H32 extracellular probe used in the study. B. Simultaneous dual-cell patching and extracellularrecording (Adapted from Mitelut et al., 2014 with permission).#2) and a full spike-sorting test on all spikes from all cells (>27,000 total spikes) for three differentoperators (Test #3). Two recorded neurons from this dataset are shown in Fig 3.3 including theirintracellular traces, their intracellular and extracellular PTP amplitudes and their extracellulartemplates.A couple of important notes should be made. First in the absence of current driving, there wasno observable spiking in the cold slices in these experiments (see Anastassiou et al., 2015). Also allspikesorting results (e.g. Test 2 and 3) were obtained following completed spikesorting of the databy a human operator and thus contain subjective splitting and merging decisions.Test no. 1The first test using in vitro datasets was aimed at qualitatively characterizing how lower eventdetection thresholds affect the percentage of correctly identified spikes from a cell while increasingthe False Positive (FP) rate. FP was defined as the number of incorrectly detected events persecond per electrode (another way to normalize is to divide by the SNR or standard deviationof the electrophysiological signal, see discussion below). However, this first test used FP rate asdefined in this more limited way in an attempt to broadly quantify the relative error rate differencesacross units of different PTP amplitudes as a function of event detection threshold. Accordingly,for this event detection-only test, no spike-sorting was done and only an event-detection algorithmwas run in order to see how the ratio of neuron events vs. noise events increase for lower thresholds.The test was also performed on each neuron’s extracellular record without mixing data from otherneurons.Using all extracellular spikes from 9 (selected) cells (Fig 3.4) and a threshold of 3.2 (see above,32Figure 3: Characterizing extracellular/intracellular datasets. A. Example of patchedneuron membrane voltage during multiple DC ramp current injections. B. Intracellular peak-to-peak amplitudes of neuron in (A) computed using the second derivative of Vm reveals within ramp(≈1.4min duration) spike-amplitude adaptation and longer time-scale adaptation (i.e. following10 current ramps). C. Extracellular peak-to-peak amplitude showing extracellular spike-amplitudeadaptation. D. 50-extracellular spikes aligned to maximum amplitude channel on single shank ofH32 probe. E-H. Same as A-D for a different cell with lower PTP amplitude. (Note: cell id numberscorrespond to original recording date and were left in for reference purposes).33also Methods) resulted in spikes from the largest PTP amplitude neurons (n=5) being detectedvery cleanly: no events from other units were present in those detected units (Fig 3.4B points atbottom of graph, i.e. FP=0). This threshold value was selected specifically as the approximatevalue FP rates began to rise significantly. As expected, decreasing the detection threshold (Fig3.4C-E) results in additional spikes being detected, for example, spikes detected for the light-blueunit (46.0µV PTP amplitude) increase from ∼20% (3.2 threshold) ∼30% (3.0 threshold). But thecost of those additionally detected events is that there error rate went from ∼0 (3.2 threshold) to∼30 events per minute of recording (3.0 threshold). Importantly, as the threshold values drop thereis a substantial increase error rates for all neurons including (up to hundreds of incorrect eventsper minute of recording).A few additional points should also be made. While these FP rates are provided pre-clustering(i.e. the detected events have not been clustered and cleaned; see spike sorting section above), theynonetheless warn of an uneven trade-off where increasing the number of neuronal events (i.e. spikes)passing threshold may lead to exponential increases in noise for clustering stages. While some of thenoise (i.e. background activity or spikes from other neurons) may be removed by subsequent sortingstages, there is no (analytical) limit on how much noise can be removed by a sorting process. Thisis an expected result due the (gaussian) distribution of extracellular potential values (not shown)and these elementary tests characterize this distribution using real tissue recorded data in a spikesorting toolbox using real event detection methods. Second, decreasing thresholds adds potentialerrors (i.e. increases FP rate) even for high-amplitude units (e.g. red, orange units in Fig 3.4).Lastly, it should be noted that all in vitro recordings were done in relatively cold slices (i.e. ∼20oC)and there was largely no spiking observed in the record other than the single patched neuron. Thisis not the case in vivo where a single electrode can detect a few nearby cells (though electrodecontact resistance can affect this number). Additionally, in cold slice, due to de-afferentiated anddisconnected tissue, the standard deviation of the signal was lower than in vivo and possibly skewedsome of the results towards higher-error rates due for such in vitro preparations.In sum, spike detection varied as expected (i.e. lowering threshold values increased errors)with a strongly nonlinear increase in FP rates near a specific threshold (3.2 here; note: for otherrecordings, this thresholding value may depend on the STD of the original raw voltage signalsand the latest version of SpikeSorter uses an automated thresholding method based on voltagedistributions on each channel).Test no. 2The second test carried out was aimed at arguably the most elementary question that can beasked of a spike sorting algorithm: what is the closest spacing of two cells that still allows theirspikes to be separated using spike-sorting methods? The ideal experimental data for investigatingsuch a question would come from simultaneously patched, nearly electrically and morphologicallyidentical, physically neighbouring neurons while also recording their extracellular potentials using34Figure 4: Detection threshold vs. false positive rates. A. Peak-to-peak extracelluar po-tential amplitude (on the maximum amplitude channel) of 11 in vitro recorded single cells. B.False-positive rate (FP; computed as # of incorrectly clustered events per second per electrode)vs. percent correct spikes detected for an event-detection run using a threshold of 3.2 (see maintext) showing very low FP rate for higher amplitude units. C-E. Same as B for lower detectionthresholds (Adapted from Mitelut et al., 2014 with permission).a high-density electrode <50µm away. Not only that, but such an experiment would need tobe reproduced for many different extracellular electrode penetrations to capture different relativelocations of the two neurons in the extracellular record (i.e. different angles between the neuronsand the face of the electrode). Unfortunately such data is not only not available but may not beobtainable for a long time given the current state of art of single neuron patching.Given that such an ideal test is not possible, a modified approach using available in vitrodata was implemented herein where a hybrid spike recording was made using an original singleneuron record plus a spatially shifted spiking record based on a linear interpolation procedure. Theadded neuron spikes essentially simulate an extracellular space shift in the 2-dimensional planeperpendicular to the extracellular probe. In particular, voltage values from spikes from a single35neuron were linearly interpolated (only in the vertical, i.e. y-direction) to simulate spikes from aneuron at a slightly shifted location on the probe (Fig 3.5A). For example, spatially interpolatinga voltage value on channel c at time t, 10µm towards channel c-1 would require an interpolation:V ct ⇒ Vc−1t1030+ V ct2030(4)The linear-interpolation method was implemented for each neuron by first identifying 50 samplepoints (@20Khz sampling rate) for every spike’s multi-channel extracellular waveform. Each spike’swaveform was then interpolated using equation (3.1) and then added back - at a random, non-overlapping location - to the original extracellular recording. The final extracellular record thuscontained the original neuron with n spikes and the shifted neuron record also with n spikes for atotal of 2n spikes. This shifted-spiking record poses a more difficult sorting task: both ”neurons”to be sorted have very similar extracellular spike waveforms (but for the minor scaling) which isnot expected of in vitro or in vivo recordings making the recording more challenging for sortingalgorithms to analyze.Using spatial shifts of 15µm, 10µm and 5µm all spikes from 8 (selected) neurons were interpolatedusing the method described above (Fig 3.5). However, the approach was slightly different from Test#1 above. First, only neurons that could be well isolated after the maximum shift condition, i.e.15µm, were chosen: i.e. neurons that had >50% correct spike identification. The reason was thatneurons with higher errors would be very poorly sorted for smaller spike shift conditions (and notuseful herein). Second, the goal of the test was not to identify the number of neurons as those werefixed at 2 for each dataset. Rather, the goal of the test was to detect as many of spikes as possibleand separate them into two clear units (i.e. clusters). All tests were carried by the author andinvolved full spike-sorting, i.e. event detection + clustering + operator cleanup. The FP value wasan average of the FP rate for both sorted neurons (i.e. the average of the percentage of incorrectlyidentified spikes in each of the two units) and an average of the true positive (TP) rate (i.e. theaverage of the percentage of correctly identified spikes in each of the two units).The results show that for a simulated vertical spatial displacement of 15µm, even pairs ofneurons with a PTP amplitude of ∼45µV can be relatively well separated from each other (Fig3.5B). This is a promising result suggesting that neurons that are vertically aligned (i.e. one ontop of the other) relative to the face of a high density extracellular probe can be well isolated.However, decreasing the spatial separation to 10µm substantially increases error in sorting of pairsof neurons with PTP <50µV (Fig 3.5C) and one neuron with slighly larger PTP (Fig 3.5C - cyancolour, 56.6µV PTP) which had a more limited extracellular waveform making it more challengingto sort to begin with. Decreasing the spatial separation to 5µm increases the false positive rate aswell decreasing the percentage spikes correctly identified for most neurons (except the largest PTPamplitude neuron) making most spike sorted data essentially unusable for analysis purposes.These findings are encouraging as even in the extreme, i.e. 5µm simulated shift, case the36neuron with PTP value of >110µV was relatively well isolated. This provides general support forthe stability and isolation for neurons with PTP values >110µV even when very similar neuronsnearby provide confounding spikes. However, while somata cannot overlap in space, given a nearbyextracellular probe insertion pairs of neurons can have virtually any relative position (i.e. even<5µm) to the face of the probe. That is, pairs of neighbouring neurons are not generally alignedperpendicularly to the face of the electrode: i.e. their somata will usually not be on top of each ofeach other and at the exact same distance from the face of the electrode. Thus, the distance of pairsof neurons to the electrode can be arbitrarily along a given axis, and certainly <5µm. This is asignificant caveat not only to this pair-wise study presented here, but the very concept of pair-wiseneuron separation and reflects the real complexity of properly testing spike sorting algorithms usingreal patch clamp data.These considerations caveat the results presented here and suggest other, perhaps more com-prehensive studies including in silico recordings (see next Section) are required to address thecomplexity of what is likely the most fundamental spike sorting test that can be carried out.Test no. 3In the last test using in vitro data, three operators sorted the combined (i.e. randomly con-catenated) extracellular spiking data from 10 neurons (Fig 3.6). For this test, the multi-channelwaveforms for spikes from each neuron were extracted from the original (single neuron) record as500-1000 sample points (i.e. 10-20ms sections) and randomly concatenated with other neurons andrandom length noise from the original recordings. This resulted in a final recording that containedall spikes from the original 10 neurons in random temporal order. To note is that although someneurons were located on the same electrodes, they did not temporally overlap (i.e. no two spikeswere added together at a point in time).The error rates are reported as purity and completeness metrics (see Methods). Briefly,purity identifies the proportion of spikes in a sorted unit that come from the principally identifiedneuron (i.e. most number of spike matches), and completeness reflects the proportion of spikes fromthe identified principal neuron that are captured in the sorted unit. Thus, the ideal sort shouldhave large purity and high completeness values.Neurons in this in vitro dataset had spiking rates of 3Hz-12Hz (Fig 3.6A). Clustering of theextracellular spikes using PCA in SpikeSorter (following event detection and feature selection)showed in many cases clear grouping (see Fig 3.6C). Three operators of varying skill (Fig 3.6D:grey: ≈1 year sorting experience, green: 5 years sorting experience, and blue: 10+ years of sortingexperience), sorted the data using SpikeSorter (Swindale and Spacek, 2014: Fig 3.6D-F - grey andblue) and KlustaKwik (Kadir et al., 2014: Fig 3.6D-F - green) sorting suites (Fig 3.6D-F). Thepurity values were high for all sorts ranging from 88 to 93%. This meant that on average, each sortedunit contained between 88% to 93% of spikes coming from a single neuron. Completeness ratesvaried substantially more: 53-70% indicating that on average each sorted unit captured between37Figure 5: Spatial separation test. A. Conceptual single neuron separation showing extracellularwaveforms from a single neuron in the original and vertically shifted condition. B. 15µm cellseparation test showing the % spikes correctly identified (i.e. true positive, TP) vs. false positive% for 8 neurons (see main text also). C,D. Same as B for different spatial separation tests. (Adaptedfrom Mitelut et al., 2014 with permission).3853% to 70% of its principal assigned neuron’s spiking activity. Considering the number of unitssorted (Fig 3.6E-left) there is also variability: one operator found all 10 units whereas the other two(blue and green) only found 8 units further suggesting that those operators did not set sufficientlylow thresholds. Thus investigating multiple threshold settings may be required as part of spikesorting strategies. However, it is also important to note that the grey operator (i.e. the author)was aware of the presence of 10 units and was therefore biased during sorting by this extraneousinformation. Interestingly, however, despite finding 20% less units (i.e. 8 instead of the original10), one operator (green) identified >17,000 spikes whereas the grey operator found all 10 units butonly identified ∼16,500 spikes. It is unlikely that this was a chance result as the green operator’sresults were also very good when considering other tests including multiple operators and spikesorting tests using simulated data (see below) suggesting the green operator’s skill was the mainreason for these higher sorting metrics.Finally, a single value metric (SVM) was used to reduce the overall sorting quality to one value.The SVM was defined as the sum over the product of purity and completeness values for all unitsi :Σi(purityi ∗ completenessi) (5)The SVM revealed a ranking that was as expected: the grey operator who identified all 10 unitshad the greatest overall SVM metric, followed by the green operator who identified more spikesthan everyone else with the blue operator being ranked last.SummaryAlthough very challenging to obtain and thus rarely available, simultaneous extracellular-intracellularin vitro datasets allow for the elementary testing of spike-sorting detection and sorting methods.Applying spike detection and sorting algorithms to a unique 10 cell dataset (Anastassiou et al.,2015) several novel findings were made:• false-positive rates increase with decreasing spike detection threshold values which suggestsattempting to sort lower-amplitude units becomes significantly more challenging to mixing ofother units’ spikes and background noise;• while higher amplitude neurons (e.g. >50µV PTP amplitude) can be separated from bio-physically similar and nearby (i.e. ≥10µm distance) neurons, neurons that are farther fromthe electrode (and pairs of neurons not lying perpendicular to the face of the probe) are likelychallenging to sort. Additionally, the minimum-distance test is not easy to interpret or applyand better spatially diverse metrics need to be investigated);• varying event detection threshold values should be carried out during each sort (or animalrecording) and automated methods for determining optimal event detection threshold should39Figure 6: In vitro datasets multi-operator sorting results. A. Firing rate distributions of invitro recorded (excitatory) neurons used for spike sort testing. B. Example of extracellular traces(50ms) on a single shank of the H32 extracellular polytrode. C. Examples of cluster separationfor units sorted from the dataset. D. Purity and completeness metrics from three operators whosorted the datasets. E. Number of units sorted (left) and number of spikes sorted (right) for threeoperators. F. Sorting quality using a single-value-metric (see main text) for the three operators.(Adapted from Mitelut et al., 2015 with permission).be eventually implemented (see e.g. latest version of SpikeSorter, Swindale and Spacek, 2014);• purity values across operators sorting in vitro recordings were similar and quite high (∼90%),however, completeness values varied much more suggesting that operator skill plays a signif-icant role in spike sorting.Using novel in silico datasets for spike sorting validationWhile the in vitro recordings discussed above provide ground-truth (i.e. spike times) for singleneurons recorded extracellularly, the datasets are limited in a number of ways. First, the datasetscontain only single neurons (that need to be combined artificially to generate multi-neuron record-40ings) and were acquired in ”cold” (i.e. 20o) cortical slices which have no background spiking activityfrom other neurons or the patched neuron. Additionally, the recordings were obtained using onlyone type of electrode configuration (i.e. the Neuronexus H32 layout) and cannot be adapted toother more common multi-column high-density electrodes.In order to provide more complex and flexible datasets, biophysically detailed network modelswere employed to generate simulated extracellular recordings on H32 probes and higher density(e.g. 20µm spacing) electrode layouts. In this section a brief review of single neuron and networkmodeling is provided before discussing biophysically detailed in silico datasets and spike-sortingresults across several tests and electrode layouts. Data generated and analyzed here comes fromin silico simulations prepared for a number of studies: Mitelut et al., 2014; Mitelut et al., 2015;Hawrylycz et al., 2016; Jun et al., 2017a; Jun et al., 2017b. All simulations and figures weregenerated by the author (the exception is Fig 3.18 based on data generated by the author andsorted by J. Jun; see also Jun, Mitelut et al., 2017a).Background - computational models of neurons and networksOne approach to investigating brain function has been to build comprehensive forward models(Buzsa´ki et al., 2012) - i.e. models that take into account all or most of the known single neuronbiophysical properties to build single neuron and cortical network models. The network modelsare constrained using results for pair-wise neuron and network connectivity experiments and largeensembles of connected neurons can be used to simulate neuronal activity at multiple spatial andtemporal scales.Developing models of single neurons that account for synaptic input to compute the somaticresponse, i.e. somatic action potentials, has been an ongoing project for many decades (see Fig 3.7for morphology and synapse model for a mouse V1 neuron). In very early studies only the somaticcompartment was considered with synaptic input directly affecting the soma (i.e. McCulloch andPitts neurons; McCulloch and Pitts, 1943). The Hodgkin and Huxley model of the action potential(Hodgkin and Huxley, 1952) introduced more complex nonlinear properties of membranes but itwas not until Wilfrid Rall’s adoption of cable theory to model single neurons (i.e. compartmentalmodeling; Rall, 1964) that the contribution of distant dendritic processing (including both passiveand active membrane properties) was taken into account. Modern computers further advancedmodeling with arguably the most successful and commonly used single neuron modeling software,i.e. NEURON, (Hines, 1986; Hines and Carnevale, 1997) eventually being adopted to parallelcomputer hardware (Migliore et al., 2006) and used in large scale cortical simulations such as theAllen Institute and the BBP.In parallel to theoretical model development, single neuron recordings have lead to the estab-lishment of databases for 3D neuron morphologies (i.e. www.neuromorphy.org; Ascoli et al., 2007)which at the time of writing contains 62,304 single neuron reconstructions, from 206 different brainregions from dozens of species contributed by 278 groups around the world. Other databases (e.g.41Figure 7: Single neuron morphologies and synapse locations. A. Examples of single neuronmorphology with nonspecific synapse distributions. B. Same as (A) but zoomed in example. (Datagenerated using netbuilder and visualized using biovis).ModeDB; McDougal et al., 2017) provide complete models of single neurons, i.e. they provide bothmorphology and descriptions of membrane conductances in models that can be readily simulated(over 1100 published articles as of the time of writing).However, the development of good single neuron models requires - at a minimum - in vitro or invivo patching of single neuron somata (and dendrites if possible) and the recording of somatic actionpotentials under different DC and AC current loads (actually, fitting neurons using extracellulardata is better, i.e. more constraining, Gold et al., 2007; however, ideal fitting requires measuringactivity at all dendrites which is not currently possible using available experimental tools). Sub-sequently fitting multi-compartment single neuron models, i.e. finding the optimal parameters ofactive and passive membrane conductances and distributing them along the dendritic and somaticcompartments to replicate in vivo and in vitro responses of single neurons - is challenging. Onlyvery recently has single neuron model fitting become standardized and automated led in large partby the BBP and the Idan Segev group (Hay et al., 2011, Hay et al., 2013). These efforts resultedin the development of ”evolutionary” algorithms that can even generate single neuron models withcomplex behaviours such as Ca2+ spikes and back-propagating action potentials.While initial single neuron modeling software suites (e.g. NEURON) were developed to com-pute the intracellular and transmembrane currents for a given description of a neuron, computing42Figure 8: Extracellular potentials of multi-compartment neuron models. Examplesof extracellular potentials computed in a single plane during a single spike from two BBP ratsomatosensory Layer 5B thick tufted neurons show the exponential decay of extracellular potentialwith distance from the soma. Note the colour scaling indicates the maximum PTP amplitude of thespike as per the colour scheme. Inset shows the extracellular waveform in the highlighted section(black outline) decaying exponentially as a function of distance from the soma. (Adapted fromMitelut et al., 2015, with permission).extracellular potentials at an arbitrary location in space from single neuron intracellular activity isa more recent addition (e.g. the addition of the extracellular mechanism in NEURON). The mostcommon approximation for computing the extracellular potential of multi-compartment neurons isthe line-source-approximation (LSA) which can approximate the potential at any location in spacefrom single neuronal compartments with specific transmembrane currents and known lengths (Fig3.8; Plonsey, 1974; Holt and Koch, 1999; Gold et al., 2006, 2007).Computing the extracellular potential (using Ohm’s law: V=IR) at a point in space basedon multi-compartment single neuron models requires the computation of the voltage from trans-membrane currents of every 3-dimensional (cylindrical) compartment of a (simulated) neuron. TheLSA reduces the distribution of trans-membrane current from the surface of each (cylindrical) com-partment to a line segment that passes through the centre of each segment. Thus, for each neuron,the extracellular potential V at electrode site i is computed by summing over the transmembrane43Figure 9: Allen Institute network model of mouse V1. Allen Institute network model ofmouse V1 using biophysically detailed model multi-compartment single neuron models (shown at10% in vivo neuronal density; adapted from Mitelut et al., 2015 with permission).current from each simulated neuron segment j :Vi = Σj(RijIj) (6)where Rij is the resistance:Rij =14πr(7)and has to be computed for every distance r along the neuron segment (i.e. by integrating oversegment). The LSA simplifies this computation as the integral can be computed analytically (formore details see Holt and Koch, 1999; Gratiy et al., 2017).Using this approach it is possible to generate simulated extracellular recording data for singleneurons to test our understanding of active and passive channels in neuron models (Holt and Koch,1999; Gold et al., 2006, 2007).Over the past several years, the Allen Institute for Brain Science has been developing methodsto characterize and catalog morphologies and physiological properties of large numbers of single44neurons from mouse V1. In parallel with these efforts, the modeling group at the Allen Institutehas developed Python-based interfaces to NEURON (which is a C/C++-based low-level simulationenvironment) to describe and simulate network models of V1 at different levels of granularity(i.e. detail; Hawrylycz et al., 2016). The lowest-level, most detailed network models use multi-compartment single neuron models (with passive and active) dendritic conductances that can besynaptically connected and driven by background and in vivo-like inputs recorded from LGN duringawake and anesthetized mouse experiments (Fig 3.9; Gratiy et al., 2014,Gratiy et al., 2015, Gratiyet al., 2017).The initial biophysically detailed Python-based network model (Gratiy et al., 2014) did notcontain an extracellular potential calculator. Accordingly, some of the in silico datasets presentedand analyzed here required a temporary solution (i.e. a Fortran-based calculator implemented bythe author; see results below; see also Mitelut et al., 2014). However, over the past two years aPython-based simulator and extracellular potential calculator has been developed (Gratiy et al.,2015, 2017) and several simulated datasets have been developed and used to test spike-sorting algo-rithms (Mitelut et al., 2015, Jun et al., 2017a) and the next-generation of high-density extracellularelectrode layouts (Jun et al., 2017b).Currently, a multi-purpose, cortical area-independent interface that uses a graph theoretic ap-proach to describe cortical networks (i.e. using nodes and vertices) has been developed (Fig 3.10;Gratiy et al., 2017 in preparation). This most recent version of contains a simulator component -Bionet (main author: S. Gratiy; Gratiy et al., 2017) in addition to a network description buildingtool - Netbuilder(www.github.com/netbuilder; author: Y. Billeh) and an OpenGL based 3D visu-alization tool - Biovis (www.github.com/catubc/biovis; author C. Mitelut; see Appendix ). Theapproach first requires a description of a network including morphological types, synapse types andconnectivity matrices; and then enables simulations of the network using the provided connectivityin addition to time-varying synaptic inputs (e.g. LGN-like inputs simulating thalamic input to V1).Results of sorting in silico datasetsSpike sorting methods were applied to in silico recordings in four tests: a single operator sortingtest using a 500 neuron patch recorded on an extracellular H32 probe layout (Test #4); a 3-operatorsorting test of a ≈3,200 neuron patch simulation recored on the H32 probe (Test #5); a 7-operatorsorting test of a ≈3,200 neuron patch simulation recored on a 30-channel high-density probe (Test#6), and a single operator sorting test of 1-column vs 2-columns using a 30-channel high-densityprobe (Test #7). The simulations were run using early development versions of BioNet (Gratiy etal., 2014, 2015) using hundreds to several thousand neuron networks. The single neuron modelscame from 12 unique neuron morphologies (7 pyramidal and 5 basket cells) from BBP-relatedstudies (Hay et al., 2011, Hay et al., 2013, Hu et al., 2009, Norenberg et al., 2010). There weremultiple electrode layouts simulated but only two are discussed in this thesis: the H32 probe (usedfor in vitro recordings previous section) and a 2-column, 30-channel high-density electrode (20µm45Figure 10: Bionet network examples. A. Examples of an instantiated networks showing onlythe somata location in a mouse V1 column with additional extra-columnar ≈L4 neurons (note:L1 neurons have brighter colours, all other layers are dimmed). B. And example of a verticalslice (width ≈100µm) showing neurons in L4, L5 and L6. (Data generated using netbuilder andvisualized using biovis).spacing in the x- and y-axes) which was based on a new generation of high-density electrodes - i.e.IMEC probes - that were in development (soon to be available to the public, see Jun et al., 2017b).The simulations were run on supercomputer clusters through the Neuroscience Gateway (NSG)at the University of California, San Diego - Super Computer Centre. Many different simulationswere run (only some of them are presented in this chapter) ranging from 30 seconds to 10 minutesin duration and using electrode layouts with 30 to as 7,500 sites. The simulations generated upto several hundred thousand single spikes (see Appendix A for more details). Several of thosesimulations were selected for spike sorting by multiple operators using different spike sorting suitesand approaches. The overall findings are that electrodes with higher channel count and higherspatial density substantially improve sorting quality and quantity. However, operator skill wasalso a significant factor for sorting quality. Surprisingly, 2-column high-density probes may onlymarginally improve sorting quality compared to single-column configurations suggesting that futurehardware development should focus on lower footprint, possibly multi-shank (3D), but single columnelectrodes.The simulations, analysis of spike-sorting rasters and figures presented in the remaining sectionsof this chapter were all generated by the author.46Figure 11: Simulated neuron morphologies and physiology. A. Examples of an intracellulartrace (100ms) from a biophysically detailed actively spiking single neuron. B. Partial morphologyand soma (black) of a neuron near an 8-channel extracellular electrode (red). C. Intracellulartraces (blue) and spike rasters (black) of neuron in (A) over the entire 60sec simulation period.D. Extracellular traces of all spikes of the sorted neuron. E-H. Same as (A-D) but for a differentneuron. (Adapted from Mitelut et al., 2014 with permission).Test 4This 4th spike sorting test relied on in silico datasets created using an early Python-based networksimulator (Gratiy et al., 2014) and was previously presented (Mitelut et al., 2014). It was largelycarried out as a proof of principle to show that the simulations can generate realistic test datasetsfor spike sorting purposes (Figs 3.11,3.12).The test relied on a simulation of 500 biophysically detailed and synaptically connected neuronsbased on 7 morphologically and functionally detailed multi-compartment, biophysically detailedmodels (Hay et al., 2011, Hay et al., 2013). The simulations were run on a dual-Xeon CPUworkstation with 32 cores and 128GB of ram and required ∼48 hours for simulating the activityand saving transmembrane currents to disk (using a ∼1TB of space on solid-state-drive) withthe computation of extracellular waveforms requiring an additional ∼168 hours to compute (usingan offline Fortran calculator). The simulated dataset was 60 seconds long and contained >37,000extracellular spike waveforms. The data was analyzed by the author using SpikeSorter and revealedthat the extracellular waveforms had typical characteristics of in vivo recoded spike waveforms (Fig3.11D,H).The result of this test is summarized in Figure 3.12B as unit detection accuracy vs. PTPextracellular amplitude. Detection accuracy (a metric identical to purity) for all neurons sorted47Figure 12: 8-channel probe simulated traces and sorting results. A. Examples of simulatedextracellular traces recorded on a single shank of an H32 probe (8-channels). B. Unit detectionaccuracy (i.e. TP rate; see also main text) as a function of unit PTP amplitude. Note that someneurons are outliers (arrow) despite having large PTP. (Adapted from Mitelut et al., 2014 withpermission).revealed an expected asymptotic relationship with accuracy increasing with increasing PTP am-plitude. However, a few neurons seem to deviate substantially from this asymptotic relationship(e.g. Fig 3.12B - small green cluster with >300µV amplitude but <25% detection accuracy) sug-gesting either a combination of similarity in extracellular waveform or over-splitting of units due tospike-amplitude adaptation during burst spiking (a mechanism present in the BBP single neuronmodels). These initial simulation and sorting results validated the use of simulated extracellulardata by confirming the expected dependence of sorting on extracellular PTP amplitude which wasobserved for in vitro datasets previously tested.Test no. 5The next test was a 3-operator sorting test basd on simulated data for the H32 probe layout.This test was the same as Test 3 (which used an in vitro dataset; see above) but using simulateddatasets that contained several thousand simulated neurons for which ground-truth rasters wereknown (Fig 3.13).The activity of 3,198 inter-connected neurons were simulated and a 4 minute recording wasgenerated. The neurons had spiking rates of 5Hz-10Hz and clustering of data (using SpikeSorterselected features and PCA) revealed clear clusters in most cases (Fig 3.13C). Three operators (sameas Test 3) sorted the data using SpikeSorter (Swindale and Spacek, 2014: Fig 3.13D-F - grey andblue) and KlustaKwik (Kadir et al., 2014: Fig 3.13D-F - green). The purity values were 81% to 85%(somewhat lower than the in vitro sorts of 88% to 93%). Completeness rates were: 55-67%, similarto in vitro results (53% to 70%). Considering the number of units sorted (Fig 3.13E - left) thereis substantial variability: one operator found 27 units (blue) whereas another operator identified35 units (green). This ∼30% increase in additional units (over the blue operator) is substantial48Figure 13: In silico datasets multi-operator sorting results. A. Firing rate distributionsof 3,198 in silico excitatory and inhibitory neurons used for spike sort testing. B. Example ofextracellular traces (50ms) on a single shank of the H32 extracellular polytrode. C. Examples ofcluster separation for units sorted from the dataset. (Note clusters were separated using a basick-means algorithm for display purposes only). D. Purity and completeness metrics from threeoperators who sorted the datasets. E. Number of units sorted (left) and number of spikes sorted(right) for three operators. F. Sorting quality using a single-value-metric (see main text) for thethree operators. (Adapted from Mitelut et al., 2015 with permission).and (as discussed below) strongly suggests that operator skill is very important in sorting. Asexpected, the number of spikes clustered also varied with each operator: from ∼45,000 spikes (greyoperator) to ∼64,000 spikes (green operator) constituting an increase of 42% more spikes acrossresults. This is a significant difference which suggests substantial room for improvement in sorting(for the grey operator) either in subjective decision making (e.g. setting thresholds or splittingclusters) or cluster cleanup and is briefly discussed further below. Finally, the SVM value (i.e.the sum of purity x completeness for all units; see discussion above) revealed a ∼22% improvedperformance of one operator (green) over the lowest performing operator (grey). This suggests thatwhile the best operator (green) found 30% more units and 42% more spikes, a portion of the spikes49Figure 14: High-density electrode simula-tion layout. Simulation layout using L5B neu-rons (and hippocampal basket neurons) indicat-ing the location of the electrode and neuron bod-ies displayed at 0.01% original simulation density.(Adapted from Mitelut et al., 2015 with permis-sion).and units were not correctly identified resulting in a lower overall SVM than expected.Test no. 6The spike sorting tests discussed so far used older electrode layouts (i.e. the H32 polytrode). Thetests below (Test 6 and Test 7) were carried out on high-density extracellular electrode layouts.These electrode layouts were based on real electrodes being developed in a multi-year collaborationbetween the Allen Institute, University College London, Howard Hughes Medical Institute (JaneliaFarms) and Interuniversity MicroElectronics Center (IMEC) in Belgium. (Note: at the time ofwriting, the manuscript describing this largely electronics engineering effort are under review, seeJun et al., 2017b).The simulations used in Test 6 relied on a similar 3,198 neuron network patch (used in Test5) but on a 2-column 20µm spacing (x- and y-axes) 2nd generation IMEC electrode layout (Fig3.14). (Note: additional simulations testing variations on this IMEC-polytrode layout in supportof Jun et al., 2017a were carried out but are not presented here. The simulations had similar firingrates, voltage traces, cluster distributions and unit templates as in vivo recordings; Fig 3.15). Theextracellular simulation used a previous version of the Allen Institute biophysically detailed model(Gratiy et al., 2015) and was run using the Neuroscience Gateway service (ssh access) established50Figure 15: High-density simulated extracellular recordings. A. Firing rate distributionsof 3,198 in silico excitatory and inhibitory neurons used for spike sort testing. B. Example ofextracellular traces (50ms) on the 2-column IMEC polytrode. C. Examples of cluster separationfor units sorted from the dataset. (Note clusters were separated using a basic k-means algorithmfor display purposes only). D. Example sorted units from simulated data. (Adapted from Mitelutet al., 2015 with permission).at the San Diego Super Computer Centre (Sivagnanam et al., 2013; see also Appendix A). Thesimulations provided 4 minutes (240 seconds) of extracellular data on a 30-channel version of theIMEC polytrode sampled at 20Khz.The datasets were sorted by 7 different operators using 4 different spike sorting suites: 2 op-erators used SpikeSorter (Swindale and Spacek, 2014), 1 operator used a Python-based version ofSpikeSorter with some modifications (https://github.com/spyke/spyke), 2 operators used KlustaK-wik (Kadir et al., 2014), 1 operator used SpikingCircus (Yger et al., 2016) and 1 operator useda development version of Kilosort (Pachitariu et al., 2016). An additional operator relied on adevelopment version of JRClust (Jun et al., 2017a), but the results were not incorporated into thisdataset as JRClust was not fully developed at the time (i.e. the results were quite poor). In orderto encourage participation, the agreement with the operators was that names and suites used werenot to be identified.51Figure 16: High-density simulated extracellular recordings - sorting results. A. Purityand completeness metrics for the 7 operators sorting the data. B. Number of sorted units (left) andnumber of sorted spikes (right). C. Single value sorting quality metric for all operators. (Adaptedfrom Mitelut et al., 2015 with permission).The results are presented in Fig 3.16. Purity values (i.e. the average purity across all unitssorted by each operator) ranged from 83% (cyan operator) to 99% (red operator) and completenessvalues ranged from 52% (grey operator) to 74% (pink operator) (Fig 3.16A). The number of sortedunits, however, varied widely from: 22 units (cyan operator) to 61 units (green operator) with otheroperators identifying different numbers of units in that range (Fig 3.16B). Similarly, the numberof spikes sorted also varied from ∼47,000 spikes (cyan and red operators) to >91,000 spikes (greenoperator). The SVM sorting metric found the expected rank of sorters based on the number ofunits and number of spikes. In particular, the SVM correctly identified the best (green) and worst(cyan) operators, with the worst operator doing more than 2 times worse than the best operatorand lagging substantially behind all other operators.The variable findings with respect to the number of spikes and number of units are of significantconcern as it suggests substantially different units and spikes are identified by different operatorsand suites. What’s quite surprising is that the best (green) and worst (cyan) sorts were carriedout using the same spike sorting suite (name not disclosed). It is also of some concern that thedistribution of number of spikes and number units is not only broad, but no 2 operators identifiedthe same number of units. This suggests that highly variable sorting results are likely present for invivo sorts adding substantial variability to results reported in publications. This conclusion cannotbe understated especially as the number of spike sorting suites continues to increase with time.A final caveat should be noted: while the best performing operator (green) identified sub-stantially more units (≈20) than other operators, the low PTP amplitude units identified by thisoperator had substantially lower purity and completeness values (not shown). For in vivo applica-tions, it cannot be known whether the addition of such lower-quality units helps with analysis (e.g.by increasing the clusterability of firing-rate state space distributions - e.g. Mazor and Laurent,52Figure 17: Sorting results - 2-column vs 1-column electrode layouts. 2-column IMECprobe layout vs a 1-column version of IMEC probe (using the same vertical density). B. Spikesorting results reveals only minor spike sorting differences between the two layouts (see also maintext; adapted from Mitelut et al., 2015 with permission).2005) or whether the addition of these units would confounds analysis such as in single visual neu-ron properties. Thus, it is unclear what the usefulness of such units is without further tests and invivo applications. However, it is important to note that this question could be pursued further viasimulated datasets and ground-truth for low-PTP amplitude units.Test no. 7The last test presented here is aimed at addressing a way for reducing electrode damage whichcontinues to be a concern and active efforts to minimize damage are studied (e.g. Kozai et al.,2010). This final test was carried out to determine whether narrower electrodes would result ina loss of sorting precision. Specifically, the 2-column IMEC probe was tested against a 1-columnversion of the probe. This test was implemented in a simple manner by sorting simulated dataacquired on the 2-column IMEC probe layout and comparing with sorted data from the samesimulation but ignoring (i.e. removing) one of the columns from the extracellular voltage record.Surprisingly, sorting the data using the 1-column version of the probe yielded very similarsorting error rates (Fig 3.17). An interpretation of this result is that given the size of the somas inthe simulations (i.e. ∼20µm to ∼35µm diameter) from the perspective of single-unit spike sorting,electrode density may reach a saturation point beyond which increasing vertical density (or evenhorizontal density) provides limited benefits.Thus, this test suggests that future polytrode development efforts should focus on minimizingpolytrode size (i.e. width) to minimize damage to cortex which may be a more important goal inelectrode development. In fact, polytrode fabrication techniques have already reached extraordinarydensities (e.g. 1µm inter-site-spacing and 30µm electrode thickness; private communication EdBoyden lab).53Figure 18: Sorting results are similar across in vitro, in vivo and in silico datasets.A. Definition of false negative and false positive metrics used for comparing spikesorting resultsacross in vitro, in vivo and in silico datasets. B. The number of sites on which each sorted neurontemplate has a voltage peak (Vp) amplitude > than 3 x the RMS of the signal (i.e. Vrms). C.Data from (B) reveals no statistical difference between sorting real vs. simulated data. D. Falsepositive rates across different dataset types show similar trends. Inset shows the distribution ofsimulated units as a function of PTP amplitude (i.e. Vp) divided by the noise RMS (i.e. Vrms).E. False negative rates are also similar across the datasets. F. There are no statistically significantdifferences between real and simulated data. (Adapted from Jun et al., 2017a with permission).Ongoing efforts for spikes sorting suite developmentAs part of an ongoing effort to develop independent datasets for spikesorting tests several dozenbiophysically detailed network simulations have been generated and analyzed over the past threeyears. Several of those datasets have been analyzed and presented in the sections above (Tests#1-7). Simulated datasets reproduce spatio-temporal aspects of high-passed extracellular poten-tial data (i.e. spike records) and confirm (as expected) spike detection and sorting accuracy decaysubstantially with decreasing spike PTP height (i.e. single neuron distance from electrode). Sim-ulated recordings also have signal noise standard deviation of 8-11µV similar to in vivo recordings(though simulations with higher firing rates can have slightly higher standard deviations). Overthe past 1.5 years many of the simulated datasets were made available on the author’s freely acces-sible website (www.spikesortingtest.com; see also Appendix ). The website serves as a cloud-basedvalidation tool for testing spikesorting algorithms. Users are able to download simulated datasets,sort them using their preferred spike sorting suite and upload the sorting results. They receivea report usually in 30-60 seconds. The website was implemented using multiple cloud computingtools including Django, and runs on Python and Fortran modules on the server side.54Lastly, a new spikesorting toolbox (JRClust; Jun, Mitelut et al 2017) has been developed overthe past two years which implements a novel signal transformation approach, clustering methodand automated drift tracking and seeks to increase sorting automation (i.e. reduce operator input).JRClust relies largely on large simulated datasets for testing and uses graphics-processing-unit(GPU) based code to achieve faster than real-time sorting even in recordings with more than100-channels. Importantly, JRClust was used to sore more than dozen simulated datasets and itwas shown that spikesorting results of simulated datasets were statistically indistinguishable fromsorting the available hybrid-in vivo (12 neurons) and in vitro (4 neuron) datasets (Fig 3.18). Inparticular, it was shown that the physical extent of each neuron’s template, i.e. the number of siteson which each neuron spike has a voltage peak (Vp) amplitude > than 3 x the RMS of the signal(i.e. Vrms) was very similar between the different datasets (Fig 3.18B, C). Second, false positive(Fig 3.18D) and false negative (Fig 3.18E) rates are similar across the datasets and there are nostatistically significant differences(Fig 3.18F).DiscussionThe general findings of sorting in silico datasets are that:• most operators and sorting suites have high purity (i.e. low error) rates and moderate com-pleteness (i.e. missed spikes) rates;• the number of spikes and units sorted vary substantially across operators with as much asthree times less units (e.g. 61 vs 22) sorted by one operator vs another even when using thesame sorting suite;• simulated datasets are a good representative of in vivo and in vitro recorded data as theyhave similar signal-to-noise ratio of extracellular voltage signal, similar extracellular templatesize distribution and similar neuron yield;• as expected, higher density probes provide lower error (e.g. higher purity) rates than lowerdensity probes;• electrode density may saturate beyond a value (e.g. <20µm) and further increases in channeldensity may not be useful for unit isolation while potentially being more damaging to tissue.Large-scale simulations (i.e. containing many neurons and electrode sites) can provide previ-ously unavailable ground-truth data making it possible to test both novel spike sorting methodsand novel electrode layouts. The availability of large ground-truth datasets may also enable theanalysis of more complex issues such as the role of operator skill - including identifying and opti-mizing automation based on subjective decisions. The ability to generate arbitrary electrode layoutmay even be used to test the limits of extracellular electrophysiology, i.e. determine what electrode55Figure 19: Allen Institute active conductance neurons have small spatial propagation.Example extracellular potentials (∼40channels) showing that single neuron spikes are largely local-ized to single channels and do not propagate to nearby channels making them more challenging tosort using sorting algorithms (Note: only spikes from the first 500 simulated neurons are highlighteddue to limitations in visualization in SpikeSorter.layouts are required in order to capture all - or the vast majority of - neuronal spiking in a patchof tissue.However, modeling extracellular voltages has not been perfected. For example, while datapresented in this chapter comes from BBP single neuron models,the Allen Institute has also releasedtheir own mouse V1 neuron models that also contain active conductances. While arguably moreappropriate for use with the Allen Institute network models, unfortunately, the Allen Institutesingle neuron models generated extracellular spikes that were difficult to sort due to very smallextracellular waveforms (Fig 3.19). That is, the extracellular waveforms for each spike was confinedmostly to one channel resulting in an extracellular record that looked more like it was obtainedfrom multiple single neuron (extracellular) recordings rather than using high-density electrodes.There are several important remaining challenges in spike sorting including: capturing low-amplitude neurons, increased automation of drift correction and improved overall sorting automa-56tion (i.e. decreased human supervision). Because of the approximately gaussian distribution ofextracellular potential values around Ve=0µV, it will continue to be challenging to capture loweramplitude neurons. Even though lower noise hardware is being developed, this will not overcomethe 1/r dependency of extracellular signals. Thus, distant neurons will always contribute noiseto recordings of closer, potentially more sortable units (e.g. 50µm to 100µm from an electrode).Automation will continue to be a challenge - but likely only for a smaller proportion of neurons,i.e. those neurons that have very similar (usually low-amplitude) templates and which will needto be compared manually. Drift correction may in fact be the easiest to tackle as ground-truthdatasets can be used to improve tracking of electrode movement (see Jun et al., 2017a for an imple-mentation of drift correction). One important result from this chapter is that spike sorting resultsusing SpikeSorter - the main sorting toolbox used in this thesis - is at the same level as other sort-ing tools. Overall, however, more discussions need to be had in the neuroscience community andfurther standards need to be developed where neuron yields are evaluated against neuron qualityand decisions are made about what types of error rates are acceptable. It is also important toemphasize the independence of spike sorting development and data generation/ground truth (seee.g. approach above and also Neto et al., 2016). Importantly, there is also a need for the develop-ment of some standards of analysis where independence is preserved between data generation, i.e.between experimentalists who patch neurons or developers who run simulations, and sorting suitedevelopers.57Optical Mapping of SpontaneousSingle Neuron Activity“Using voltage sensitive dye imaging, we previously established a close link betweenongoing activity in the visual cortex of anaesthetized cats and the spontaneous firing ofa single neuron...We suggest that dynamically switching cortical states could representthe brain’s internal context, and therefore reflect or influence memory, perception andbehaviour.“Tsodyks et al., 1999Spontaneous neural activity is present across all spatial and temporal scales. For example,single neuron spiking can be studied during stimulus-free periods using intracellular or extracellularrecording methods while the activity of large populations of neurons across entire cortical areas(e.g. visual cortex) can be studied using recent optical imaging methods (e.g. GCaMP6 or VSDs)or functional-magnetic resonance imaging (fMRI).Over the past couple of decades, a number of studies sought to relate spontaneous and stimulusevoked neuronal activity across spatio-temporal scales. The studies have been generally carriedout by simultaneously recording single neuron activity along with activity of large populationsof neurons making it possible to correlate activity at the micro-scale (e.g. spiking neurons) withmacro-scale neuronal activity. Two early studies (Tsodyks et al., 1999; Kenet et al., 2003) relatedspontaneous and stimulus evoked neuron spiking (using extracellular electrodes) to ongoing dorsalcortex activity (using VSDs over ≈2mm x ≈6mm regions) in cat V1. These studies made a num-ber findings including that stimulus evoked and spontaneous neuron spiking correlated with thesame pattern of cortical activity of large numbers (i.e. millions) of neurons as observed in VSD.Additionally, it was shown single neuron firing rates depended on the spatial pattern of ongoingpopulation activity in cortex suggesting that the population activity modulates the firing rate of asingle neuron. Interestingly, the studies identified the presence of ”dynamically switching corticalstates” (e.g. activated patterns that mimicked ocular dominance columns) which emerged spon-taneously suggesting ”dynamically switching cortical states could represent the brain’s internalcontext” (Kenet et al., 2003). In other words, the spontaneously occuring activity patterns werereminiscent of stimulus evoked patterns hinting that ongoing cortical activity may function as anactivity-space repertoire from which stimulus representations are selected.These findings suggest that functional studies across spatial scales is not only possible, but canyield insightful results that are otherwise not obtainable. They also confirm a strong link acrossspatial scales and that spontaneous activity, even at large spatial scales, can yield insight abouthow information is processed in the brain.58These pioneering studies suggest new directions for inquiry about micro-to-macro scale neuronalinteractions - and also about how such findings can be extended using more recent experimentalmethods. For example, do other imaging methods, e.g. based on [Ca] reporters, confirm or extendthese findings in other cortical areas and in other animals that do not have cortical structures such asocular dominance columns? Recently completed mouse anatomical connectivity database (i.e. AllenMouse Brain Atlas) document with extraordinary detail cortical and subcortical monosynapticconnections. Can additional multi-scale functional studies use this information and target largeareas of cortex and even subcortical neurons to define functional relationships across large areasof cortex? The work in this chapter expands on the pioneering cross-scale functional connectivitystudies discussed above to anesthetized and awake mice recordings using the latest transgenic [Ca]reporters, i.e. GCaMP mice. The work also relies on high-density (i.e 64-channel) extracellularprobes which not only decrease the single neuron spiking detection error rates (i.e. improve spikesorting) but also yield simultaneous spiking from dozens of cells while providing spatial informationacross multiple cortical laminae and subcortical nuclei. Wide-field optical mapping (WFOM; Maet al., 2016) techniques are used to simultaneously capture calcium activity across large areas ofmouse dorsal cortex including visual, barrel, motor and medial cortical structures and relate themto the spiking of single neurons. Methods for computing ”spike-triggered-maps” (STMs) usingGCaMP6 (with preliminary findings using VSD reporters) are also explored.The general findings of these investigations are that neurons from different cortical depths(but within the same area) yield similar dorsal cortex GCaMP6 STMs conforming to the expectedmono-synaptic anatomical connectivity. However, subcortical neurons - from the same or differentthalamic nuclei - exhibit substantial differences to their expected mono-synaptic connectivity mapsas well as temporal dynamics. In particular, subcortical neurons are more likely to be co-activatedduring dorsal cortex activity of areas that are not known to be monosynaptically connected basedon mouse brain connectivity information (e.g. Lein et al., 2007; Oh et al., 2014). The workpresented in the first section of this chapter (i.e. barrel cortex and thalamic neuron mapping) havebeen published (Xiao, Vanii, Mitelut et al 2017). The majority of experiments in this section werecarried out by Dongsheng Xiao and more than half of the analysis and figures were prepared by theauthor. For clarity all figures in the first section of this chapter will be captioned with initials ofthe author who made them as follows: CM: Catalin Mitelut, DX: Dongsheng Xiao, MV: MatthieuVanni.The findings presented in the second section (i.e. visual and auditory cortex neuron GCaMP6mapping) have been partially presented at a conference (Mitelut et al., 2016). All experiments andanalysis presented in the second section were carried out by the author. The findings presented inthe third section (i.e. auditory and visual cortex neuron VSD mapping) have not been presentedbefore. All experiments and analysis presented in the third section were carried out by the author.59GCaMP6 mapping of spontaneous activity of barrel cortex andthalamic neuronsSummaryUnderstanding brain function requires knowledge of cortical operations over wide-spatial scales fromsingle neurons to large populations. In order to investigate the relationship between spontaneoussingle neuron spiking and mesoscopic cortical activity, in vivo, wide-field imaging and sub-corticaland cortical cellular electrophysiology were carried out in GCaMP mice. A rich set of cortical activ-ity motifs were identified in spontaneous activity in anesthetized and awake mice. Using geneticallytargeted indicators of neuronal activity, mesoscale spike-triggered averaging allowed the identifica-tion of motifs (i.e. spatio-temporal patterns) that were preferentially linked to individual spikingneurons. Single thalamic neuron spiking correlated with cycles of wide-scale cortical inhibition andexcitation. In contrast, single cortical neurons correlated with mesoscale spatio-temporal mapsexpected for regional cortical consensus function. The approach can define network relationshipsbetween any point source of neuronal spiking and mesoscale cortical maps.BackgroundNeural activity ranges from the microscale of synapses to the macroscale of brain-wide networks.Mesoscale networks occupy an intermediate space and are have been widely studied in cortex form-ing the basis of sensory and motor maps (Bohland et al., 2009). These networks are largely definedby co-activation of neurons and have been evaluated with a variety of statistical approaches thatcapitalize on detecting synchrony. The study of large scale networks (meso-to macro-scale) hasbeen mostly restricted to functional magnetic resonance imaging (fMRI), or magnetoencephalog-raphy that can capture whole-brain activity patterns (de Pasquale et al., 2010; Kahn et al., 2011;Logothetis et al., 2012), but lack high spatial and temporal resolution and sensitivity. To overcomethese limitations, alternative approaches including mesoscopic intrinsic signal, voltage, glutamate,or calcium sensitive indicator imaging have been employed (Kleinfeld et al., 1994; Kenet et al.,2003; Ferezou et al., 2007; Chemla and Chavane, 2010; Chen et al., 2013b; Mohajerani et al., 2013;Stroh et al., 2013; Vanni and Murphy, 2014; Carandini et al., 2015; Chan et al., 2015; Madisen etal., 2015; Wekselblatt et al., 2016; Xie et al., 2016). New preparations using large scale craniotomies(Kim2016b) and large format imaging systems (Tsai et al., 2015; Sofroniew et al., 2016) providethe ability to link mesoscale activity patterns to individual neurons. However, these measures arerestricted to superficial layers of cortex and cannot assess functional connections to sub-corticalstructures. While developments in fiberoptic technology allow local optical functional assessmentof brain activity in sub-cortical structures (Hamel et al., 2015; Kim et al., 2016), they cannot simul-taneously resolve cortex over large fields of view. Although the evolution of imaging has revealednew aspects of cortical processing in identified neurons (Harvey et al., 2012; Chen et al., 2013a;60Chen et al., 2013b; Fu et al., 2014; Guo et al., 2014), the electrically recorded action potential isstill a signal of prominence given its temporal precision and ability to reflect the output of neuronalnetworks (Buzsa´ki, 2004).Extracellular recordings of single units were made in cortex, thalamus, and other sub-corticalsites with simultaneous mesoscopic functional imaging in transgenic mice expressing the calciumindicator GCaMP (Zariwala et al., 2012; Vanni and Murphy, 2014; Silasi et al., 2016). While slowerthan protein-based or small molecule voltage sensors, GCaMP imaging offers a high signal to noiseratio and is associated with supra-threshold activity which in some cases is a more direct reflection ofspiking activity. This work extends pioneering studies investigating the relationship between singleneuron spiking and local neuronal population activity assessed by voltage-sensitive dye imaging.Specifically, spike-triggered averaging (STA) has been previously used to identify the local activityprofile related to the spiking activity of a single neuron within a population (Arieli et al., 1995)and it was further demonstrated that this activity profile could reveal the instantaneous spatialpattern of ongoing population activity related to a neurons optimal stimulus in the visual cortex ofanesthetized cats (Tsodyks et al., 1999). The work presented herein extends these approaches andalso exploits the main advantage of mesoscopic imaging allowing the simultaneous measurementof brain activity in multiple regions across most of cortex simultaneously rather than activitysurrounding the recording site. This multi-scale strategy can help define temporal relationshipsbetween the activity of single neurons at the microscopic scale and mesoscale cortical structuralprojection maps (Zingg et al., 2014; Madisen et al., 2015). Furthermore, multisite silicon probesused here enable the assessment of long-distance activity relationships between multiple subcorticalsingle neurons and mesoscale cortical population activity. Spontaneous activity in awake andanesthetized mice was exploited as a source of diverse cortical network activity motifs (Mohajeraniet al., 2010; Mohajerani et al., 2013; Chan et al., 2015). Application of spike-triggered averaging inspontaneous cortical (calcium) activity linked the activity of single neurons to mesoscale networks.Single thalamic neuron spiking was found to functionally link to more diverse sensorimotor mapswhereas cortical neurons spiking was largely associated with consensus cortical maps. Thalamicneurons were found to both predict and report (firing before and after) specific cycles of wide-scalecortical inhibition and excitation, while cortical neuron firing was usually associated with excitation(and in some cases multi-second refractory depression). These results are consistent with an activecomputational role of thalamus in sensory-motor processing (Theyel et al., 2010; Hooks et al., 2013;Petrus et al., 2014; Sheroziya and Timofeev, 2014; McCormick et al., 2015), as opposed to merelyserving a relay function. The findings are also consistent with a diverse role of the thalamus infeed-forward sensory processing. Thalamocortical transmission can dynamically and differentiallyrecruit local cortical excitation and inhibition based on thalamic neuron firing patterns and wherethalamocortical feedforward inhibition is a critical feature(Galarreta and Hestrin, 1998; Swadlowand Gusev, 2001; Gabernet et al., 2005; Cruikshank2007; Hu2016). This spike-triggered cortical andsubcortical neuron mapping technique, exploiting mesoscopic calcium imaging, can be extended to61Figure 20: Experiment setup: multichannel electrode recordings and spike-triggered-averaging (DX,CM). A. Set-up for simultaneous wide-field calcium imaging and single unit record-ing using a glass pipette or laminar silicon probe. B. (i) Top view of wide-field transcranial windowand (ii) cortical atlas adapted from the Allen Institute Brain Atlas. C. Example of (i) cortical and(ii) subcortical pairs or spike recordings from separate channels showing cluster isolation in the firsttwo principal components. D. Procedure for generating a spike triggered average map (STM) for aunit located in barrel cortex.62Figure 20: (continued from previous page) E. (i). STM generated from single neuron with 1158spikes recorded in right barrel cortex. (ii) Red traces: Spontaneous calcium activity recorded fromtwo different cortical areas (BCS1 and HLS1). Blue trace: spontaneous spiking activity recordedsimultaneously from right BCS1. (iii) STM time courses generated from average of calcium activitytime-locked with each spike (red) and random spike (see Methods, black, blue: subtraction of spikeand random spike evoked responses) in region-of-interest (ROI). (Note: these examples are frommice under anesthesia).Figure 21: Spectral distribution of spontaneous activity (MV). Average power (±SEM) ofthe resting state activity of green fluorescence (black curve, n=20) and green 532 nm reflectance(gray curve, n=11) within barrel cortex of awake GCaMP6f mice.any brain location where electrodes can be placed to identify functionally linked cortical mesoscalenetworks.ResultsLinkage of individual spiking neurons to specific mesoscopic cortical mapsWide field of view mesoscale cortical imaging was used in GCaMP transgenic mice (Madisen etal., 2015) in combination with cellular electrophysiology recordings to derive cortical networks thatreflect activity at targeted point sources of neuronal spiking throughout the brain. Cortical andsub-cortical neuron spiking activity was recorded electrically while simultaneously imaging cortical63mesoscopic activity across a ≈9mm x ≈9mm bilateral window that encompassed multiple areas ofthe mouse dorsal cortex including somatosensory, motor, visual, retrosplenial, parietal associationand cingulate areas (Fig 4.1A,B). Spectral decomposition of the mesoscopic spontaneous activityusing GCaMP6 revealed the presence of information below 10 Hz that was distinct from non-specific green light reflectance (Fig 4.2). Given the slow Ca2+ binding and unbinding kineticsof GCaMP6, imaging dynamics are expected to be prolonged compared to actual spike records.Additionally, deconvolution was employed as described elsewhere (Pnevmatikakis et al., 2016) topossibly improve the time course of raw calcium signals relative to the spike-triggered-mappingtechniques used herein (Fig 4.3). While deconvolution improved the temporal dynamics of thedecay of the calcium signal, spike triggered analysis was only marginally affected and deconvolutionwas not implemented throughout.Spiking signals were initially recorded in multiple brain areas using glass electrodes (n=8 mice)to minimize obstruction of cortical imaging and reduce potential for damage from electrode place-ment. Subsequently, laminar probes (16 channel with 0.1 mm contact spacing) permitted theresolution of more spiking neurons simultaneously, and facilitated the recordings in multiple sub-cortical regions (n=16 mice). Given the invasive nature and the long duration of recordings, initialdata was obtained from urethane (n=4) or isoflurane (n=12) anesthetized adult mice, with addi-tional awake recording periods (n=12, see Methods). The spike triggered average maps (Fig 4.1D,E)obtained under both these conditions were qualitatively similar and this observation was consistentwith previous work using VSD imaging (Mohajerani et al., 2013). To perform these assessmentssingle neuron spikes were identified from extracellular recordings using spike sorting methods basedon clustering of principal components distributions of spike signals on sets of adjacent channels(Swindale and Spacek, 2014) (Fig 4.1C).Initial characterization of GCaMP6 (Chen et al., 2013b) suggested a calcium excitation risetime of ≈100ms and a minimum decay time of ≈150-200ms; accordingly, analysis of dynamics isusually limited to below 10 Hz. Additionally, [Ca] time-frequency power analysis on the imagingdata was used to evaluate power at different frequencies in the GCaMP6 signal and compared tothe green reflectance signal - which is not expected to carry neuronal activity information. Thisanalysis showed the presence of [Ca] activity power up to 8-10Hz which was substantially higher(5-10 times) than the power in the reflectance signal - validating the use of fluorescence signals upto this frequency limit (Fig 4.2).Spike triggered average maps (STM) were computed from simultaneously acquired wide-fieldcalcium imaging and single neuron recordings to investigate how single neuron spiking activityat a specific cortical or sub-cortical location was related to regional cortical activity (Fig 4.1D;see also Methods). For each individual spike, cortical imaging frames were considered from 3sbefore to 3s after the spike. The frames were normalized as dF/F0 by subtracting and dividingthe average calcium activity during the 3s preceding the spike (see also Methods for alternativedF/F0 calculations). Next, all spike triggered frame stacks were averaged into a multi-frame motif64Figure 22: Deconvolution does not substantially change STMs (DX). A. Top: Original(dF/F0) calcium image and time course of calcium dynamics in a region of interest (ROI). Bottom:deconvolved calcium image and time course in the ROI. B. Top: spike triggered average map (STM)and STM temporal dynamic (STMTD) of original data in ROI. Bottom: spike triggered averagemap (STM) and STMTD of deconvolved data in the ROI for spikes from the same neurons. (Note:neurons were recorded in thalamus under anesthesia).ranging from -3sec to +3sec centred on spiking (e.g. 6sec x 30frames per second = 180 frames;see also Fig 4.16A for a full-temporal resolution motif example). The STM - computed as asingle image representation - was then defined as the maximum (or peak) activity at each pixelwithin the time window of ±1 second of the single neuron spiking. This pixel-specific peak-valuemethod better captures correlated activity than merely averaging over the ±1 second intervalwhich smooths out highly activated but short duration - activity increases. The approach revealedthat the activity recorded from a single right barrel cortex neuron, for example, yielded an STMshowing strong and specific GCaMP signal in barrel and motor cortices of both hemispheres (Fig4.1E(i)). STMs calculated by averaging calcium activity centred on spiking activity accordinglyrevealed spatial specificity correlating with neuron spiking that was not present when comparedwith calcium activity in reference region (e.g. hind limb; Fig 4.1E(ii)) or random spike averaging(Fig 4.1E(iii); see also Methods).65Kinetics of spike-triggered mappingBy computing spike-triggered calcium image averages, the contribution of neurons which fire outof phase is substantially reduced. Analysis of STM time dynamics indicate slower time to peak(100ms) than postsynaptic potentials evoked by a single synaptic connection (time to peak 20ms)(Bruno and Sakmann, 2006). Slower dynamics are expected given the kinetics of GCaMP6 (Chen etal., 2013b) and deconvolution (Pnevmatikakis et al., 2016) can be used to take into considerationthe slower kinetics and compensate accordingly. Using this approach, a significant accelerationof raw data was observed but very modest effects on STMs indicating that STMs may alreadyrepresent accelerated activity relative to GCaMP6 kinetics perhaps due to the statistical natureof spike-Ca2+ transient temporal convergence. It is also possible that slower dynamics reflectsequences of spiking activity propagating through specific polysynaptic circuits. This speculationwas supported by the similarity in time dynamics during cue-triggered recall of learned temporalsequences shown in other studies (Xu et al., 2012). The kinetics of imaging can be improved infuture studies using faster sensors such as organic voltage sensitive dyes (Shoham et al., 1999;Mohajerani et al., 2013), or genetically encoded voltage (Carandini et al., 2015; Gong et al., 2015;Abdelfattah et al., 2016) or glutamate sensors (Xie et al., 2016) - but at the cost of lower SNR dataand more ambiguous interpretations of activity (e.g. glutamate activation; see also Chapter 4, lastsection on VSD STMs).Validation of GCaMP6 correlation with single neuron activityThe specificity of STMs computed in GCaMP6 mice (reflecting underlying neuronal activity) wasconfirmed by imaging Thy-1 GFP-M mice (n=6 mice) that lacked calcium-dependent neuronalfluorescent signals. Thy-1 GFP-M mice failed to produce functional maps using the same STMprocedure outlined above (Figure 4.4A,B). Additionally, the stability of STMs as a function of thenumber of neuron spikes was measured by quantifying the similarity of STMs generated from thetotal number of spikes vs subsets of spikes. Stable STM maps were generally observed using atleast 256 spikes (Fig 4.4C-E). The stability of STMs was also confirmed by comparing the mapsgenerated by splitting a unit’s spikes into two halves, or into odd and even groups which yieldsimilar STMs.Thalamic neurons show more diverse STMs than cortical neuronsThe next step was to apply the STM method to multiple recordings in cortex and thalamus andidentify mesoscopic networks associated with single neurons. First, the anatomical location of singleneurons was confirmed by labeling probes with Texas red-dextran or DiI as to visualize tracksand approximate the location of each sorted neuron on the electrode (Fig 4.5A, subcortical trackand nucleus identification from 4 different mouse experiments, see also Figure 4.4A,B). Spikingcortical neurons had very similar STMs in each cortical recording and were linked to consensus66Figure 23: Control recordings and convergence of STMs with # of spikes (DX) A.Simultaneous calcium and spiking activity recording in GCaMP6f mouse and STM yielded fromsingle unit recorded in barrel cortex.67Figure 23: (continued from previous page) B. Simultaneous GFP fluorescence and spiking activityrecording in Thy-1 GFP-M mouse and STM yielded from single unit recorded in barrel cortexresulted in no clear regional map. C. STMs generated from a subset of spikes (2 2048, on theleft) randomly chosen in one experiment. Correlation coefficients (r-value, on the right) betweenSTMs were used to evaluate the consistency of mapping. In this example, STMs generated bymore than 64 spikes generated a correlation >0.9 and were very similar between the pairs of SPMsmade using the same number of spikes. D. Distribution of correlation values between pairs of STMfor an increasing number of spikes. No significant change in r-value distribution was observed for512 spikes in comparison to 256 spikes (Mann Whitney test, P=0.126, U=948.5, 256 spikes groupn=58, r-value=0.97±0.01, mean±SD; 512 spikes group n=40, r-value=0.98±0.01, mean±SD). E.STMs and profile of responses computed using spikes divided into halves or even-odd sets. (Note:neurons in examples come from recordings performed under anesthesia).local and long range cortical networks (Fig 4.5B-left: single STM examples of cortical STMs; right:contours of all single neuron STMs recorded in each recording shows substantial overlap across alldepths). Consensus cortical networks have been discussed previously (Mohajerani et al., 2013) andare defined by using correlated activity mapping techinues (i.e. seed pixel analysis: each pixel’scorrelation value with every other pixel is computed over the entire recording). Consensus mapsreflect major mono-synaptic intra-cortical axonal projections and generally demarkate either singlesensory areas (e.g. visual cortex) or sensory areas and their major mono-synaptic projections (e.g.barrel and motor cortex; see also Mohajerani et al., 2013). Barrel cortex neurons have STMsthat substantially or completely overlapped with regional GCaMP signal changes in barrel andmotor cortex, as well as showing signals in homotopic areas of both hemispheres, consistent withthe previously observed pattern of long distance mono-synaptic connections (Ferezou et al., 2007;Mohajerani et al., 2013; Guo et al., 2014; Vanni and Murphy, 2014; Chan et al., 2015). Thesimilarity across cortical neuron STMs can be observed in ”contour” maps of each STM (computedas the border of full-width-half-max value STM) which largely overlap with each other (Fig 4.5B,C-contours).The patterns observed in thalamic neuron STMs also supported the presence of a functionallink between specific thalamic nuclei and consensus cortical projection areas (Hunnicutt et al.,2014; Oh et al., 2014; Zingg et al., 2014), but also showed additional variability and complexity tocortical neuron STMs. Thalamic neurons were associated with bilateral hemispheric signals withinmultiple primary sensorimotor and higher order brain areas (Fig 4.5C; Note: some unilateral STMsare observed generally in contralateral (to probe) areas and it is likely this is due to electrode damagein the ipsilateral hemisphere rather than unilateral mono-synaptic thalamic neuron connections tocortex - which have not been previously described).The use of multichannel probes enabled the recording of multiple single neurons across multiplecortical layers and multiple thalamic nuclei simultaneously. Accordingly, additional relative (andabsolute) laminar location analysis could be carried out. Cortical neurons in the same cortical area -68Figure 24: Topographic properties of neurons and neighbouring similarity metric (CM,DS). A. Electrode track for sample recordings (Blue channel: DAPI, yellow: DiI). B. STMs andoverlay contours of neurons recorded in barrel cortex with each color representing one neuron’sSTM border (see main text and Methods). C. STM and overlay contours of neurons recorded inthalamus in the electrode track presented in panels A and B. Color bar on the right side indicatesthe depth of each recording site. D. Diversity of overlap of STMs between neurons on neighbouringlaminar electrode channels. (i). Example of overlapping STMs (red area) between two corticalneurons recorded on adjacent channels. (ii). Example of overlapping STMs for neighbouring pairsof neurons recorded subcortically showing differences across depth.69Figure 24: (continued from previous page) (iii). Average neighbouring cortical neuron mapoverlap (blue: 93%) and neighbouring sub-cortical neuron overlap (78%) show significant dif-ferences (Mann Whitney test, p<0.0001, U=617408.0, mean percentage overlap of corticalSTM pairs=92.77±0.23%, mean±SEM, n=966; mean percentage overlap of sub-cortical STMpairs=78.11±0.61%, mean±SEM, n=1936). These results are from awake mice (except top row,Mouse #1).but from different depths - exhibited substantial similarity (i.e. overlap) and there were only subtletopographic changes in STMs across depth: static STMs were largely similar between and withinsuperficial and deep layers (see contour maps in Fig 4.5B; see also Fig 4.5D(iii) overlapping analysisresults). Subcortical neurons, had more diversity even within the same sub-nucleus in thalamus(see examples in Fig4.5C where neurons from neighbouring electrodes could have different STMs;see also examples in Fig 4.7 SUA STMs). To quantify the difference of variability of STMs acrossdepth for both cortical and thalamic recordings the percentage overlap of static STMs (Fig 4.5D;red areas) was computed for all pair-wise neighboring channel neurons (i.e. 100µm apart). STMsderived from neighboring electrode contacts in cortex were largely similar (mean of 93% across allcortical recordings) whereas sub-cortical STMs were more varied (mean of 78% Fig 4.5D(iii)).Additional methods of generating event triggered maps were explored including using eventsfrom MUA rasters or using LFP amplitude to trigger (i.e. scale) imaging frames. MUA-triggeredSTMs were computed using thresholded spiking (i.e. 4 times standard deviation of the signaldivided by a scaling factor Swindale and Spacek, 2014). Additionally, band-passed LFP triggeredSTMs were computed for Delta (0.1-4Hz), Theta (4-8Hz) and Gamma (25-100Hz) bands (see alsoMethods). Briefly, LFP triggered STMs were computed using data from 60 second recording periodsby scaling (i.e. multiplying) each imaging frame by the average LFP amplitude (at each recordingdepth) at that point in time. Thus, imaging frames during which the average LFP amplitude waslarge and positive contributed substantially to STMs, whereas frames where the average LFP valueswere closer to zero did not (negative LFP values were clipped). (Note: negative LFP values werealso explored with somewhat different results - not shown here).The MUA-triggered method revealed that both cortical and subcortical MUA triggered STMswere largely similar to single-unit STMs and did not vary substantially across laminar depth (Figs4.6, 4.7)). An additional method was implemented (see Xiao et al., 2017 reviewer comments online)where a normalized MUA-triggered STM was subtracted from single-unit STM with the hypothesisthat such methodology would reveal unique STMs not present in single neuron STMs. Unfortu-nately the resulting STMs appear largely as very noisy maps (i.e. maps with <1% dF/F0 valuesthat have no ROI specificity). This is due to the similarities between single-unit and MUA STMsbefore subtraction which results in cancellation of the vast majority of high-SNR calcium signal.The LFP-band triggered STM methodology revealed STMs that were substantially different fromsingle cell and MUA triggered STMs but stereotyped across depth (with some minor differences;7010mm-1    0    1Figure 25: LFP and MUA triggered STMs: cortical neuron examples (CM). STMs werecomputed at different depths of the electrode using single cell spikes, Multi-Unit-Activity (MUA)and LFP amplitude. Single unit STMs were computed (see Methods) for up to three representativecells at each depth. MUA STMs at each electrode were computed using all spiking activity overa threshold (math>4 times the standard deviation of the high-pass record divided by a scalingfactor - see main text). LFP triggered STMs were computed by scaling each image by the averageLFP amplitude - and averaging over a 60 second recording period. Thus, imaging frames wherethe average LFP amplitude was large and positive contributed substantially to the STM, whileframes where the average LFP values were low did not. The various band-passed LFP values usedwere delta: 0.1-4Hz, theta: 4-8Hz and gamma: 25-100Hz. Single cell STMs at each depth andacross depth have similar motifs to each other and MUA triggered STMs. LFP triggered STMs aresubstantially different from single cell and MUA triggered STMs and across different LFP frequencybands (see main text).71for example, see Fig 4.6 Theta band-triggered STMs change from left barrel cortex activation toleft limb cortex areas with depth). The LFP-triggered STMs were also different across LFP fre-quency bands. However, it is important to note that LFP amplitude- (or power-) triggered mapshad very low dF/F0 values. That is, most LFP-amplitude triggered STMs had peak dF/F0 valuesmath<0.05% dF/F0 whereas single neuron or MUA-triggered STM had peaks of 1-5% dF/F0. Thisindicates that the LFP-triggered method requires averaging over highly-variable dynamics (e.g. theLFP signal might vary substantially more than the GCaMP6 signal). Nonetheless, LFP-triggeredSTMs appear stable across depths and are highly unlikely to result from averaging random activ-ity. In fact, LFP-triggered STMs reveal anatomically discrete maps that are stable across multipledepths in cortex and subcortical areas with only small - but systematic - differences across layers(Figs 4.6,4.7). Whereas LFP-amplitude triggered STMs for delta band activity were similar tosingle neuron STMs, STMs associated with higher frequency bands showed different patterns sug-gesting a different type of functional connectivity present at higher LFP frequencies across spatialscales (Figs 4.6,4.7). As LFP contains mostly synaptic activity and only limited spiking activity(i.e. spiking only from within a region with radius of ≈250µm; Buzsa´ki et al., 2012) LFP-triggeredSTMs likely represent average synaptic activity from large neighbouring regions thus blurring theoverall effect of local LFP activity. However, because delta band activity is where most GCaMP6functional imaging indicator power is located (Chan et al., 2015), higher frequency componentsare closer to hemodynamic and other noise sources making interpretation of the results above 4Hzchallenging. Other, higher temporal resolution imaging methods (e.g. VSD) will likely yield addi-tional insight into the relationship between mesoscale maps and LFP activity across layers and insubcortical areas and is a topic to be investigated in future work.Sub-cortical neuron STMs are more diverse than cortical neuron STMsTo quantitatively determine how single cortical and sub-cortical neuron STMs relate to intra-corticalnetworks, single neuron STMs were compared (using cross-correlation) against a cortex-wide libraryof seed pixel correlation maps (Mohajerani et al., 2013; Vanni and Murphy, 2014; Chan et al.,2015) generated iteratively for all pixel locations in the same spontaneous activity recording (Fig4.8A,B). To create SPMs, the cross-correlation coefficient r values between the temporal profilesof one selected pixel and all the others within the field of view were calculated (Mohajerani et al.,2013; Vanni and Murphy, 2014; Chan et al., 2015; see also Methods). To evaluate the similaritybetween static STMs and SPMs, the correlation coefficient between pixels of both types of maps wascomputed for all possible SPMs in the library. The best matching SPM was selected, i.e. the SPMwith the highest correlation between a given STM and the library of SPMs. The library of SPMsis expected to reflect cortical consensus activity motifs (areas undergoing temporally-correlatedactivity) and can be largely attributed to underlying intra-cortical axonal projections (Mohajeraniet al., 2013). Single cortical neuron derived STMs were more similar to SPM correlation maps(Fig 4.8A), i.e. the correlation values were high between SPMs and cortical STMs (Fig 4.8C). In72Figure 26: LFP and MUA triggered STMs: subcortical neuron examples (CM). Thala-mic STMs: Spike vs. LFP. STMs were computed at different depths of the electrode using singlecell spikes, Multi-Unit-Activity (MUA) and LFP amplitude (see also Fig 4.6).73Figure 26: (continued from previous page) In contrast to cortical STMs, single thalamic neuronsat some depths have more varied motifs (e.g. electrodes 4, 11, 12, 15), while MUA triggered STMsappear similar across large thalamic regions (e.g. electrodes: 4-16). LFP triggered STMs aredifferent from single cell and MUA triggered STMs and across different LFP frequency bands (seemain text).contrast, thalamic STMs were more complex and corresponded to more unique distributions ofcortical patterns and had lower correlations with the cortical consensus SPM library (Fig 4.8B,C).For example, VPM, VPL and CP neurons can functionally link to multiple cortical areas and thisdiverse connectivity was not always present in SPMs (i.e. SPMs made from seeds in BCS1, HLS1,RS or other areas). It is possible that sub-cortical neuron STMs neurons are the super-position of2 or more cortical networks defined by SPMs. For example, some subcortical neuron STMs can bebetter described using pairs of SPMs which when added together provide a more complex map thatis arguably more similar to the single subcortical neuron STM (see Fig 4.8D for examples neuronSTMs constructed from combination of SPMs).To better understand the underlying structural circuit basis of distinct thalamic STMs, thalamicneuron location and projections were analyzed using the Allen Mouse Brain Connectivity Atlas(Oh et al., 2014) (Fig 4.9; see also Mohajerani et al., 2013). The 3D-atlas data was used to matchthe composed anatomical 2D maps with 2D static STMs. As expected, STMs of spiking corticalneurons corresponded with underlying structural axonal projections, i.e. monosynaptic connections(see BCS1 example in Fig 4.9, also Mohajerani et al., 2013). In contrast, not all sub-cortical neuronSTMs had monosynaptic projections from sub-cortical to cortical areas that accounted for theircortical mesoscale patterns. There are such examples coming from non-thalamic neurons: e.g.while hippocampus (HPF) does not have a strong direct anatomical link to RS, one identifiedHPF neuron had an STM that was dominated by RS co-activation. Another example showed thata caudoputamen (CP) neuron had a strong hindlimb/forlimb map (BCS1/HLS1) yet there wasno established direct monosynaptic links between CP and such areas suggesting that sub-corticalSTMs could represent polysynaptic links to cortex (Hunnicutt et al., 2014; Oh et al., 2014).Cortical and sub-cortical neuron firing is tuned to cortical network dynamicsspanning millisecond to multi-second time scalesAfter the static STM spatial analysis considered above, the next step was to consider time dynamicsand determine whether single neuron STMs had inhibition-activation patterns that revealed addi-tional, novel information about micro-to-mesoscale functional connectivity (Figs 4.10-4.15). STMTemporal Dynamics (STMTDs) were computed by identifying regions-of-interest (ROIs) with highactivation and tracking the time course of the maximally activated/depressed cortical pixel withineach ROI from 3s before to 3s after spiking. As all extracellular recordings were in right barrelcortex and predominantly right subcortical structures, the left barrel cortex (LBCS1) was used74Figure 27: STMs vs SPMs (MV,DX). A. Cortical STM (left) and the best fitting SPM (right)according to correlation coefficient (cc) values for different neuron locations (left of panel). Simi-larity was calculated by measuring the r-value Pearson coefficient between each pair of map pixels.Group data from 12 GCaMP6f mice are reported in panel C. B. Sub-cortical STM (left) and themost similar SPM (right). Cell #2 and #4 were from GCaMP6s mice. Cells #3, 5-7, 11 werefrom GCaMP3 mice. Other cells were from GCaMP6f mice. These examples were performed underanesthesia. C. Distribution of r-values (Mann Whitney test, pmath<0.0001, U=5227, sub-corticalgroup n=246 r-value=0.64±0.18, mean±SD; cortical group n=168 r-value =0.85±0.04, mean±SD).D. Examples of sub-cortical STMs compared with pairs of SPMs for seed indicated by a and b.75Figure 28: STM comparisons with Allen Brain Atlas projection maps (DX). A. Fromtop to bottom, example STMs from neurons recorded in BCS1, VPM, VPL, HPF and CP, respec-tively. B. Example projection maps (2D surface and 3D) reconstructed from Allen Brain Atlas withinjection sites (Oh et al., 2014) in the same region as the recorded neuron (in A). For example,for a spiking neuron recorded in BCS1 anterograde labeling of GFP emanating from an injectionsite in BCS1 shows mono-synaptic projections to motor cortex and is present across cortex in the2D surface plot of cortex. This projection pattern for BCS1 matches the STM map well and isconsistent with previous work (Mohajerani et al., 2013). In contrast, for sub-cortical injectionsof GFP tracer in HPF there was less overlap between STMs and projection maps perhaps sug-gesting STMs represent polysynaptic as opposed to merely monosynpatic pathways. (Website:2015 Allen Institute for Brain Science. Allen Mouse Brain Connectivity Atlas. Available from:http://connectivity.brain-map.org.76as the primary ROI as it contained the most activated and depressed calcium activity patterns.It should be noted that ipsi-lateral cortex (e.g. left barrel cortex in a left barrel cortex electrodepenetration) was also highly activated, but even for cortical neuron generated STMs, it showedsubstantially less consistent activation than the its opposite hemisphere homotopic area (likely dueto damage done by the electrode in the ipsilateral hemisphere).As observed for static STMs, the cortical calcium dynamics associated with cortical neuronswere relatively homogeneous with a peak in activity within ≈100ms-200ms following spiking and areturn to baseline (Figure 6A). However, some cortical cells (≈20%) participated in multi-seconddepression dynamics (see distribution of profiles in Figs 4.10, 4.17B,C and 4.20A). In contrast,STMTDs obtained from thalamic neuron recordings were more varied and were dominated bydepression dynamics (≈80% of cells) lasting up to 3 seconds (Fig 4.10B, 4.17B,C and Fig 4.20B).Bursting vs. tonic spiking modes reveal similar STMTDs for both cortical andsubcortical neuronsThe results above indicate that averaging GCaMP cortical motifs from all spikes of a single neuronproduces converging STMs and STMTDs for both cortical and subcortically recorded single neu-rons. However, thalamocortical synapses are known to be prone to synaptic depression and burstpattern firing may yield altered cortical responses and STMs (Castro-Alamancos, 1997; Gil et al.,1997). It is thus conceivable that averaging all of single neuron’s spikes over a period of severalminutes (or longer) could mask spatial heterogeneity observed when averaging over (functionallyor otherwise defined) sub-groups of spikes. The question is thus whether averaging over all of aneuron’s STMs is a correct (or representative method) of identifying a single neuron’s contributionto (or correlation with) ongoing cortical activity (see also reviewer’s comments in Xiao et al., 2017).Therefore, additional analyses were performed where spike triggered STMs and STMTDs werecomputed using only sub-groups of spikes either representing spiking modes (bursting vs tonic) ormotif similarity in a high-dimensional space with the goal of determining whether during differenttypes of active cortical dynamics single neurons contribute different types of STMs or temporaldynamics (Figs 4.11-4.15).Arguably the most obvious method for separating spontaneous neuron spiking into groupsis to partition them into bursting versus tonic modes (Figs 4.11-4.13; see Methods). The firstmethodology implemented was previously described for defining the main spiking modes of a single(thalamic) neuron (see Sherman and Guillery, 2006, Fig 6.5, pg 236). Briefly, the method requirescomputing the distribution of each spikes inter-spike-interval (ISI) between the previous (x-axis)and following (y-axis) spike (Fig 4.11). The resulting pre- and post-spike 2D distribution is thenplotted using logarithmic scales. Naturally arising groups are then clustered as the main spikingmodes of the neuron. In particular, spike groups occurring in approximately each quadrant of theplot indicate different spiking modes: first spikes in a burst (bottom right), burst spiking (bottomleft), last spikes in a burst (top left), and tonic spikes (top right). The vast majority of cortical77Figure 29: Cortical and Subcortical Motifs and STMTDs (CM,DX). A. Top: right hemi-sphere barrel cortex neuron motif and STMTD in a left hemisphere barrel cortex ROI. The maxi-mally activated pixel in the ROI (red arrow) is tracked over time and reveals dynamics which risequickly at spike time t=0 and decays in ≈1sec followed by a further 1-2sec cortical depression (redcurve in right plot).78Figure 29: (continued from previous page) Bottom: Additional examples of right hemispherebarrel cortex neuron motifs and STMTDs show similar barrel-motor cortex activation patternwith peaks in cortical activation shortly following spiking and a return to baseline or prolongeddepression. B. Top: Same as in A, but for a thalamically recorded neuron which correlates stronglywith motor cortex activation shortly after spiking. Bottom: Additional examples of thalamicneurons reveal both the spatial diversity (i.e. different STMs) and temporal diversity (i.e. differentSTMDs).neurons recorded in barrel cortex did not exhibit multiple classes of spiking modes, however 2examples are provided where some spike-mode clustering is present (Fig 4.11A,B). Despite theapparent spike-mode differences, the spike sub-groups representing different spiking modes yieldedvery similar motifs (Fig 4.11 A(ii),B(ii)) and similar STMTDs in the ROI of highest activation, i.e.left barrel cortex (Fig 4.11A(iii),B(iii)). Two additional examples of thalamic neurons are providedwhich show spiking modes (Fig 4.11C,D). However, much like cortical neurons the thalamic neuronsidentified had STMs and STMTDs that were largely stable across the clustered spiking modes.The isi-based methodology for identifying spiking modes did not reveal clear spiking modes forthe 3 cortical and 4 thalamic cells provided as examples (see Fig 4.10) nor for additional examplesconsidered (see examples in Fig 4.11). An additional methodology was implemented for previouslyconsidered neurons (i.e. neurons in Fig 4.10): bursting mode initiation was identified as spikingthat is preceded by at least 500ms of silence (Figs 4.12,4.13). Using this heuristic to define spikingmodes, single-neuron motifs and STMTDs were computed across all-spikes vs heuristically definedbursting mode spikes. The results were similar to the findings above: both motifs and STMTDs aresimilar across all-spike and heuristically defined bursting modes for both both cortical and thalamiccells examined.STM-space grouping and decomposition reveals single neurons participate instereotyped STMs that are present during spontaneous neural activityThe next step was to determine whether single spike STMs (rather than STMs generated as anaverage of STMs over all of a neuron’s spikes) fell into distinct clusters or groups - possibly indicatingunderlying spiking modes or network changes that could not be captured by traditional spikingmode analysis. The method used was to compute the STM for each spike, convert it to a highdimensional vector and determine whether some spikes gave rise to GCaMP STMs that were similarto each other and could be clustered or grouped (see Methods). However, none of the single neuronsconsidered showed evidence of naturally occuring clusters or groupings (not shown; but see Fig4.14B for single spike motif distributions using PCA; see also Methods). This suggests that singleneurons fire spikes during ongoing or ”spontaneous” mesoscale activity (i.e. activity not necessarilyassociated with our single neuron activity) and that single neuron spikes do not associate only withspecific mesoscale activity patterns.79Figure 30: Firing modality-defined STMs and STMTDs (CM) 80Figure 30: (continued from previous page) A(i), B(i): Cortical cell spiking modes determined bygrouping the distribution of each spikes inter-spike-interval between previous (x-axis) and following(y-axis) spike. The four quadrants indicate different firing modes (see also main text and Methods).Cortical cells (barrel cortex) generally did not have clear spiking modes but two examples areprovided where clusters were present and spikes were grouped accordingly. A(ii), B(ii): Six-secondmotifs generated using spikes from different spiking modes in part (i) are largely the same forcortical cells. A(iii), B(iii): STMTDs of left-hemisphere barrel cortex tracked across time for allspiking modes were largely similar. C, D: Same as in A, but for thalamic cells where burstingmodes are more readily found.Figure 31: Heuristically defined bursting reveals similar STMTDs for cortical neurons(CM). Left: Spatio-temporal motifs for the 3 cortical cells presented in Fig 4.10 considering con-tributions of all spikes from each cell (top motif) versus just the bursting condition for each cell(i.e. only spikes that are preceded by a >500ms silent period; see also Methods). Right: The timecourse of the peak signal in the left hemisphere barrel cortex (see Figs 4.10,4.11 also) for all spikes(blue) and first spikes in a burst (red curves). Both the motifs and time course curves are similarfor both conditions.As natural clusters are not present in STM-space, an alternative ”partitioning” approach wasimplemented to determine how sub-groups of similar-STMs compared to the overall average. Themethod essentially splits single spike STMs into groups of similar ongoing ”network” activity thatare present during active single neuron spiking. The hypothesis was that even though single spikeSTMs are diverse and very different from the average neuron STM (i.e. average over all spikes),perhaps when grouped by similarity the average ”sub-group” STM would be similar to the averageSTM for all spikes and/or reveal further information about the types of ongoing activity thatcorrelates with single neuron spiking. However, the results were similar to those above: subgroups81Figure 32: Heuristically defined bursting confirms similar STMTDs for subcorticalneurons (CM). Left: Spatio-temporal motifs for the 3 of the 4 cortical cells presented in Figure6 considering contributions of all spikes from each cell (top motif) vs. just the bursting conditionfor each cell (i.e. only spikes that are preceded by a >500ms silent period; bottom motif). Right:The time course of the peak signal in the left hemisphere barrel cortex (see Figs 4.11-4.12 also) forall spikes (blue) and first spikes in a burst (red curves). Both the motifs and time course curvesare largely the same for both conditions. The bursting condition for the 4th cell in Fig 4.10 had anSTM with a peak dF/F0 value of ≈0.25% and was excluded from comparison here.of spikes gave rise to different STMs (Fig 14.4D) further supporting the hypothesis that singleneuron spiking occurs during varying ongoing cortical activity. Given that STM sub-groupingrevealed different patterns of calcium activity during single neuron spiking, the next step was todetermine whether it is possible to remove the ongoing or ”spontaneous” activity contributionsfrom single spike STMs. The goal was to isolate the contribution (or ”correlation”) of single spikesto observed mesoscale GCaMP activity.The first step was to replicate the grouping method above and generate several (e.g. 4) func-tional sub-network partitions from the STM distributions (Fig 4.14B). The resulting 4-partitionSTMs were substantially different from each other and the all-spike STM average (with the excep-tion of STM resembling the all-spike average STM; Fig 4.14D; all spike average is the bottom STM).The inter-spike-interval (ISI) distributions during activation of these sub-networks was also similaracross all partitions and had a poissonian distribution - further confirming that the similarity-basedpartitioning method captured similar patterns of spiking and did not selectively group burstingvs tonic periods of activity (Fig 4.14C). The sub-network partition STMs thus contained spikingcontributions to ongoing cortical activity particular to each STM-space partition. The next stepwas to remove the spontaneous component (i.e. component not specifically related to spiking)82Figure 33: Single neurons participate in stereotyped network activity patterns duringall spontaneous activity (CM). A: Five examples of STMs generated from 5 single spikes from asingle neuron reveal substantial variability during spiking and high activation (peak dF/F0>5%).B: Distribution of all single spike STMs (3779 spikes for example neuron) from a single corticalcell visualized in 2-Dimensions using PCA does not reveal natural clusters and is partitioned usingneighbouring distance (i.e. K-Means, n=4) into 4 sub-networks (coloured dots) each with a distinctcentre (larger colour dots). C: Spike rasters for the 4 sub-networks reveal no natural spike-timingrelated clusters: spikes in each sub-network are distributed in time inter-spike-interval (ISI) distri-butions are similar for all 4 sub-networks (green, red, blue and magenta colours) and the all-spikecondition (black colour). D: STMs generated from the 4 sub-networks reveal substantial differencesin the sub-network dynamics (top 4 STMs) with the sum of all STMs providing the average STMpattern (bottom STM) (Note: partitioning the data randomly does not reveal these sub-networksbut patterns similar to the all-spike STM). E: Same as (A) but STM examples are from all possiblespontaneously occurring STM during the recording (9439 possible STMs in a ≈5.2mins recordingat 30Hz). F: Same as in (B) but STM distributions are for all spontaneous STMs with the centresof the sub-networks obtained from the single-cell STM sub-networks in (B). G: Same as in (C)but for all spontaneous STMs. The ISI histograms peak at ≈33ms (i.e. a single frame-interval)indicating that spontaneous STMs group naturally into sub-networks and are dominated by burstsof similar STMs each separated by single frames. H: Same as in (D) but for the 4 sub-networksgenerated from spontaneous STMs (Note: as expected the sum of all spontaneous STMs is ≈0.0%dF/F0, see bottom STM). I: Subtracting the nearest spontaneous sub-network STMs from spikinggenerated sub-network STMs reveals that single cell STM contribution is largely uniform despitespiking occurring during vastly different activated functional networks(see D).83by finding the most similar types of STMs time considering STMs from all spontaneous activityperiods. Thus, the same partitioning method was used but this time spontaneous calcium activitywas used to generate STMs: essentially converting every frame during into an STM (Fig 4.14.E-H).The distribution of spontaneous STMs also has substantial variability (Fig 4.14F). The sponta-neous motifs were then projected to STM-space but were partitioned based on the spike-triggeredSTM sub-network centres derived previously (Fig 4.14B). This guaranteed that spontaneous ac-tivity STMs would be the closest (i.e. most similar) to our spike-triggered STM partitions andallowed for matching each spiking partition STM to an all-frame-partition STM. Because spon-taneous STMs were computed based on single frames, the spontaneous activity motifs showed astrong ISI distribution peak at ≈33ms which is equivalent to the single inter-frame-interval of theimaging system (i.e. 30FPS; Fig 14.4G). This confirms that neighbouring frames were substan-tially more likely to be in the same region of STM-space and are thus grouped together. This isexpected as transitions in GCaMP6 dynamics usually last several frames at a 30Hz sampling rate.Importantly, the sum of the 4 spontaneous STMs yield an STM with an approximately 0% dF/F0value - which is the expected result when averaging all spontaneous motifs (or frames) during arecording (Fig 4.14H bottom STM). The last step was to remove the spontaneously activity STMs(Fig 4.14H) from the single cell spike-triggered sub-networks (Fig 4.14D). The resulting differenceSTMs are very similar to the average STM computed using all spikes.These results suggest that single spike STMs contribute (or correlate with) a stereotyped patternof mesoscale activity that can be recovered by removing very similar STM cortical patterns thatoccur during non-spiking periods. In other words, single neuron spikes correlate with (or participatein) activated functional networks that ride on top of other types of neuronal activity.The STM partitioning method was next applied to two cortical and two thalamic neurons withsimilar results (Fig 4.15). Additionally, a 12 sub-group partition was carried out (Fig 4.15C) andmuch like in the 4-partition cases, subtracting spontaneous activity STMs resulted in differenceSTMs that were largely similar to the all-spike averages. While not investigated further here, itwould be interesting to determine where along the continuum of partitions (i.e. from 1-partition(all spikes) to all possible partitions (single spike partitions) this methodology breaks down. Suchan investigation could use correlation r-values between sub-partitioned STMs obtained after spon-taneous STM removal and the total-spike STMs and characterize the r-value curves to determinewhether they are linear, contain non-linearities or show plateau or asymptotic behaviour.In sum, using various methods for dividing single neuron spikes and STMs it was shown thatall-spike STMs are either preserved or can be recovered from the data. This supports the averagingbased STM methodology used here as representative of a single neuron’s contribution to ongoingcortical activity. The findings further suggest that: (i) cells fire during many different ongoingcortical states; and (ii) that despite the high degree of variability single cell spikes appear tocorrelate with (or possible contribute to) similar overall dorsal cortex activity patterns (as observedin STMs).84Figure 34: Cortical and subcortical examples of STM-space grouping and decompo-sition (CM). A. Two examples of single cortical neuron STMs recovered using the partitionedsub-network approach: i) 4 spike-triggered sub-network STMs; ii) spontaneous sub-network STMs;iii) difference between cell-triggered and spontaneous motifs reveal single cell contributions toactive sub-networks. B. Same as (A) but for two thalamic neurons. C. Same as in (A) but par-titioning data into 12 sub-networks also reveals that average STMs are largely recoverable fromactive sub-networks.85Figure 35: Example of STM vs. Variance STM (CM). A. Spatiotemporal motif of spiketriggered average map of a cortical neuron showing activation near t=0 and following for approx-imately 300ms. B. Spatiotemporal dynamic of spike triggered variance map of the same neuronshowing variability increases post spiking (i.e. t=0) but at partially different ROIs including the≈ RS region (medial-posterior) not present in STMs. Note that the amplitude of the variance issmall (dF/F0 from -0.2% to 0.2%).86An additional test was performed comparing spike-triggered [Ca] motifs against spike-triggered[Ca] variance maps. Because of the high variability across single-spike STMs (and the corresponding180-frame motifs) it was important to consider how variance changes over time (Fig 4.16; see alsoMethods). While spike triggered activity maps were of interest because they showed spike-timelocked responses and revealed distinct spatial patterns, variance maps do not have the temporalspecificity of STMs. An example is provided where an STM is compared to the variance mapwhere both the spatial and temporal differences are clear (Fig 4.16). The variance map does havespatial specificity that is present throughout the recording period but increases following spiking.Interestingly, the RS region (medial-posterior association cortex) shows increased variance across alltimes suggesting that this association area is highly activated and depressed during single neuronspiking - but that such variance is averaged out in STMs (likely due to the methodology). Itshould be noted, however, that the variance map dynamics are much smaller (dF/F0 ≈0.2%) whencompared to the STM range (df/F0 ≈4.4%).STMTD clustering suggests novel single-neuron physiological propertiesThe next step was to investigate whether cortico-cortical and thalamocortical temporal dynamicsrevealed by the functional mapping approach could be further characterized by considering allanimals and recordings (Fig 4.17). Because single neuron imaging motifs (i.e. 6 second, 180 framestacks, 256 x 256 pixels) contained time-varying activation in multiple regions, the approach takenwas to track cortical dynamics in only one specific ROI: contralateral (i.e. left) barrel cortex(L-BCS1; Fig 4.17A; Note: alternative ROI methods were explored but were found to be toonoisy-not shown). L-BCS1 was chosen as the homotopic region corresponding to the electrodeinsertion site (i.e. the electrode was always inserted into right hemisphere barrel cortex and righthemisphere sensory thalamus) and L-BCS1 was activated for most STMs. For this analysis dataacross all animals was combined thus covering a variety of individual track penetrations. Ratherthan choosing a specific (i.e. absolute) pixel location for analysis across all data, the location ofthe maximally activated (or depressed) cortical pixel was selected for each neuron and its activitywas tracked over time. This method was viewed as a more neutral means of comparing dynamicsacross all neurons which could account for variations in recordings: including variations in probeinsertion angle, animal variability, unit yield and variable probe damage (Fig 4.17A - red dots).The method employed to track dynamics at a specific ROI pixel was identical to the STMTDapproach described above. After generating all STMTDs, each STMTD was converted to a high-dimensional vector (i.e. 180-Dimensions for a -3sec to +3sec period, i.e. 180 frame stack). Thedistribution of all 428 STMTDs from both cortical and subcortical cells was then plotted togetherand visualized using PCA (Fig 4.17B; see also Methods). The time courses fell into 3 broad patternsand were clustered using a K-means algorithm (with n=3). Based on their temporal relationshipto spikes, the three patterns had specific characteristics: pattern #1: spike-triggered-excitation;pattern #2 spike-triggered inhibition; pattern #3 inhibition triggered spiking followed by inhibition87(Fig 4.17C). The distribution of these 3 patterns was not evenly divided in cortical and thalamicneuron populations (Fig 4.17(ii)). Specifically, ≈80% of cortical neurons were associated witha purely cortical excitation profile (pattern #1) whereas 20% correlated with spiking-triggeredinhibition (pattern #2). In contrast, only 20% of thalamic neurons associated with post-spikingcortical excitation (pattern #1), while 80% of were associated with cortical inhibition patterns(45% with pattern #2 and 35% with pattern #3). Notably, the purely inhibitory pattern (i.e.pattern #3) was only identified for subcortical neurons. Neither cortical cell depth, subcortical celllocation (e.g. VPM vs VPL), nor cell-type classification (inhibitory and excitatory types; Connorsand Gutnick, 1990; Nowak et al., 2005)) revealed any significant correlations between the patternclusters and cell-classification (not shown).These findings using STMTD clustering suggest that single neurons have unique - likely discrete- properties that relate their spiking to mesoscale activation patterns. As such, STMTD classifica-tion might be a novel intrinsic single-neuron physiological property and that such properties aredistributed across all neurons in cortex (and subcortical regions). Future work with significantlylarger datasets (from multiple cortical and subcortical areas) could further expand these findingsand determine whether additional classes of temporal dynamics are present in cortical cells orwhether a continuum of spatio-temporal relationships is involved.STM hemodynamic corrections are small compared to SNR of calcium activityBecause increases in blood volume are expected to decrease both excitation and emission light, it hasbeen previously suggested that the relative contribution of hemodynamic activity vs neurnal activityneeds to be accounted for (Wekselblatt et al., 2016; Ma et al., 2016). Here additional experimentswere done using a multi-wavelength approach similar to corrections performed in recent studies (seeMethods). Briefly, previous approaches implemented a strobed light presentation using alternatingblue and green lights to capture GCaMP epi-fluorescence (blue illumination) and green reflectance(dim green illumination) in alternating frames. The approach here was similar but methodologicallysimpler to implement: short blue reflected light was simultaneously monitored (with green epi-fluorescence) as a reference signal for fluorescence changes due to hemodynamics. This methodwas implemented as there was some concern the strobed light approach looses some time resolutionas well as requiring camera timing signal programming which are not trivial to implement forall acquisition systems. Additionally, flashing (strobing) lights could potentially stimulate thevisual system of mice (although blocking the excitation/reflectance lights from reaching eyes whilealso preserving access to the eyes for visual stimulation is a potential solution). Accordingly, themethod developed here (Fig 4.18A,B) relies on using a color RGB camera (Picam) (Murphy et al.,2016) which allows for simultaneous acquisition of a short blue light reflected signal (447 nm LEDand 438/24 nm filter near an isosbestic point for hemoglobin), and a green epi-fluorescence signal(GCaMP).In a control experiment it was shown that the short blue reflected light signal (i.e. 438 nm88Figure 36: STMTD clustering suggests STMTD dynamics represent discrete singleneuron physiological properties (CM). A. Example of two STMTDs from pixels within L-BCS1 (left barrel cortex) from a single mouse recording (both cortical and subcortical neuronsSTMTD pixel locations shown). Each recorded neuron STMTD has a slightly different maximumpixel amplitude location, but all fall within the L-BCS1 region. B. STMTD PCA distribution fromall 428 cortical and subcortical neurons recorded from all mice separated using KMEANS (n =3). C (i). STMTD (±SD) classifications from (B) with the number of neurons from cortex andthalamus used for the average are presented in title. (ii) Distribution of STMTD classificationbetween cortical (clear) and subcortical (hashed) neuron generated STMTDs.89reflected light) was strongly correlated with 532 nm green reflected light (r = 0.93; Fig 4.18(D)).For the examples provided (Fig 4.18C-G) the short blue reference signal correlated significantly withapparent blood volume artifacts revealed by parallel experiments using green reflected light imaging.Consequently, the ratio of green over blue signal greatly reduced the blood volume hemodynamicresponse. To determine whether the short blue correction strategy was effective, data from GFP-mmice was examined where green fluorescence signals are not expected to be the calcium dependentas in GCaMP6 mice (Fig 4.18E). The ratio of dF/F0 green signal to blue reflected light signaldF/F0 could then be used to reduce nonspecific signals observed in GFP mice. Consistent withprevious work (Ma et al., 2016; Wekselblatt et al., 2016) blood volume artifacts were greatly reducedusing this strategy. This approach was then applied to GCaMP fluorescence data. Although theapproach was effective at removing smaller non-specific signals in GFP mice, GCaMP6 mice havea much larger activity-dependent signal and only a relatively small apparent contribution of bloodvolume to cortical and subcortical STMs was observed consistent with previous work (Vanni andMurphy, 2014; Murphy et al., 2016; Silasi et al., 2016) (Fig 4.18F,G). Furthermore thalamic STMsand STMTDs still indicated cases where thalamic spiking was associated with cortical inhibition.These comparisons were all performed after signals were converted a dF/F0 value and using agreen/blue weight of 1 (Fig 4.18C). Weights >1.0 yielded some aspects of the kinetics that wereover-corrected. This additional correction dampens non-specific fluctuations associated with bloodvolume changes which were aggravated in the awake state. Overall, the approach supports a shortblue reflected light approach: shot blue light signals can be used as a surrogate for a green reflectedsignal without the need for alternating cycles of light. Thus the short blue reflected light signalis an equivalent strategy similar to previous approaches (Sirotin and Das, 2010; Ma et al., 2016;Wekselblatt et al., 2016).After implementing the corrective strategy to additional RGB Picam recordings, only minorchanges were observed in spike-triggered map activity consistent with control investigations offunctional connectivity or task-related connectivity in GFP animals in experiments done previously(Vanni and Murphy, 2014; Murphy et al., 2016; Silasi et al., 2016). While the multi-wavelengthcorrection strategy yielded some changes to the STMs, the corrections only altered cortical orthalamic maps or dynamics peaks by less than 10% dF/F0. Furthermore, notable features, suchas some areas showing apparent cortical inhibition (reductions in calcium activity) were preservedin the corrected maps. Given that the correction does not have a large impact and failed tochange the appearance of the maps, the original figures were left intact and not modified further(Note: originally acquired data also cannot be corrected offline as it was captured using a singlewavelength).STMs are not affected by (minor) body movementsThe primary goal of this study was to assess cortical functional connectivity based on coincidencebetween individual neuron spiking and ongoing spontaneous activity in awake and anesthetized90Figure 37: STM hemodynamic corrections are small compared to SNR of calciumactivity (MV).91Figure 37: (continued from previous page) A. Diagram of the experimental setup used to eval-uate the contribution of blood volume to the GCaMP green fluorescence. Local blood volumewas evaluated by measuring the change in short blue (447nm) and green reflectance (525nm, forvalidation control only, see panel B) while hemoglobin is known to absorb blue and green light.Green fluorescence of GFP or GCaMP was excited by using 473nm excitation light when 525nmlight was turned-off (and 447nm light turned-on). Short blue reflectance and green reflectance orfluorescence was imaged using a RGB color camera. B. Spectrum of excitation and emission ofGCaMP (black curves, similar for GFP) and absorption of hemoglobin (red). The transmission ofthe blue and green channel depends on the window of the triple band-pass filter (dash gray boxes)and the transmission of the blue and green channels of the CCD camera (dotted blue and greencurves). C. Spontaneous activity of green and short blue reflectance (cross-correlation: r=0.93).Blue channel DC was adjusted to fit with green channel baseline and both channels were low passfiltered (0.033Hz). D. First and Second lines: cortical motifs generated from green reflectance (us-ing green LED) and short blue reflectance (using short blue LED) from a GCaMP mouse. Thirdline: Ratio of the two motifs green over short blue. Left: motif values within BCS1 for greenand short blue channels (cross-correlation: r=0.88±0.04, n=7; present example: r=0.92) as wellas ratio. E. First and Second lines: cortical motifs generated from GFP fluorescence (using blueLED) and short blue reflectance (using short blue LED) from a GFP mouse. Third line: Ratio ofthe two motifs green over short blue. F,G. Same as C but for a GCaMP mouse and for corticaland subcortical motif.mice. However, in awake (head-fixed) mice neuronal activity is rarely entirely spontaneous andperiods of volitional movement are interspersed within largely quiescent but longer intervals. Ob-servations indicated that limb twitches as well as tail and facial movements were present in therecordings. In order to evaluate the impact of body movement, behavioural video recording wasused to track movement. STMs were then generated from spikes during periods of quiescence andcompared with STMs from all spikes (4.19). Periods of quiescence and movement were identifiedby measuring behaviour collected simultaneously with neurophysiological data (see Methods, Fig4.19A). STMs generated from periods of quiescence did not differ from those generated using allspike (see Methods, Fig 4.19B). This suggests that movement either minimally contributes to STMsor contributes diverse patterns which are generally averaged out by the methodology. The analysisalso indicated that periods of high movement were relatively rare in awake head-fixed mice underthe conditions employed here and contributed negligibly to overall maps. Therefore, brain imag-ing activity obtained in awake states is mostly indicative of a quiescent, awake state and is notprimarily movement-related activity.STMTD diversity is substantially greater in subcortical than cortical neuronsLastly, all neuron motifs across all recordings were compared to assess large trends across the twotypes of data. Because of the large number of neurons across all recordings it is not possible to plotall motif or STM profiles side-by-side for such a comparison (i.e. the motifs would be small and92Figure 38: Body movement contributes minimally to STMs (MV). A. Image insets: pic-tures of the frame average and standard deviation showing the location of movement over oneentire recording. Yellow box: region of interest used to quantify movement. Graph: Black curveis movement density calculated for each frame by measuring the average of the absolute gradientwithin the region of interest (yellow box). Standard deviation (std) and median of the profile werecalculated and period of quietness and movement are identified by selecting periods of time below[median+std/10] (green) and above [median+std] (red) respectively. B. Motifs generated for allspikes (top, black curve) and spikes only from quiescent periods (bottom, green curve). C. Maxi-mum and positive peaks amplitude for motifs from all and quiescence periods showing no overallchange trends (paired t-test: p=0.108 and 0.431 respectively, n=31).93illegible). Accordingly, a method was devised to convert the 180-frame, 256x256 pixel image stacksfor each neuron to several 2-Dimensional (time vs. amplitude) STMTDs which capture the mostactivated (and easy to identify) ROIs for comparison across all data (Fig 4.20). Thus, STMTDsfor 8 major ROIs across all mice and recordings were plotted by aligning the peak activation (red)or peak depression (blue) in various ROIs (Fig 4.20). This allows for a comparison of temporalrelationships for all data that is initially 3-Dimensional using only 2 dimensions. As observed forindividual neuron STMTDs, cortical neurons STMTDs generally peaked in excitation in barrel andmotor cortex (BCS1 and M1) and were followed by a return to baseline or depression in barrel cortex(Figure 4.20A). In contrast sub-cortical neurons were linked to diverse cortical activity profiles, inparticular longer depression across multiple cortical ROIs (Figure 4.20B).DiscussionUsing wide-field spontaneous calcium imaging data, mesoscopic cortical maps defined by the spik-ing activity of individual cortical and sub-cortical neurons were characterized. The results demon-strate that STMs reveal functional cortical architecture related to the activity of individual cor-tical and subcortical neurons. STMs for cortically recorded neurons reflect the cortical state inmono-synaptically connected areas during spiking activity. STMs of sub-cortical neurons havehigh variation than maps attributed to spiking cortical neurons. For example, sub-cortical STMpatterns for neighboring neurons were more diverse than those of neighbouring cortical neurons,and were less likely to match intra-cortical consensus activity patterns defined using seed pixelcorrelation mapping. Sub-cortical-neuron derived STMs revealed multiple areas of activation andmultimodal kinetic behavior, while intra-cortical spiking neuron networks were simpler in struc-ture and kinetics. Furthermore, spiking sub-cortical neurons reflected diverse cortical multi-phasicexcitation-inhibition timing patterns that were reflected in dynamic STMTDs. In contrast, mostspiking cortical neurons were linked to a single phase of cortical excitation.Event triggered mesoscale mappingPreviously, spike-triggered averaging of local field potentials has been used to investigate of howsingle neurons in visual cortex were linked to on-going state-dependent activity (Nauhaus et al.,2009), however, this work only examined such correlations locally (i.e. within visual cortex) anddid not assess regional connectivity using imaging or investigate differences with individual sub-cortical neurons. Other similar applications where single neuron spiking was recorded and related tospontaneous activity using calcium imaging have been restricted to in vitro brain slices (Aaron andYuste, 2006)). The study presented here extends previous in vivo work that assessed spike-triggeredmapping using voltage sensitive dye imaging (Arieli et al., 1995; Tsodyks et al., 1999) to encompassa larger spatial scale, higher density electrode arrays, awake recordings, and selective geneticallyencoded indicators of activity. While being important seminal findings (Arieli et al., 1995; Tsodyks94Figure 39: ROI-specific dynamics reveal cortical stereotypy and subcortical diversity(CM). A. Top: Normalized STMTDs (as in Fig 4.17) from multiple ROIs (HLS1, FLS1, BCS1,RS, V1, M1, PTA and ACC, see Table 1) for 255 cortical cells. Each horizontal line represents asingle neuron’s STMTD in each of the eight ROIs normalized to the overall maximum or minimumactivation. Bottom: average (± std) of STMTD within each ROI for all neurons. B. Same asA, but for all thalamic neuron generated STMTDs reveals thalamic STMTDs are more diverse,less temporally precise, and contain longer depression epochs. (Note: these results are from awakemice).et al., 1999), previous spike-triggered averaging work was largely confined to the visual system,performed under anesthesia, and was unable to define how multiple brain areas interact. Theapproach here is most analogous to event-triggered MRI imaging from the standpoint of largerspatial scale (Logothetis et al., 2012) where it was observed that during hippocampal ripple statesthat cortex exhibited net positive bold responses and thalamus net negative BOLD responses. Thisanti-correlation is consistent with observations of thalamic spiking activity in the current studycorresponding with cortical temporal dynamics exhibiting slow depression of calcium signals andmay point to a larger coordinated network involving other brain structures.The current study has advantages over MRI signals which lack temporal resolution and can be95more difficult to relate to neuronal activity than GCaMP signals that are isolated within excitatoryneurons of GCaMP6f mice using specific promoters (Chen et al., 2013b; Madisen et al., 2015).Unique to the approach presented here is the ability to assess the functional connectivity andtemporal dynamics between specific sub-cortical neurons and areas of cortex not predicted byprevious knowledge such as linkages between thalamic neurons and cortical state as defined byGCaMP signal dynamics. The aproach provided here can be further refined when more selective cre-dependent CGaMP6f transgenic mice are available to allow for the expression of calcium indicatorsin particular neuron types (Madisen et al., 2015). Furthermore, 2-photon microscopy could be usedto provide information about behavior of individual cells within the context of larger maps (Chen etal., 2013a; Guo et al., 2014; Okun et al., 2015). Because of the high sensitivity of the indicator andthe possibility of measuring the activity of dozens of single-units using multiple electrode channelssimultaneously, a large number of functional connections can be mapped in only a few minutes ofrecording. Although only a single electrode shank was used here, future experiments could furtherincrease the yield using higher density electrodes and multiple shanks to collect spikes from moreneurons simultaneously.Cortical and sub-cortical neuron derived maps reflect different functional rolesThe results indicate that neocortex contains discrete subdivisions where individual spiking corticalneurons generally belong to spatial-temporal maps that follow a consensus function that can bedefined using correlation as in previous work (Mohajerani et al., 2013; Chan et al., 2015). Incontrast, single thalamic neurons tend to fire when cortex is in more kinetically-diverse states. Themore diverse dynamics between thalamic neurons and cortical mesoscopic networks indicate thatsub-cortical thalamic neurons play an instructive role with respect to cortical state, particularly withrespect to feed-forward cortical inhibition (Stroh et al., 2013; Urbain et al., 2015), whereas corticalneurons may serve as relay endpoints or amplifiers (Douglas et al., 1995). A better understanding ofthese dynamics may yield insight into how disorders, such as epilepsy, and dementia, emerge wheninteractions between brain areas are disrupted (Paz et al., 2013; Busche et al., 2015; McCormicket al., 2015). The diversity in sub-cortical spiking derived maps may also reflect differing receptivefield properties in thalamus and cortex based on varying types of functional convergence describedpreviously (Miller et al., 2001). Indeed, in the somatosensory whisker barrel system, evidence for”ensemble convergence” has been described where input from the thalamus can extend outsideof the boundaries of the corresponding cortical receptive field (Simons and Carvell, 1989; Lindenand Schreiner, 2003). The larger diversity of maps derived from the spiking of different thalamicneurons may be expected because of the smaller size of thalamic nuclei compared to the cortexand the recording of thalamic neurons from more varied structures. Another potential source ofvariance may arise from the diversity of thalamocortical impulses that can be comprised of patternsof activity ranging from tonic, ’relay’ transmission consisting of high regular rates of firing toburst-like activity where firing rates are low and interspersed with high-frequency events (Steriade96and Llina´s, 1988; McCormick and Feeser, 1990; Sherman and Guillery, 1996). Thalamic burstingcan powerfully activate neocortical circuits and has been suggested to serve a ”wake-up” signalto sensory cortices (Sherman and Guillery, 1996; Swadlow and Gusev, 2001). While segregatingrecordings into various firing configurations did not reveal substantially different STMs or STMTDs,it may be that other properties (not accessible here) such as specialized synapses may account formap diversity. Interpreting these results is also caveatted by the mesoscale resolution and calciumdynamics present in the recorded data.Applications of spike-triggered mappingMapping the functional connectivity of spiking neurons is important for understanding brain func-tion and finding therapeutic targets for brain stimulation or brain machine interfaces. Identificationof networks linked to individual neurons may help reveal the mechanism of brain machine inter-faces where key signals are often attributed to only a small number of neurons (Stanley et al., 1999;Serruya et al., 2020; Taylor et al., 2002; Guggenmos et al., 2013). Other applications include under-standing of how small groups of epileptic neurons (Paz et al., 2013) are coupled to brain networksleading to seizure propagation. Given that reciprocal connections between mesoscale structures arewidespread, the cortical maps associated with a spiking neuron in a sub-cortical structure such asthe sub-thalamic nucleus may provide clues as to how cortical activity can be manipulated to affecta sub-cortical target. This hypothesis can be tested by recording sub-cortically using electrodearrays while stimulating regions of cortex that show coincident STMs using Channelrhodopsin-2or other opsin-activity sensor pairs (Lim et al., 2012; Rickgauer et al., 2014; Zou et al., 2014;Abdelfattah et al., 2016; Kim et al., 2016).Extension to behaviorally driven activityThe same approach applied here could be extended to generating STMs during specific behaviors.However, major shifts in area map boundaries as defined during spontaneous activity are not ex-pected, as these are largely determined by anatomy (Mohajerani et al., 2013; O’Connor et al., 2013;Oh et al., 2014; Zingg et al., 2014) and in the case of sub-cortical neuron maps (HPF for exam-ple) poly-synaptic connections. During behavior more nuanced changes in the weighting, timing,and frequency-dependence of STM networks might be present during an active task. It is alsopossible that specific behaviors will reveal the superposition of multiple cortical motifs associatedwith progression through a task. STM mapping of cortex would be particularly interesting in thecontext of rhythmic whisking-related centers within the medulla and thalamus and their linkage tocortical maps within barrel-motor areas (Moore et al., 2013; Deschnes et al., 2016; Sreenivasan andPetersen, 2016).97SummaryIn sum, single neuron spiking activity reliably reflects mesoscale activity transitions within mousecortex. STMs together with connectomic information (Hunnicutt et al., 2014; Oh et al., 2014; Zingget al., 2014; DeNardo et al., 2015) may help bridge the gap between single neuron function andlarger networks. The findings presented herein reveal that thalamic neurons interact with cortexduring specific state transitions that are reflected by typical consensus cortical neuron behavior. Thepresence of such long-range relationships in spontaneous activity may suggest new opportunitiesand routes by which brain stimulation and inhibition can be applied to affect synaptically connectedareas.GCaMP6 mapping of spontaneous activity of auditory and visualcortex neuronsBackgroundThe findings provided above (also published as Xiao et al., 2017) related primarily to barrel cortexand thalamic neurons. There were a number of limitations in that study including that the corticalneurons recorded come solely from barrel cortex and the use of low density electrodes (singlecolumn, 16 channel 100µm vertical spacing electrodes) limited the spatial analysis and spike sortingquality. In this section additional data is presented to (briefly) investigate whether the STMmethods implemented in the first section can be applied to auditory and visual cortex neurons. Thedata also comes from higher-SNR GCaMP6s mice recordings and was obtained using high-densityextracellular probes in visual and auditory cortex. Part of the work in this section was presentedpreviously (Mitelut et al., 2016). All experiments and analysis in this section were carried out bythe author.ResultsUsing methodology described above (and in Xiao et al., 2017) cortical neuron STMs were computedfor recordings from the visual and auditory cortices of 2 GCaMP6s mice (Fig 4.21-23). The datawas acquired using a Raspberry Pi 3 camera with custom Python code used for tracking singleframe times for offline alignment with the imaging system (see Appendix ). The findings confirmthat all cortical neurons (with >25 spikes in a recording period) can generate high-SNR STMs. Inparticular, stable STMs (i.e. qualitatively discrete maps, many of which were similar to consensusmap) were obtained using as little as 25 spikes to as much as several thousand. These findings(discussed in detail below) confirm that cortical neurons excluded in the previous section (due tolow SNR) were likely due to experimental setup challenges and not cortical physiology.Auditory cortex neuron motifs showed diversity across depth and almost all contained pre-spiking inhibition in medial-posterior areas prior to spiking (Fig 4.21-left). This depression com-98Figure 40: Single neuron STMs - auditory and visual cortex.99Figure 40: (continued from previous page) Left: single auditory cortex neuron motifs plottedby depth starting with most superficial neurons (top) and normalized to dF/F0 peak (left values)reveal ≈500ms cortical depression precedes most spiking. The auditory cortex motifs reveal theinvolvement of a medial area being co-activated with auditory cortex areas (lateral-posterior acti-vated areas. Right: single visual cortex motifs plotted by depth reveal substantially stereotypedSTM patterns accross all layers of cortex.Figure 41: Auditory cortex neurons - firing rates and STM-space distributions. A.Firing rate distributions for recorded auditory cortex neurons. B. Peak dF/F0 for auditory cortexneurons. C. STMTD-space distributions (shown using PCA) for all auditory cortex neurons (viridiscolour scheme) against controls using time-scrambled (red) and the origin (i.e. 0-vector; black) showdiversity of activation dynamics.ponent is larger than those observed in barrel cortex where almost no pre-spiking inhibition wasobserved for cortical neurons. This longer pre-spiking depression suggests auditory cortex neuronsmay form an additional STMTD class of their own - possibly suggesting that neurons in differentcortical areas may have different temporal relationships with dorsal cortex. Interestingly, the medialcortex area where pre-spiking depression occurs is also co-activated with the auditory cortex areafollowing spiking. This suggests that auditory cortex neurons may have mono-synaptic correctionswith medial cortex neurons - a functional relationship not fully characterized previously (to theauthor’s knowledge).In contrast, STMs for visual cortex neurons were very similar across all layers. Assuming that100the recordings were correctly aqcuired and were not dominated by aberant activity in Cre-dependentGCaMP6s mice (Steinmetz et al., 2017), these findings suggest that visual cortex neurons form partof functional networks that - at the mesoscale - appear to engage the entirety of visual cortex duringexcitation and depression. However, it must be noted that both of the recordings (i.e. auditoryand visual) reported here come from urethane anesthetized mice where large numbers of UP-statetransitions underlie most spiking activity. Accordingly, it is expected that single neurons duringsuch synchronized cortical states (see Chapter 5 for discussion of such states) spike simultaneouslywith large parts of cortex.Firing rate and dF/F0 distributions did not reveal depth dependent trends in the auditorycortex neurons (Fig 4.22A,B). And distributions of STMTDs for each neuron motif (see Fig 4.17)against randomized STMTD traces qualitatively confirmed the variability observed in the auditorycortex motifs. Visual cortex neuron firing rates and dF/F0 distributions were somewhat higherindicating a potential issue with the cortical recording in that mouse (i.e. visual cortex neuronsshould fire more sparsely than other areas). The STMTD distributions were strikingly narrowreflecting the qualitative similarity observed in the STMs (Fig 4.21-right).While no clear clustering was observed in the distributions of the STMTDs, this was expected asthe neurons were recorded in cortex during anesthetized states (i.e. barrel cortex neuron STMTDsfell two classes with one class constituting ≈80% of the total dynamics).SummaryIn sum, the single-neuron spike-triggered cortical motifs of auditory and visual cortex neuronsconfirm that the STM methodology applies to all cortical recorded neurons (with >25 spikes).The better temporal resolution of the imaging system (i.e. software) used here coupled with themuch higher-density extracellular probes validate the general applicability of the spike-triggered-mapping procedure discussed in the previous chapter. Interestingly, the variability in GCaMP6sauditory cortex motifs across depth not observed in visual cortex motifs suggests future avenues ofinvestigating mesoscale maps and differences across areas of cortex.VSD mapping of spontaneous activity of auditory and visualcortex neuronsBackgroundThe findings provided in the two previous sections (also published Xiao et al., 2017, Mitelut et al.,2016) were obtained from recordings in GCaMP6 mice. The limitations in GCaMP6 imaging arediscussed above and include the relatively slow rise (≈100ms) and decay time (≈100-200ms; Chenet al., 2013b) of calcium dynamics. Additionally, intracellular [Ca] is a partially biased reporter ofspiking activity: [Ca] increases exponentially during bursting as opposed to tonic spiking activity.101Figure 42: Visual cortex neurons - firing rates and STM-space distributions. A. Firingrate distributions for recorded visual cortex neurons. B. Peak dF/F0 for auditory cortex neurons.C.STMTD-space distributions (shown using PCA) for all auditory cortex neurons (viridis colourscheme) against controls using time-scrambled (red) and the origin (i.e. 0-vector; black)In this section additional results are provided from spike-triggered mapping explored in VSDpreparations. VSDs that report membrane depolarizations have been used for over 40 years (Cohenet al., 1974) and in contrast to intracellular calcium sensors such as GCaMP6, membrane potentialreporters such as VSDs are ideal for reporting the state of a neuron as they have fast responsesand can report sub-threshold activity (Peterka et al., 2010). However, most VSDs currently in use(and the one used herein) suffer from low-SNR issues: the dF/F0 signal generally peaks around0.5% which is low when compared to GCaMP6 signals that can reach up to 50% dF/F0 values.However, averaging over many events can improve SNR substantially.The data presented in this section comes from recordings used for analysis in other chapters(see Chapter 5) and is presented as a brief - but complementary - section that shows the power ofspike-triggered mapping at the mesoscale can overcome the lower SNR of VSDs. Additionally, high-density probes are used and it is shown that large-dorsal cortex craniotomies can yield meaningfulVSD STMs (i.e. such crainotomies are not required for GCaMP6 recordings that were carried outthrough the intact skull).The experiments and analysis in this section were carried out by the author.102Figure 43: Single neuron VSD STMs - visual cortex. Visual cortex VSD motifs aligned bydepth (top: more superficially recorded neurons) reveal variations in spatial patterns and temporalactivation times with strong co-activation of medial cortex often dominating the mesoscale motif.103ResultsUsing the spike-triggered-averaging method described above, VSD imaging (150FPS, i.e. 6.7msresolution) motifs were computed for single neurons recorded in the visual and auditory cortex of2 mice. Additionally, motifs from mouse subcortical neurons (unspecified area) are reported. Allneurons recorded in visual cortex which had >100 spikes during a recording period generated motifswith temporal and spatial specificity (Fig 4.24-26; partial selection of neurons). The visual cortexneuron motifs showed activation from ≈-100ms to +100ms following spiking. Overall, neuronsrecorded across all layers showed medial co-activation with some visual cortex specific activation(with left hemisphere activation being stronger than right hemisphere likely due to electrode damageto the right hemisphere).The VSD activation patterns in visual cortex show substantial variation when compared tothe GCaMP6 patterns (Fig 4.21-right) which were stereotyped in time and spatial dynamics (Fig4.24). In addition, the vast majority of visual cortex neurons show substantial co-activation ofmedial (anterior and posterior) cortex during visual cortex spiking. Because VSDs report sub-threshold activity, VSD imaging provides additional details about the preferred depolarizationprofile of single visual cortex neurons during large meso-scale depolarization events (e.g. UP-statetransitions) that occur during anesthetized and/or synchronized state recordings explored here.However, the diversity present suggests that large-scale investigations of single-neuron VSD mapsin visual (and other cortical areas) may reveal a much wider range of preferred spatio-temporalmesoscale activity patterns (i.e. motifs) than those observed in GCaMP6 mice. Such patternsmay constitute a comprehensive representation-space of mono-synaptic connectivity between singleneurons and dorsal cortex. VSD activation patterns in mouse auditory cortex (single hemispherepreparation; see also Methods) also indicate differences across cortical depth. Auditory cortex VSDSTMs also confirm the involvement of a medial cortical area during spiking that was observed inGCaMP6 STMs (4.21-left). This co-activated medial area also seems to often lead the auditoryarea activation suggesting a possible causal relationship.In contrast to cortically recorded neurons - but consistent with GCaMP6 findings in thalamicneurons - subcortical neurons (likely recorded from auditory thalamus, i.e. medial geniculate nu-cleus and surrounding nuclei) had a wider range of motifs (Fig 4.26). There were several strikingpatterns not observed in GCaMP6 motifs including: peak motif activation that was offset from t=0by up to 30-60ms; motifs that had anterior-posterior oscillation cycles on top of which auditorycortex-like motifs occured; and even anti-motifs where medial depression occurs during auditorycortex activation. There are also some purely depressive motifs which were also observed in thalamicneuron GCaMP recordings (Fig 4.17C). Two example VSD motifs were expanded to full-temporalresolution (150FPS, 6.7ms per frame; Fig 4.27 and 4.28).104Figure 44: Single Neuron STMs - Auditory Cortex Neurons. Auditory cortex VSD motifsaligned by depth (top: more superficially recorded neurons) reveal similar spatial patterns butdifferent temporal profiles.SummaryIn addition to GCaMP6, VSD optical mapping of single neuron spiking activity is a viable methodfor investigating functional connectivity across spatio-temporal scales. The VSD single neuronmotifs reveal additional insight and constitute a complementary method for investigating corticalfunction.105Figure 45: Single Neuron STMs - Subcortical Neurons. Subcortically recorded neuronsaranged by order of depth show substantial diversity.106Figure 46: Example (#1) subcortical neuron VSD STM - complete motif. Exampleof single subcortical neuron STM computed at full acquisition resolution (i.e. 150FPS, 6.7ms perframe) shown from -1sec to +1sec following spiking reveals large-scale cortical depression with noclear cortical activation patterns centred on spike time, i.e. t=0ms.DiscussionThis chapter explores single-neuron triggered meso-scale calcium and VSD activity patterns. Thefindings of spike-triggered-mapping are discussed at length above (see Discussion in first section).Future work should focus on a number of improvements in both experimental acquisition andanalysis. First, cortical states were not fully explored in these datasets and a greater effort shouldbe made to compute maps as a function of synchrony index (see Saleem et al., 2010; see also Chapter5). Higher spatial resolution could be implemented using two-photon microscopy to compute spike-triggered-maps and higher signal-to-noise ratio VSDs could reveal substantially more interesting107Figure 47: Example (#2) subcortical neuron VSD STM - complete motif. Example ofsingle subcortical neuron STM computed at full acquisition resolution shows spiking correlates withthe onset of a posterior depression pattern which evolves into a partial auditory-cortex neuron-likemotif before returning to the anterior-posterior depression/activation shape.dynamics present at the mesoscale. Lastly, blood volume corrections (e.g. Ma et al., 2016) need tobe implemented in a more formal manner to address potential confounding results.108Multiple Classes of UP-stateTransitions and Non-Stationary SingleNeuron Firing Order Revealed byLocal-Field-Potential EventClustering During Slow Oscillations inMouse and Cat Cortex“Most of the information [in cortex] turns out to be encoded by the firing rates of theneurons, that is by the number of spikes in a short time window.“Rolls and Treves, 2011“When rephrased in a more meaningful way, the rate-based view appears as an ad hocmethodological postulate, one that is practical but with virtually no empirical ortheoretical support.“Brette, 2015“... transient, sequentially organized packets of activity could constitute a basicbuilding block of the cortical code. ...cortical activity is composed of coherent andstructured packets of population activity lasting a few hundred milliseconds. ... the finetemporal structure of packets is largely conserved across spontaneous andstimulus-evoked conditions, and across different brain states, and ... variations on acommon sequential structure can encode information about sensory stimuli. “Luczak et al., 2015Understanding how spiking cortical neurons represent information is one of the most centralquestions in all of neuroscience. The rate coding, also known as the ”firing rate”, doctrine hasdominated theoretical and experimental work since the early part of the 20th century and suggeststhat neurons represent information through their firing rates (Adrian, 1926). Yet an increasingnumber of both theoretical and experimental studies over the past two decades support spiketiming based codes (or the relative coordination of spikes) as important to information processingespecially during stimulus presentation (Abeles, 1991; Izhikevich, 2005; Singer, 1999; Thorpe etal., 2001; Deneve, 2008). This recent work has prompted some to argue that firing rates have a109merely correlational role in information processing and only spike timing neural codes can be causal(Brette, 2015).One particular line of studies that supports the importance of spiking timing has hypothesizedthat the cortical circuits and neuronal mechanisms engaged by slow oscillations (Steriade et al.,1993a,c; Steriade and Amzica, 1998) occuring during slow-wave sleep (SWS), quiet awake andanesthetized states in mammalian cortex - could be the unifying paradigm for the study of corticalfunction (Sanchez-Vives et al., 2017, see also Neske, 2016). This hypothesis is supported by severalrecent studies from Arthur Luczak (e.g. Luczak et al., 2007, 2009; Luczak and Bartho, 2012;Luczak et al., 2013, 2015) discussed at length below, but was initially articulated almost twodecades ago when it was suggested that ”sensory stimuli trigger K-complexes by addressing the samecortical machinery that produces spontaneous K-complexes during the slow (<1 Hz) oscillation...sensory evoked K-complexes may be the exception rather than the rule” (Steriade and Amzica,1998). As such, studying spontaneous activity during slow-oscillations and what are now calledUP-state transitions (aka ”K-complexes” Steriade et al., 1993a,c) in active cortical areas (e.g. ratsomatosensory cortex) revealed that some (i.e. high-firing rate ) neurons are activated in specificfiring orders (i.e. ”packets”; see also Luczak et al., 2015) during both spontaneous and stimulusevoked periods. Such findings of a cortical neuron firing order pose a significant challenge tofiring rate coding theories which argue that only the firing rate of a neuron is important for therepresentation and relaying of information (Adrian, 1926; Shadlen and Newsome, 1994; Rolls andTreves, 2011).This chapter is aimed at extending previous findings of firing order to the visual cortex of catand mouse - areas not previously investigated by others likely due to the sparse firing properties ofvisual cortex neurons. A novel method is developed similar to spike sorting which can identify andclassify UP-state transitions in cat V1 and mouse sensory cortex using local field potential (LFP)from extracellular recordings. Additionally, it is also shown that >90% of all neurons exhibit strongpeaks in spiking during LEC-defined UP-state transitions even in visual cortex - many more neuronsthan in previous findings in rat somatosensory and auditory cortex. Perhaps most interestingly, itis shown using multiple methods that some neurons change their relative firing orders over periodsof many minutes (e.g. 30-120minutes or longer).These findings provide a novel, single-neuron independent and temporally precise definition ofUP-state initiation that identifies multiple types of UP-states and enables the study of single neuronlatencies during UP-state transitions across all cortical areas.There are a few notes to be made before discussing the results. First, in addition to cat experi-ments (see below), approximately 27 mice (mostly GCaMP6s) were used to carry out simultaneouselectrophysiology and widefield optical imaging experiments (all experiments carried out by theauthor). These experiments were specifically made to investigate issues arising within this project(see below). Unfortunately, a May 2017 bioRxiv pre-print article (Steinmetz et al., 2017) claimsthat all (or most) Cre-Emx1 dependent GCaMP6s mice have large amplitude aberrant epileptiform110Figure 48: Power spectrograms and stimulus annotations - Cat C3. 4 tracks fromone isoflurane/N20 anesthetized cat (C3) with power spectrograms (middle), stimulus annotations(right) and synchrony index (left) reveals all tracks have prolonged synchronized state periods(blue trace sections in synchrony index) compared to desynchronized states (black trace sections).A synchrony index threshold value of 0.5 was used to define most synchronized state recordingperiods (dashed black line in synchrony index traces; see Methods for stimulus descriptions; seealso main text).11102468101214112Figure 49: (continued from previous page) Power spectrograms and stimulus annotations -Cats C1 and C2. 2 tracks from one isoflurane/N20 anesthetized cat (C1) and 2 tracks from anotherisoflurane/N20 anesthetized cat show that the recordings come from mostly desynchronized corticalstate recording periods (black trace sections) with relatively brief synchronized state periods.LFP electrical activity events in cortex. As the current chapter focuses on identifying and groupingstereotyped, large amplitude electrical LFP events occuring during UP-state transitions in cortex,the grouped events are likely to overlap with aberrant electrical activity events reported in thisrecent article. Accordingly, until the issue of Cre-EMX1 dependent GCaMP6s aberrant electricalactivity is resolved (i.e. the publication passes peer review and the findings are validated), several ofthe sections in this chapter have been limited to cat electrophysiology recordings with only limitedexamples from GCaMP6 mouse recordings (as indicated in the main text below). One additionalawake mouse cortical recording was provided by an external lab (see Methods) to complement theexisting results (Figure 5.25).Furthermore, in contrast to the GCaMP6 mouse experiments which were designed specificallyfor this chapter to have long spontaneous activity periods with periodic identical stimulus periods,cat recordings analyzed in this chapter come from ”unbiased” experiments: i.e. experiments carriedusing using randomized stimulus order. Cat experiments were made by Nicholas V. Swindale andMartin Spacek between 2009 and 2012 (described in Swindale and Spacek, 2015; see also Methods)and come from a variety of cat anesthetic preparations (see Methods) and cortical states (Fig-ure 48-Figure 51). Importantly, they have largely randomized visual stimulus periods which werenot ideally structured for use in this Chapter (e.g. neural activity during sequential spontaneousand natural scene recordings and synchronized cortical states). As can be seen in the power spec-trograms and stimulus order (Figure 48-Figure 51; discussed below) only some of the cat recordingscontain synchronized states that span multiple hours and contain longer spontaneous activity pe-riods. Accordingly, the focus in the later sections of this chapter (e.g. sections evaluating firingrate order across recording epochs and various stimulus types) was dictated in part by the natureof the available cat data (i.e. varying cortical states and stimulus periods).The Chapter begins with a brief overview of SWS, UP-state transitions as measured using singleneuron recordings and LFP, and recent work by Arthur Luczak on neuronal firing order during UP-state transitions in rat somatosensory and auditory cortex. Next, the method for clustering LFPevents into LECs is described and examples from cat and mouse recordings are provided. LECsare shown to have discrete CSD profiles that can be grouped across tracks and animals. Single unitlatencies during UP-state transitions are then identified using peaks in LEC-triggered histograms.Two examples of LEC-triggered dorsal cortex maps are provided using VSD imaging in mice asmesoscale correlates of UP-state transitions. The remaining sections focus on tracking latency andspiking distributions during UP-state transitions as well as tracking spike order within and outsideUP-state transitions. First, it is shown that latency peaks and distributions for some neurons113Figure 50: Power Spectrograms and Stimulus Annotations - Cat C4.114Figure 50: (continued from previous page) 3 tracks from one propofol anesthetized cat (C4) showsthat the recordings come from mostly desynchronized cortical state recording periods (black tracesections) with relatively brief synchronized state periods.change systematically over time. Next, state space analysis is used to track latency for selected(stable) neurons in randomly selected recordings to show changes over time. In-depth analysis using6 multi-hour recordings from 2 cats are further used to show systematic changes in peak latency anddistributions (i.e. changes in firing order) over periods of a few hours. A novel, ordered synchrony,metric is developed and it is shown that spiking order is also partially preserved outside UP-statetransitions. Lastly, spiking order and synchrony metrics from recent studies are applied to somedatasets to reveal similar gradually changing firing order across neurons. All analysis, figures andtext in this chapter were made by the author.BackgroundSWS and UP/DOWN-statesOver the past several decades, single neuron patch clamping methods have enabled the tracking ofsomatic membrane potentials of mammalian cortical and thalamic neurons across various animalstates including anesthetized, naturally sleeping and awake conditions (Sakmann and Neher, 1984;Steriade et al., 1993a). These single neuron studies have identified two states: a desynchronizedstate present during awake and attending periods and rapid-eye-movement sleep (REM) duringwhich neurons fire largely independently of each other (Harris and Thiele, 2011); and a synchro-nized state present in slow-wave sleep (SWS), quiet waking and anesthesia where neural activity hasa slow oscillation (0.2Hz-0.9Hz) with individual neurons periodically oscillating between a depo-larized (spiking) state and a hyperpolarized state (non-spiking) known as UP- and DOWN-states,respectively (Steriade et al., 1993a,c; Sanchez-Vives and McCormick, 2000; McCormick and Yuste,2006; Neske, 2016; Sanchez-Vives et al., 2017). (Note: while single neuron membrane patchingstudies provided novel findings, behavioural and cortical states have been studied for almost 100years using electro-encephalography -EEG and other methods). UP- and DOWN-states are knownto be present in cortex, thalamus, hippocampus, striatum and cerebellum and could be ”the defaultactivity pattern of the entire cortical mantle”. That is, all other types of cortical activity under-lying stimulus processing and even complex behaviour may be instantiated in cortex as a type ofUP-DOWN state transition (Neske, 2016). UP-state transitions in particular (i.e. the transition ofneurons from a DOWN-state to an UP-state) have been linked to multiple functions: facilitatingflexible processing of information (McCormick et al., 2004; McCormick and Yuste, 2006; Haider etal., 2006); involvement in rapid changes in functional connectivity during waking behavior (Neske,2016); and memory replay (Wilson et al., 1994; Sirota et al., 2003; Sirota and Buzsaki, 2005). While115Figure 51: Power spectrograms and stimulus annotations - Cat C5.116Figure 51: (continued from previous page) 4 tracks from one propofol anesthetized cat (C5)shows that the recordings come from mostly synchronized cortical state recording periods (bluetrace sections).UP-state transitions can be evoked by sensory or thalamic activation (Amzica and Steriade, 1998a;Steriade, 2001) they also occur spontaneously (Amzica and Steriade, 1995; Destexhe et al., 1999;Volgushev et al., 2006).UP-state circuits and genesis are still being debatedStudies over the past two decades have identified Layer 5 (L5) in cortex as a possible initiation sitefor UP-states with initial evidence from current-source-density (CSD) localization (Steriade andAmzica, 1996) and additional intracellular in vitro studies (Sanchez-Vives and McCormick, 2000)which suggested that neither thalamic input nor upper cortical layers were involved in genesis.Additional in vivo studies also suggest L5 cells are involved in UP-state genesis as they appear to:lead UP-state transitions (Chauvette et al., 2010), are intrinsically resonant to lower frequencies<15Hz present in the slow oscillation (Agmon and Connors, 1989; da Silva, 1991), contain a sub-type of pyramidal cells that bursts and have high spine density and wide arborization within L5.A recent optogenetic stimulation study (Beltramo et al., 2013) also found that stimulation of L5but not layer 2/3 (L2/3) neurons caused longer and more persistent transitions to UP-states. Twomechanisms by which L5 neurons could become activated and thus initiate UP-states could be (i)persistently active pacemaker L5 neurons which could initiate the cascade themselves or (ii) sponta-neous initiation due to stochastic integration of synaptic activity reaching critical thresholds. Whilesome studies suggest L5 cells ramp slowly (i.e. stochastically) while other cells depolarize quicklyduring UP-states (Chauvette et al., 2010) there is some evidence that L5 cells continuously fireeven during DOWN-states thus possibly re-initiating UP-states following a post refractory period(Sanchez-Vives and McCormick, 2000; Hasenstaub et al., 2007; Sakata and Harris, 2009; thoughsee a recent in vivo study found no spontaneous spiking in L5 during DOWN-states, Chauvette etal., 2010). However, intrinsic pacemaker resonance properties of L5 cells are also found in L2/3neurons (Le Bon-Jego and Yuste, 2007) with persistent sodium currents which can act like central-pattern-generators (CPGs) independently of input suggesting L2/3 could also play a central rolein UP-state transitions. Additional in vitro studies also identified core neurons in layer 4 (L4)that contribute to UP-states either spontaneously or via thalamus (MacLean et al., 2005; Rigasand Castro-Alamancos, 2007) leading some to suggest that there are as many as three sources forUP-state transitions: a cortical L4-L5/6 circuit and independent thalamocortical (TC) and retic-ular nucleus (NRT) populations (Crunelli and Hughes, 2010). Recent work has also found thatastrocytes can trigger slow-wave-oscillation states (Poskanzer and Yuste, 2016).117UP-state durationAlthough neurons spike the most during the early phase of an UP-state initiation, UP-states canlast from 100s of milliseconds to as much as a few seconds and it is uncertain whether cell-intrinsicor synaptic mechanisms are critical, though there is some evidence for the later: in vivo currentinjections (inhibitory or excitatory) do not affect UP-state duration or rhythmicity (Steriade et al.,1993a; Sanchez-Vives and McCormick, 2000); and membrane potential and inter-spike interval vari-ances during UP-states are high suggesting persistent synaptic input. Other mechanisms involvedlikely include: enhanced inhibitory neuron activity, synaptic depression of excitatory synapses, orhyperpolarizing conductances becoming active. One study (Steriade et al., 1993) showed that UP-states can be extended by stimulating brainstem cholinergic projectors (muscarinic signaling).Thalamus, brainstem and basal forebrain contribute to UP statesDespite slow oscillations surviving thalamic lesions (Steriade et al., 1993) and occurring in corticalslabs (Timofeev et al., 2000) and in vitro (Sanchez-Vives and McCormick, 2000) more recent studiessuggest thalamus has an active role in UP-states. In fact, a number of studies have shown thatTC relay cells, in particular, fire post-inhibitory rebound spike-bursts before UP-state transitionsin cortex suggesting thalamus initiates (Contreras and Steriade, 1995) and possibly synchronizesUP-states across cortex (Amzica and Steriade, 1995). Additionally, sensory stimulation is knownto evoke UP-states, even in anesthetized animals using drift gratings (Anderson et al., 2000; Jiaet al., 2010) or whisker stimulation (Petersen et al., 2003; Hasenstaub et al., 2007) supportingthe role of TC neurons. Further, UP-state frequency is decreased by severing TC axons in mousebarrel cortex slices (Rigas and Castro-Alamancos, 2007) as does pharmacological blockage of TCspiking in anesthetized and sleeping rats (David et al., 2013). And fast-oscillations occurring dur-ing UP-states (discussed below) are reduced along with multi-neuron synchronization by thalamicinactivation (Lemieux et al., 2014). Interestingly, sensory thalamic nuclei relay neurons are highlyinhibited after UP-state initiation but non-sensory nuclei are not (Sheroziya and Timofeev, 2014)likely because TRN neurons only connect to sensory thalamus and avoid non-sensory nuclei (Barthet al., 2002). This leaves open the possibility that non-sensory thalamus neurons (matrix neurons;Jones, 1998) may be involved in persistence and termination of UP states.Additionally, brainstem nuclei send chonlinergic, noradrenergic and serotonergic axons to cortexand thalamus, with some findings suggesting they constitute at least half of all synaptic inputs tosome thalamic nuclei (e.g. cat lateral-geniculate-nucleus, LGN; Erisir et al., 1997). Basal forebrain(nucleus bassalis, and accumbens and others) also sends cholinergic axons and it seems some basalforebrain neurons prefer either UP or DOWN-states (Detari et al., 1997; Manns et al., 2000; Mena-Segovia et al., 2008; Eschenko et al., 2012; Schweimer et al., 2011).The function of slow oscillations and UP-states are largely unknownWhile it is unclear how cortico-thalamic circuits give rise to slow oscillations and UP and DOWN-118state transitions, even less is known about the role of such oscillations in mammals, though a num-ber of suggestions are offered. Slow oscillations may facilitate synaptic plasticity and in particularmemory consolidation. Increased synaptic plasticity during slow oscillations has been previouslysuggested (Moruzzi, 1965; Steriade and Timofeev, 2003) and is supported by REM sleep and per-formance studies (Karni et al., 1994) and findings of replay of patterns of hippocampal place cell(Wilson et al., 1994) as well as newer studies (Waters and Helmchen, 2006; Tononi and Cirelli,2015). Such memory consolidation mechanisms may involve large dendritic calcium influx (Yusteand Tank, 1996) triggering synaptic plasticity mechanisms (e.g. CaMKII; Soderling and Derkach,2000). It has also been suggested that during UP-states hippocampus transfers information backto cortex: sharp wave ripples (SWR) appear in hippocampus right after UP-state initiation (Siapasand Wilson, 1998; Sirota et al., 2003) and hippocampal activity patterns are transferred to corticalnetworks during the UP-state while cortical plasticity is high (Sirota and Buzsaki, 2005). Addition-ally, cellular restoration (Vyazovskiy and Harris, 2013) may be occurring during DOWN states asit is known that death occurs in familial insomnia where SWS is impossible (Cortelli et al., 1999).Another interesting hypothesis is that during UP-states the membrane properties of neurons aremodified (McCormick et al., 2003) to possibly better detect sub-threshold inputs, or function asbetter gain control mechanisms (McCormick et al., 2004; McCormick and Yuste, 2006; Haider andMcCormick, 2009) but in a highly dependent fashion on the state of the neuron.UP-state transitions have LFP correlatesIn two publications entitled Cellular substrates and laminar profile of sleep K-complex and Elec-trophysiological correlates of sleep delta waves (Amzica and Steriade, 1997a, 1998a) it was shownthat K-complexes (now called UP-state transitions) recorded in cats during natural sleep were simi-lar to those recorded under ketamine-xylazine anesthesia. Importantly, it was shown that UP-statetransitions engage large parts of cortex and that single neuron depolarizations across cortex man-ifest as large amplitude negative depth LFP deflections (Figure 52). Additional studies over thepast several years have further shown that UP-state transitions do have single (Saleem et al., 2010)and multi-channel LFP correlates (Chauvette et al., 2010) and additional work in rat hippocampusslices has also identified that some LFP waveforms have stereotyped shapes that can be groupedtogether (Reichinnek et al., 2010). One particular study using LFP and intracellular recordings innaturally sleeping cats (Chauvette et al., 2010) has also shown that during UP-state transitions,large amplitude multi-laminar LFP deflections correlate with single neuron UP-state initiation(Figure 53A). Importantly, however, the extracellularly recorded LFP was shown to have a morestereotyped shape during UP-state transitions than individually patched single neurons which hadhigher somatic voltage variability relative to each other or to their own previous transitions - evenacross sequential UP-state transitions separated by 1-2secs (Figure 53B). Lastly, a recent study inanesthetized macaque hippocampus suggest that grouping of LFP by shape is possible (Ramirez-Villegas et al., 2015).119Figure 52: LFP negativities mark single neuron UP-state transitions. A. Local fieldpotential (top) and intracellular recording (bottom) of an area 5 neuron in a naturally sleepingcat reveals strong alignment between LFP negativities and single neuron membrane depolarization(dashed red lines) during UP-state transitions even in the absence of single neuron spiking (bluelines). B. An example of a single UP-state transition (same experiment as in A) measured usingdepth EEG (i.e. LFP), surface EEG, single neuron clamp and nearby extracellular electrode showsthe alignment of LFP to membrane voltage depolarization and nearby neurons spiking during thetransition. C. Same as (B) but the patched neuron does not spike while other nearby neurons spikesubstantially. (A-C adapted from Amzica and Steriade, 1998a with permission).Neuronal firing order during UP-state transitionsDespite the outstanding question of genesis, UP-states are becoming more commonly studied byresearchers seeking to understand neural coding in cortex (Luczak et al., 2015; Neske, 2016; Sanchez-Vives et al., 2017). A series of recent studies have focused on identifying UP-state transitions usingmulti-unit-activity (MUA) and evaluating single unit activity firing order during such transitions(Figure 54; Luczak et al., 2007, 2009, 2013, 2015; Bermudez-Contreras et al., 2013). In particular,the studies have shown that high-firing rate (>2Hz) neurons in rat somatosensory and auditorycortex are activated in a similar order (on average) during UP-state transition as they are dur-ing stimulus presentation (Figure 54A,B). It has been hypothesized that firing order of such highfiring-rate neurons encodes information and that further identifying order for all cells and in all120   UP-state          UP-stateTransition #1   Transition #2Figure 53: Multi-channel LFP is more stereotyped than single neuron potentials dur-ing UP-state transitions. A. UP-state transition (t=0sec) recorded using laminar, multi-channelLFP (left) and intracellular voltages from two neurons (right). B. Example of two sequential UP-state transitions recorded as LFP using multichannel extracellular electrodes (black trace) andintracellular voltages for two cells (red and blue traces) shows that LFP traces are much moresimilar over time than single neuron membrane voltages which have different dynamics even whenseparated by 1-2sec (A,B adapted from Chauvette et al., 2010 with permission).cortical areas may help reveal the coding strategy used by all cortex (Figure 54C,D; Luczak et al.,2015). Additionally, another recent study using cross-correlation histograms (CCH) has shown thatgroups of neurons (up to 12, average 7) recorded simultaneously in anesthetized cat visual cortexshow preferred firing sequences lasting up to ≈15ms in duration that change in part as a functionof stimulus properties (Havenith et al., 2011).While findings of firing order are evidence for spike timing neural coding theories, these firingorder studies have so far been limited to analysis of only high firing rate neurons in both UP-statetransitions and CCH-based studies. The CCH studies are limited by a number of factors includingthat they contain complex statistical tests (e.g. averaging after multi-resampling of sparse datasetsand tracking sub-millisecond firing sequence errors) and require heuristics for excluding sub-optimalstimuli. They also rely on higher-firing rate neurons for generating 1ms-bin CCHs with sufficientstructure (see analysis below and Figs 5.31, 5.32). And UP-state firing order studies have largelybeen limited to cortical areas where neurons fire spontaneously (and during stimulus presentation)at high rates (e.g. rat somatosensory and auditory cortex). Critically, using MUA to define UP-state initiation has limitations: it is potentially circular as single unit spiking is used to defineboth UP-state transitions and the single unit order that follows; peaks in MUA are dependenton heuristics such as the choice of histogram bin width and MUA threshold that indicates anUP-state transition. Furthermore, MUA-based definitions of UP-state transitions used so far canonly capture high firing rate neurons (>2Hz) for analysis (Luczak et al., 2007, 2009, 2013, 2015;Bermudez-Contreras et al., 2013). The latter limitation excludes most (>80%) of neurons recorded121Figure 54: Latency order results from other studies. A. UP-state transition triggeredsingle neuron PETHs in a urethane anesthetized rat somatosensory cortex recording with whitedots corresponding to approximate centre-of-mass (COM) metric used in the studies. B. ExampleCOM (red dots) and firing distributions (grey traces) for 90 rat auditory cortex neurons recordedduring spontaneous and stimulus evoked periods show relatively similar order across all conditions(note: neuronal order is based on COM order during spontaneous recording periods, i.e. UP-statetransition-based order). C. Hypothesized neuron firing order packets occuring during spontaneoussynchronized state activity are similar to packets during tone presentation. D. Similar hypothesisas (C) but during desynchronized (e.g. awake and attending) cortical states. (A - Adapted fromLuczak et al., 2007 with permission; B - Adapted from Luczak et al., 2009 with permission; C,D -Adapted from Luczak et al., 2015 with permission).122in sensory cortex where firing rates for stimulus and spontaneous activity are generally <2Hz (e.g.for mouse sensory cortex recordings analyzed in this chapter the median firing rate was 0.26Hz with89% of neurons spiking at rates <2Hz; in cat V1 recordings presented here the median firing ratewas 0.31Hz with 86% of neurons spiking at rates <2Hz; see also Methods). Thus, a single-neuronindependent and largely heuristic-free definition of UP-state transition that could enable the studyof all neurons’ relative firing order in all cortical areas is not yet available.This chapter presents an improved method for detecting UP-state transitions using LFP wave-forms (rather that MUA) and applies this method to visual cortex to yield novel results. Themethodology developed relies on LFP events, i.e. large amplitude LFP events 50-200ms in du-ration to provide more temporally precise UP-state transition times for computing and trackingspiking properties (e.g. latencies) across all cells, including low firing-rate cells. The focus here islargely on extracellular recordings in anesthetized cat and mouse cortex because under such con-ditions slow-oscillations dominate and there are many UP-state transitions which can be identifiedand analyzed.ResultsClustering multi-laminar LFP events reveals distinct UP-state transition classesThe approach for identifying UP-state transitions from LFP is similar to single-unit spike sortingapproaches: event detection is carried out on the raw LFP voltage record and clustering methods arecarried out on aligned multi-channel LFP events (Figure 5.8A-C; see also Methods for a completedescription). Detected multi-channel LFP events in the recordings considered here had similarextracellular waveforms across most or all channels; each event contained 1 to 2 troughs and 1peak; the events lasted between ≈50ms to 200ms; and PTP amplitudes of the events were between≈250µV to 500µV (Figure 55D(i), E(ii)). After aligning LFP events using a centre-of-gravityiterative RMS procedure applied on the waveform on the maximum PTP channel (see Methods),features were selected and visualized using PCA (Figure 55C(iii)). The observed distributionswere clustered (Swindale and Spacek, 2014) into distinct LECs, individual LFP waveforms in eachLEC were re-aligned using RMS and timestamps for each waveform was exported as the temporallocation of the centre-of-gravity of each waveform (Figure 5.8C(iv); see also Methods).LEC extracellular templates were obtained by averaging aligned waveforms in each LEC (Fig-ure 55D(i), E(ii)). The templates were distinct from each other and the PC values did not driftsubstantially over time (Figure 55(ii), E(iii)) indicating that LECs are largely stable (Figure 558D:cat visual cortex - Recording ID: C5.3; Figure 1E: mouse auditory cortex - MV1).Analysis across 15 cat visual cortex tracks yielded a total of 34 LECs (≈2.3 LECs per track;see Table 5.1) and across 20 mouse sensory cortex tracks yielded 33 LECs (≈3.1 LECs per visualcortex track; ≈1.6 LECs per barrel cortex track; and 1 LEC per auditory cortex track; see Table123Figure 55: Clustering LFP events during synchronized cortical states. A. Top: powerspectrogram for a cat visual cortex recording (deepest channel); Middle: synchrony index (red curve;see Methods) falls largely above threshold (dashed line) indicating synchronized states; Bottom:single unit rasters ordered by neuron depth with extracellular polytrode (right). B. Same as (A)but from a mouse visual cortex recording. C. LFP event clustering methodology: (i) 3-secondsynchronized state recording from a cat visual cortex recording (bottom traces) overlayed withMUA histogram shows spiking activity coincides with large amplitude negative LFP deflections aspreviously reported (Amzica and Steriade, 1997b, 1998b; Chauvette et al., 2010); (ii) LFP recordis high-pass filtered (4Hz cutoff) and events over threshold (4 x STD, see Methods) are identified(bold traces); (iii) LFP events are aligned and clustered into LECs (see Methods); (iv) two clustersidentified using PCA. D. Templates of LECs in 1C(iv); (ii) Plots of principal components for clustersin 1C(iv) vs. time showing almost no changes over time. E. Same as F, but for a mouse visualcortex recording.124Table 2: Cat recordings and clustering detailsRecording ID Anesthetic Track ID Sync Rec Time1 / Total Rec Time No. of LECsC1.1 Iso/N20 1 0.6 / 13.4hrs2 2C1.2 Iso/N20 2 1.4 / 9.4hrs2 2C2.1 Iso/N20 1 1.7 / 11.1hrs2 3C2.2 Iso/N20 2 4.7 / 12.0hrs2 1C3.1 Iso/N20 1 1.4 / 8.4hrs 1C3.2 Iso/N20 2 4.2 / 8.3hrs 3C3.3 Iso/N20 3 3.4 / 7.4hrs 2C3.4 Iso/N20 4 3.0 / 5.3hrs 1C4.1 Propofol/Fentanyl 1 3.3 / 11.9hrs 1C4.2 Propofol/Fentanyl 2 0.5 / 11.5hrs 23C4.3 Propofol/Fentanyl 3 0.9 / 4.6hrs 2C5.1 Propofol/Fentanyl 1 2.5 / 2.9hrs 3C5.2 Propofol/Fentanyl 2 4.9 / 10.6hrs 3C5.3 Propofol/Fentanyl 3 3.1 / 8.5hrs 4C5.4 Propofol/Fentanyl 4 4.6 / 6.3hrs2 4Totals 15 341. Synchronized recording time was defined as periods with synchrony index >0.5. 2. Synchronyindex threshold of 0.3 used. 3. LECs had peak-to-peak amplitude and CSD maxima belowthreshold and were excluded from CSD grouping and further analysis.5.2).LECs have discrete laminar CSDs patterns common across tracks and animalsCSD profiles for each LEC template were investigated to determine whether there were uniquesinks and sources attributed to each template (Figure 56). CSD analysis (see Methods) revealedthat while during desynchronized cortical states events with high amplitude current sinks andsources are frequent and present at multiple depths across cortex, during synchronized states cur-rent sink-source pairs were more infrequent (≈1Hz - consistent with UP-state transition frequenciesof 0.2-0.9Hz) and fell into distinct and spatially localized patterns (Figure 56A). Computing CSDprofiles from LECs revealed a more diverse laminar profile in comparison to extracellular templateswhich typically have qualitatively similar shapes (and amplitudes) across most or all channels (Fig-ure 56B: left: LEC extracellular template; right: corresponding CSD distribution). This indicatesthat during synchronized cortical states, UP-state transitions engage distinct cortical circuits withspatially localized sinks (e.g. neuronal action potentials) and sources (e.g. repolarization or distalinputs) in partial support of the multiple oscillator hypothesis for UP-state genesis (Crunelli andHughes, 2010).Computing CSDs for all cat visual cortex and mouse sensory cortex recordings LECs (Figure 57-Figure 60) revealed that similar CSD patterns could be identified across tracks from the same125Table 3: Mouse recordings and clustering detailsRecording ID Anesthetic Track ID Area Sync / Total Rec Time No. of LECsMV1 Isoflurane 1 Visual 2.9 / 3.0hrs 3MV2 Isoflurane 1 Visual 3.3 / 3.3hrs 4MV3 Isoflurane 1 Visual 2.6 / 3.4hrs 4MV4 Isoflurane 1 Visual 2.3 / 2.4hrs 3MV5 Urethane 1 Visual 5.3 / 5.3hrs 3MV61,4 Isoflurane 1 Visual 2.0 / 2.8hrs 2MV72,4 Isoflurane 1 Visual 2.6 / 3.1hrs 3MB1 Isoflurane 1 Barrel 2.8 / 2.9hrs 1MB2.1 Isoflurane 1 Barrel 2.7 / 2.7hrs 2MB2.2 Isoflurane 2 Barrel 2.3 / 2.4hrs 2MB3.1 Isoflurane 1 Barrel 2.8 / 3.1hrs3 1MB3.2 Isoflurane 2 Barrel 1.8 / 2.0hrs3 1MB4 Urethane 1 Barrel 7.4 / 7.8hrs 1MB5 Urethane 1 Barrel 1.5 / 2.7hrs 3MA1 Isoflurane 1 Auditory 1.9 / 2.0hrs 1MA2 Urethane 1 Auditory 2.0 / 3.0hrs3 1MA3 Urethane 1 Auditory 1.5 / 3.9hrs3 1MA42,4 Isoflurane 1 Auditory 0.9 / 1.2hrs 1Totals 20 331. Tetrode recordings did not yield laminar (i.e. relative depth) information and were not used forCSD analysis. 2. VSD recordings. 3. Synchronized state defined as synchrony index >0.3 (>0.5used for all other recordings). 4. Wildtype mice (all other mice GCaMP6s).animal (Figure 56C, D) or even across animals (Figure 56E). Grouping all data from cat visualcortex recordings by qualitatively similar CSD pattern showed that few CSD patterns were uniqueto one track only: total numbers of LECs grouped into sets of 2 or more: cat V1: 30/34; mousevisual cortex: 14/17; mouse barrel cortex: 8/11; mouse auditory cortex: 3/3; see Tables 5.3, 5.4 andFigure 57-Figure 60). This suggests that individual LECs do not reflect idiosyncratic circuits (i.e.animal specific) but engage common cortical circuits present in all cats or possibly all mammalsthat have cortical UP/DOWN states (note: CSD patterns were not compared across species, i.e.between cat and mouse visual cortex recordings).Lastly, in addition to the PCA space stability (see Figure 55), the temporal stability of all LFPwaveforms in each LEC group was evaluated by computing the standard deviation of the full-width-half-max (FWHM) of each LFP event on the maximum amplitude channel (Figure 56F-I). 4 LECexamples provided from a single cat V1 recording (Figure 56F) have FWHM±STD distributionmeans of 28±7ms, 22±4ms, 24±5ms and 18±5ms, respectively (Figure 56H). Further grouping ofall LECs across all cat tracks using CSD patterns (colours in Figure 56I match coloured groupingdots in Figure 57) revealed that for the first 6 LEC groups the FWHM STD distributions within126Figure 56: CSD-based LEC grouping and temporal precision. A. CSD profiles computedduring spontaneous cortical activity in cat visual cortex during a desynchronized (left) and syn-chronized (right) recording period in the same track. (Note: CSD colour scales are normalized tobe centred on current=0 µA/mm3 with -1 representing current sinks: positive current flowing awayfrom extracellular space, and +1 representing current sources: positive current flowing into theextracellular space). B. Example of an LEC template and its CSD correlate. C. Examples of CSDsobtained from LECs from different tracks in a propofol anesthetized cat. D. Examples of CSDsobtained from LECs from different tracks in an isoflurane anesthetized cat. E. Examples of CSDsobtained from LECs from tracks in two different isoflurane anesthetized cats. F. LFP templates ofLECs from Figure 55D. G. Method for computing the stability of extracellular LFP events in eachLEC using the FWHM of the first trough for each LEC event. H. Trough FWHM ±STD of all LFPevents in each LEC in (F) shows that most troughs FWHM are stable (i.e. the STD is ≈5ms; seemain text also). I. STDs (hollow circles) and average of STDs (filled circles) of all LEC templatesbelonging to the first 6 groups of LECs (see Figure 57 for colour matching; see also Methods andmain text).127Table 4: Cat visual cortex - LEC groupsRecording ID Anesthetic Groups#1 #2 #3 #4 #5 #6 #7 #8 #9 OtherC1.1 Iso/N20 X XC1.2 Iso/N20 X X XC2.1 Iso/N20 X X XC2.2 Iso/N20 XC3.1 Iso/N20 XC3.2 Iso/N20 X X XC3.3 Iso/N20 X XC3.4 Iso/N20 XC4.1 Propofol/Fentanyl X XC4.3 Propofol/Fentanyl X XC5.1 Propofol/Fentanyl X X XC5.2 Propofol/Fentanyl X X XC5.3 Propofol/Fentanyl X X X1 XC5.4 Propofol/Fentanyl X X X,X1Totals 5 5 4 5 1 4 2 2 2 31. These 2 LECs are similar to Class 6 but may also form a separate class.each group (hollow circles) and for their average (filled circles) are <10ms (Figure 56I: cat visualcortex, 23/25 LECs STD <10ms; Figure 58: mouse visual cortex 13/14 <10ms).Most neurons lock to LEC-defined UP-statesWhile a number of previous studies have shown high amplitude LFP deflections to be the globalcorrelates of UP-state transitions (Amzica and Steriade, 1998a; Steriade and Amzica, 1998; see Fig-ure 52,Figure 53) the methodology described here can identify UP-state transitions with significantprecision: ≈5-10ms for most recordings in cat and mouse visual cortex (Figure 56I,Figure 58) and≈5-15ms in mouse barrel and auditory cortex (Figure 59,Figure 60). The approach thus enables theinvestigation of single neuron firing properties (e.g. latency and order) during UP-state transitionsin visual cortex of cat and sensory cortex of mouse.The equivalence of LECs and UP-state transitions is further shown using single unit activity intwo example recordings from cat visual cortex and mouse auditory cortex (Figure 61). Peri-LEC-Event-Time-Histograms (PETH) for single neurons identified via single unit spike sorting for thetwo highest PTP-amplitude LECs (Figure 61A-C,E-F; see also Methods). PETHs were computedby convolving each spike for every neuron with a 5ms (cat visual cortex) or 10ms (mouse auditorycortex) STD gaussian. (Note: method adopted from Luczak et al., 2007, 2009, 2013 where 20mswide gaussians were used, but changed to reflect LFP event FWHM results presented above; seealso Methods). An example PETH from a single neuron in cat visual cortex computed for the two128Figure 57: LEC groupings - cat visual cortex. Cat visual cortex grouping of all clusteredLECs (34 LECs across 15 tracks in 5 cats) based on CSD profile similarity reveals that most LECscan be grouped together with other LECs from either a different track or a different cat. Thecoloured dots represent LEC groups analyzed in Figure 56I.129Table 5: Mouse Sensory Cortex - LEC GroupsRecording ID Anesthetic Area Groups1#1 #2 #3 #4 OtherMV1 Isoflurane Visual X X XMV2 Isoflurane Visual X X X XMV3 Isoflurane Visual X X X XMV4 Isoflurane Visual X X XMV5 Urethane Visual X(3)Totals 4 4 4 2 3MB1 Isoflurane Barrel XMB2.1 Isoflurane Barrel X XMB2.2 Isoflurane Barrel X XMB3.1 Isoflurane Barrel XMB3.2 Isoflurane Barrel XMB4 Urethane Barrel XMB5 Urethane Barrel X(3)Totals 3 2 3 0 3MA1 Isoflurane Auditory XMA2 Urethane Auditory XMA3 Urethane Auditory XTotals 3 0 0 0 01. Grouping labels are different for different cortical areas and across cat and mouse recordings(e.g. visual cortex group #1 6= barrel cortex group #1).Figure 58: LEC groupings - mouse visual cortex. Mouse visual cortex grouping of all LECs(17) in 5 tracks from 5 mice. Coloured dots represent LEC groups analyzed in the panel on theright for their trough FWHM STD values.130Figure 59: LEC groupings - mouse barrel cortex. Mouse barrel cortex grouping of all LECs(11) in 4 tracks from 4 mice. Coloured dots represent LEC groups analyzed in the panel on theright for their trough FWHM STD values.Figure 60: LEC groupings - mouse auditory cortex. Mouse auditory cortex grouping of allLECs (3) in 3 tracks from 3 mice. Coloured dot represents the LEC group analyzed in the panelon the right for trough FWHM STD values.1310      1Figure 61: Single neuron UP-state latencies in cat and mouse cortex. A. Examplesof LEC-defined UP-state transition triggered single neuron histograms (PETH) from a cat visualcortex recording for the two largest amplitude LECs reveal narrow spiking distributions near LECtime (t=0ms). B. Histograms for all neurons aligned by peak latency (left) or cortical depth ofneuron (right) show the vast majority of neurons responding strongly to LECs within a narrow(i.e. ≈25ms) window and have FWHM PETH distributions of ≈5-10ms. C. Same recording as B,but for a different LEC. D. Percentage of single neuron spikes locking to LECs (i.e. ±50ms fromUP-state transition) for all neurons aligned by depth (left) or neuron firing rate (right) - showsa depth correlation between neuron firing percentages and LEC type (but not firing rate). E-H.Same as A-D, but for a mouse auditory cortex recording.highest amplitude LECs (Figure 56A) reveals a very narrow spiking distribution near the definedcentre of the LEC events (i.e. t=0ms). In contrast, PETH results from a single neuron in mouseauditory cortex activity during events from the highest amplitude LECs (Figure 5.14E) reveal asharp onset but a slower decay time of 10ms to 100ms (or longer) consistent with UP-state PETHsreported previously (see Figure 54A,B and Luczak et al., 2007, 2009, 2013, 2015).Plotting PETHs for all neurons in order of PETH peak (Figure 61B-left) or cortical depth(Figure 61B-right), reveals a strikingly narrow distribution of latencies with virtually all neuronsrecorded firing with an ≈25ms window of each other and individual distributions of ≈5ms FWHM.Neurons respond to LEC #2 across a broader period of time (≈100ms), but also individually132Figure 62: UP-state latencies: 1-Sec interval. A.1-sec interval LEC triggered PETH from amouse visual cortex recording. B-D. Same as (A) but for different recordings and areas as indicatedin legends above histograms. Note: all PETHs are ordered by latency and were computed for a-500msec to +500msec window.exhibit narrow latency distributions similar to those for LEC #1 (Figure 61C). The different PETHdistributions between the two LECs further supports the observation that UP-state transitionsfall into different types. Similarly plotting mouse auditory cortex neuron PETH distributionsalso reveals that most neurons peak in spiking near the UP-state transition (i.e. t=0ms) butcontinue to spike subsequently (i.e. have longer tailed distributions; Figure 61E-G) consistent withprevious findings in rat auditory and barrel cortex (Luczak et al., 2007, 2009, 2013; Bermudez-Contreras et al., 2013). Lastly, the proportion of neuron spikes that occur within a ±50ms (catvisual cortex) or ±100ms (mouse barrel cortex) window of LEC events was investigated. In thecat visual cortex recording example provided, neurons fired from ≈10-90% of their spikes duringLEC-defined UP-state transitions. There was also evidence that LECs prefered specific classes ofUP-state transitions: superficial cells fired more during LEC #1 (Figure 61D - red coloured LEC)while deeper cells preferred LEC #2 (Figure 61D - green LEC). Cells from the mouse auditorycortex example fired ≈10-70% of their spikes during LEC defined UP-states, well above chance(dashed lines in each plot indicates chance; see Methods).Considering other cat visual and mouse sensory cortex recordings, it was found that LECscould be used to reveal strong peaks in PETH distributions for >90% of all neurons recorded (seeFigure 62 for additional examples using wider windows). PETH distributions for visual cortex cells(in both cat and mouse) had narrower (i.e. ≈5-10ms wide) PETH FWHM during LEC defined UP-state transitions, while mouse auditory and barrel cortex neurons also had strong peaks but longertailed distributions as previously reported (Luczak et al., 2007, 2009, 2013; Bermudez-Contreras etal., 2013). These findings provide further evidence that LECs are temporally precise LFP correlatesof DOWN-to-UP state transitions and are tightly linked to spiking of all neurons irrespective offiring rates or cortical area.The finding that visual cortex neurons have narrow distributions with full-width-half-maximum133(FWHM) activations on the order of ≈5-10ms and peaks near UP-state transitions (i.e. t=0ms),has not been previously made. It suggests that similar to rat somatosensory and auditory cortex(Luczak et al., 2007, 2009, 2013), visual cortex neurons fire substantially during UP-state transi-tions. Importantly, these findings allow for the investigation of relative firing order between visualcortex neurons.UP-state transitions have mesoscale correlates in mouse dorsal cortexUP-state transitions identified as LECs have thus far been shown to have: distinct PCA-spaceclusters (i.e. LECs); specific CSD patterns as well as correlating with strong single neuron spikingpeaks. Additional experiments were implemented to investigate meso-scale correlates of LEC de-fined UP-state transitions. In particular, widefield calcium activity in GCaMP6 mice (Ai93, Ai94;see Methods) was recorded simultaneously with extracellular potentials with the goal of linkingUP-state transitions with dorsal cortex neural activity. Accordingly, LEC-event triggered widefieldGCaMP6 activity maps were computed (similar to spike-triggered-averaging of imaging activity;Arieli et al., 1995; Tsodyks et al., 1999; Xiao et al., 2017) to identify spatio-temporal patterns (e.g.motifs) occuring in mouse dorsal cortex during UP-state transitions. Single-hemisphere (auditory-cortex) and bi-hemispheric (visual and barrel cortex) calcium activity (30Hz) recordings were madein 23 GCaMP6s transgenic mice while simultaneously acquiring extracellular potentials from vi-sual, barrel and auditory cortex (see also Methods). Findings revealed that LEC-triggered calciummotifs (i.e. averages of calcium signals centred on the events from each LEC) correlated with globalcortical activity patterns with most patterns showing depression-to-activation dynamics centred onthe specific area where the extracellular recording was made (not shown).However, as discussed above, the GCaMP6 transgenic mice used in these recordings employed aCre-based breeding strategy and a recent study (Steinmetz et al., 2017) specifically identified Cre-dependent GCaMP6s mice as potentially having aberrant electrical activity that is visible both inthe extracellular potential recordings and widefield calcium activity. Accordingly, LEC-triggereddorsal cortex calcium activity maps will not be discussed further to avoid drawing conclusions thatmay arise from potentially pathological cortical conditions.However, 2 additional imaging experiments were available where voltage-sensitive-dyes (VSDs)were used in wildtype mice along with extracellular recordings in visual cortex (1 mouse; Fig-ure 63A) and auditory cortex (1 mouse; Figure 63B; see also Methods). In contrast with GCaMP6imaging, VSD has the advantage of reporting both subthreshold and suprathreshold neural activ-ity and at a higher temporal resolution (150Hz). VSD motifs triggered off LECs (i.e. UP-statetransitions) in mouse visual cortex peaked in amplitude nearly simultaneously with the LEC events(typically ±30ms centred on UP-state transition times (Figure 5.16); Figure 63A(ii)) with mo-tifs revealing a gradual cortical activation preceding t=0ms (i.e. ≈-100ms to +100ms) suggesting agradual build up of membrane depolarizations prior to UP-state transition. This gradual activationis expected from previous work using single neuron patch clamping techniques (Volgushev et al.,134Figure 63: Mesoscale VSD correlates of UP-state transitions.135Figure 63: (continued from previous page) A(i). Three LEC-triggered meso-scale VSD motifs(-0.2sec to +0.2sec) from a mouse visual cortex recording (note: arrow indicates location of elec-trode insertion). A(ii). Temporal dynamics for motifs in A(i) from four cortical ROIs (L RS: leftretrosplenial, L V: left visual, L BC: left barrel, and L M: left motor). B. Same as A but for anauditory cortex recording. (Note: dF/F0 values during spontaneous VSD activity usually peak at0.5% and substantially lower when averaging over multiple events; see also Methods).2006; Chauvette et al., 2010). Additionally, VSD activity was more similar across all cortical ROIsidentified (see traces in Figure 63A,B(ii)) consistent with intracellular findings of simultaneous,global depolarization of neurons (Amzica and Steriade, 1995; Destexhe et al., 1999).Average UP-state transition latencies change over timeThis chapter’s main goal is to investigate whether visual cortex neurons have prefered latenciesduring UP-state transitions (i.e. PETH peaks) that can be systematically tracked (i.e. using amethod with a defined temporal precision) and if so, to examine how order during spontaneousactivity might be related to order during stimulus evoked periods. The aim was thus to extenda number of recent findings of preserved order of high-firing rate (i.e. >2Hz) neurons in rat so-matosensory and auditory cortex (Luczak et al., 2007, 2009, 2013; Bermudez-Contreras et al., 2013)to all other sensory cortex neurons including visual cortex which is an area of active research wheresuch findings have not been previously made. However, work in this chapter also relates to theo-ries of cell assemblies and synfire chains (Abeles, 1982b; Abeles and Gerstein, 1988; Abeles, 1991;Abeles et al., 1993; Abeles, 1982a; Ikegaya et al., 2004; Izhikevich, 2005; Schrader et al., 2008; Grunand Rotter, 2010; Gerstein et al., 2012; Torre et al., 2016; Quaglio et al., 2017) that hypothesizethat networks (or chains) of functionally connected neurons exist in cortex which can be studiedby identifying groups of neurons that repeatedly spike with similar inter-spike-intervals. Despiteongoing research, findings of synfire chains at the level of single neuron spiking have been limitedwith some even arguing that the few positive findings may be statistically insignificant (i.e. occur atchance: Baker and Lemon, 2003; or require multi-tier statistical support: Gerstein et al., 2012). Avery recent study (Russo and Durstewitz, 2017) developed a ”temporal-scale independent” methodof searching for synfire chains but also did not find a universal ”cortical coding scheme” (i.e. singleneuron-based synfire chains) in rat frontal cortex or monkey visual cortex recordings.One possibility for why strong evidence for synfire chains has not been identified is that temporalorder may be involved in sensory processing but is not stable for long enough to be detectedover longer recordings (i.e. order changes over time). Accordingly, before proceeding to furtherdetermine how latencies during UP-state transitions might be preserved between spontaneous andstimulus recordings UP-state transition latencies were checked for temporal stability. Using differentapproaches and recording examples it is shown below that even during spontaneous recordingperiods most neurons’ preferred latencies (i.e. peaks in PETHs) as well as spiking distributions136Figure 64: Average UP-state latencies change over time. A. LEC latencies for all neuronsin a mouse visual cortex recording for the first 60 minutes (left) versus the entire 300 minute (right)recording. B. LEC latencies in a cat visual cortex recording for 3 sequential recording epochs (0-30mins, 30-60mins, 60-90mins) reveal that peak-latency-times observed in the first epoch are notpreserved over time (see also main text). C. Example of single neuron peak-latency-times computedusing a 30 minute sliding window (1-minute increments) reveals the dependency of peak latencyon time window (real data: magma colour trace; shuffled data: grey trace; see Methods). Dashedgreen lines at the bottom indicate cell raster and black lines indicate LEC events, i.e. UP-statetransitions. Inset: distributions of average latencies at each minute computed using only odd LECevents vs even LEC events reveals low error in latencies over time (i.e. small difference to the x=ydashed line). D. Top: P-value distributions for pairwise KS tests of distributions in all possible30 minute windows reveals that distributions are similar between neighbouring periods (i.e. dataalong diagonal is largely similar) but can vary substantially for periods more than ≈30-minutesapart (i.e. values off diagonal; see also main text). Bottom: Same as (Top) but for shuffled data(grey traces in C(i)) reveals no statistically significant differences are present when shuffling LECtimes.137(see below) are only stable for periods of <20-30 minutes and systematically (i.e. non-randomly)drift over periods of >30-90 minutes or longer.A recent study (Havenith et al., 2011) using a cross-correlation histograms (CCH) methodpreviously introduced (Schneider and Nikolic, 2006) has shown that sequences of up to 12 neuronsfrom cat visual cortex show firing order in sequences of up to ≈15ms in duration that are referencedto internal beta/gamma oscillations (see Figure 5.31 and discussion later in this chapter). The studyalso found that firing sequences change as a function of stimulus properties (though 7 of 11 recordedneurons in an example provided maintained the same relative order; see Havenith et al., 2011) andthat reliability increased during beta/gamma oscillations. Importantly, the study also found thateven for the same stimulus ”small changes in sequence can be seen” for recordings made 3-12hours apart. While these findings of small changes were largely dismissed by the study, they areconsistent with the hypothesis that order may change over time and further findings discussed inthe following sections.Below, several approaches are employed to track UP-state transition latencies including: usingsliding windows to compute latencies at every minute and track their changes (Figure 64, Figure5.18, Figure 67); identifying examples of stable and drifting triplet histograms peaks - commonlyused in synfire chain studies (see below; Figure 66, Figure 67); comparing the stability and similarityof latency times of non-overlapping recordings (e.g. sequential recording periods) for individual neu-rons (Figure 68-Figure 70); and evaluating the stability of multi-neuron latency-state-spaces acrossspontaneous and stimulus evoked recording periods (Figure 68-Figure 70). A simple qualitativetest for latency stationarity is to compute latencies across different recording periods (Figure 64).As an example, a 300-minute mouse visual cortex recording is used and all neuron PETHs arecomputed in the first 60 minutes and recomputed for the entire 300 minute recording (Figure 64A).The distributions become broader over the 300 minute period suggesting that they are sensitive tothe duration of the recording. Such latency broadening is not expected if either (i) the latenciesof each neuron are stationary over time; or (ii) if spiking during UP-state transitions is randomlydistributed within some broader window (e.g. 100ms-200ms). A second example is provided froma cat visual cortex recording (Figure 64B). In the cat example, LEC latencies for 64 neurons werecomputed during 3 non-overlapping, sequential 30-minute recording epochs and displayed in orderof latencies present during the first 30-minute period. The loss of order is visible and computing(Pearson) correlation between the values representing the peak latencies of all 64 neurons (i.e.pairwise comparisons of 2 64-dimensional vectors) resulted in values of 0.4 (first vs second period)and 0.3 (first vs third period). This provides an initial hint that rather than being stationary orrandomly changing, spiking distributions for some neurons gradually but systematically changeover time (see Figure 70 for systematic correlation value analysis; see also Discussion).An additional method where latencies were computed using a sliding window approach wasimplemented (Figure 64C). The method involved computing the latency of each neuron within a30-minute window of a recording and sliding across the entire recording in 1-minute increments138(see Methods). This method subsumes approaches which generally compute latencies for either theentire recording or for the first vs the second half of the recording (Luczak et al., 2007, 2009, 2013;Bermudez-Contreras et al., 2013). In previous studies latencies (or more precisely COM values -see Figure 54) were only computed once for an entire recording period or experiment. Additionally,previous studies did not have the temporal precision provided by LEC events thus prohibitingfurther dividing of the data into shorter periods or computing latencies for all neurons, in particularlower firing rate neurons (personal communication with A. Luczak). Using the sliding windowapproach for the mouse visual cortex neuron example (Figure 64C) revealed a gradual change inlatency over a period of 300-minutes with most of the change occuring in approximately 100-minuteperiod (50 to 150-minute section). Two controls were implemented. First, the neuron LEC timeswere shuffled (Figure 64C, grey traces). The resulting UP-state transition spiking distributionsare the same for each UP-state transition, but when averaged over windows of 30minutes (or anyperiod of time) they will lose any inherent differences that were present in the original data thatmay gradually change over time (i.e. Figure 64C - grey trace values were relatively similar acrossthe entire recording). A second control was implemented where the peak latency time in eachwindow was computed using odd vs even spikes (i.e. selected from all spike times in that 30-minutewindow). The differences are plotted as vertical error bars at each point in time (see Figure 64C,magma trace vertical error bars). The odd vs. even latency error is also displayed in the inset:each dot represents the peak latency in each 30-minute window computed using odd vs. only evenspikes and the distance from the X=Y line constitutes the error in milliseconds (note: this is thesame controls implemented previously, e.g. Luczak et al., 2007, 2009). The result is that the vastmajority of latency differences are <5ms.Last, non-stationarity was further evaluated by comparing spiking distributions between allpair-wise 30-minute epochs using a 2-sample Kolmogorov-Smirnov (KS) test (Figure 64D). Forthe neuron considered, all UP-state transition spike latencies (i.e. times) during the 0-30minuterecording period were compared with spike times from the 30-60minute recording period. Thus, ifthe first neuron fired 100 spikes during UP-state transitions in the first 30 minutes of the recording,the relative times (i.e. relative to UP-state transition time) were compared to the transition timesin the second 30 minute period of the recording. This KS test is different from comparing peaklatencies - as was done above - as entire spiking distributions are compared and not just peak latencytimes. The results reveal that spiking distributions also change over time: spiking distributionsbetween neighbouring epochs (e.g. <30-minutes) are more statistically similar (Figure 5.17D:darker colours along diagonal) while distributions between recording epochs farther away are morelikely to be statistically different (brightest colours indicate pval >10-6). This provides a largelyindependent confirmation that spiking distributions during UP-state transitions change on the orderof 30-minutes or longer.Further computing sliding-window peak-latency-times for all neurons in a mouse visual cortexrecording revealed that most neurons have stable latencies on the order of several minutes, but139Figure 65: UP-state latency drift - visual cortex neurons. Mouse visual cortex recordingexample of all neuron PLs over a period of ≈180-minutes. The colours represent depth of recordedneuron (see right of inset for approximate location on probe) and the inset scatter plot representsodd vs. even spike PLs for all neurons across all time windows (Note: several neurons for whichthe maximum odd vs. even PL error was >10ms were excluded as they had very low firing ratesor periods during which their latency could not be adequately tracked).that the latencies drift over periods of 30 to 60-minutes (Figure 65; note: this recording was froma GCaMP6s mouse visual cortex recordings and further analysis was not pursued, however, similarlatency trends in other mouse and cat cortical recordings were observed). This computation revealsthat neurons maintain relatively similar orders - even as there is an overall correlation in drift acrossthe recording - with some neurons changing their latencies substantially more than others.The findings that for most neurons UP-state transition latencies are stable on the scale of≈30-60 minutes but for some neurons can change over longer periods are generally consistent withprevious findings for high firing rate rat somatosensory and auditory cortex neurons which largelyconsidered recording periods of ≈1-2 hrs but lacked the temporal precision to compute drift (Luczaket al., 2007, 2009, 2013; Bermudez-Contreras et al., 2013). Interestingly, the results presented here -that neurons can change their relative latency times - are consistent with previous work identifyingneuron ensembles (i.e. simultaneous firing neurons) where cell membership within each ensemblecould change over periods of many minutes (Ikegaya et al., 2004). Thus, a possible explanationfor some neurons’ dropping out from participation in ensemble activity may be that they simply140Figure 66: Triplet histograms - examples. Two examples of triplet-histograms from a (wild-type) mouse visual cortex (top) and cat visual cortex (bottom) recordings reveal that some neurontriplets have preferential spiking modes during synchronized state (i.e. aneshetized) cortical record-ings. (Scale normalized to peak of histogram; see Methods).drifted too far away to be visible in simultaneous two-photon activity. Overall, the findings in thissection point to a possibly ever-evolving underlying temporal neural code that can be tracked andinvestigated using LECs (see also Discussion). The results may also provide an explanation for whypositive results from synfire chain studies have been relatively rare.Triplet histograms during synchronized state recordingsThe findings above suggest that the average single neuron latency during UP-state transitions isstable over shorter periods (e.g. 30-60 minutes) but changes over longer periods. It would thus beinteresting to determine whether this change of stability could be observed across the spiking orderbetween pairs of neurons. One other method for tracking relative spike timing is to compute triplet-histograms distributions (Abeles, 1982a; Luczak et al., 2007; Markram et al., 2015) and search forthe peak in histogram activity, i.e. the spiking mode of neuron triplet. The method involvesselecting three neurons and computing 2-dimensional histograms where the x-axis represents thecross-correlogram (or cumulative PETH) between the first and second neurons, and the y-axisrepresents the cross-correlogram between the second and third neurons (see Methods). Peaks inthe 2D distributions indicate a preserved temporal spiking relationship between all the neurons.Triplet histograms were computed for randomly selected 1-hour synchronized state recordingsfrom cat and mouse visual cortex (Figure 66). Three higher-firing rate (>1Hz) were selected and141Figure 67: Examples of triplet-histogram drift. Triplet-histogram examples from GCaMP6mouse recordings in visual (top) and barrel (bottom) cortex reveals that spiking modes (i.e. his-togram peaks) can change over time. Note that, as expected, visual cortex histograms are narrowerthan barrel cortex. (Scale normalized to peak of histogram; see Methods).the 1-hour long epoch was divided into 4 - 15-minute chunks. The triplet histograms (5ms bins)revealed that some neuron triplets have strong histogram peaks (Figure 66) that are near the∆t=0 bin - which is expected as during synchronized state many neurons spike together. However,a systematic analysis of triplet histograms was not further pursued given the large amounts of datathat needed to be analyzed: a 1-hour cortical recording of 60 neurons results in 216,000 triplethistograms. Importantly, results presented in the above section showing that latencies drift overtime - calls into question whether strong peaks could be detected given that relative spiking ordercould be changing. Additionally, there are some concerns in the spike-time neural coding literatureregarding the correct statistical tests to be implemented to account for randomly occuring peaks intriplet histograms (Oram et al., 1999; Baker and Lemon, 2003; Hatsopoulos et al., 2003; Gersteinet al., 2012; note: most of these studies look at sparse firing during desynchronized, i.e. awakeand/or behaving states).However, as the neuron spiking considered here comes largely UP-state transitions (i.e. duringsynchronized state recordings) the number of spikes in the bins are much greater than those cap-ture during desynchronized state recordings considered in previous studies (i.e. the results here aremore likely statistically significant). The triplet-histogram was also applied to additional GCaMP6mouse visual and barrel cortex recordings where longer periods of spontaneous activity were avail-able (Figure 67; note: data comes from Cre-Emx1-dependent GCaMP6 mice). In some of those142recordings, neurons were identified that exhibited strong peaks near t=0ms that drifted up to 30msover ≈ 60 to 90-minutes confirming that while neurons can fire in precise temporal relationships,those relationships can change over time.Peak latencies change systematically over timeThe findings above suggest that most, if not all, neurons have prefered peak latencies (or PLs) nearUP-state transitions and that such latencies are stable on short time scales but change over longerperiods of time. If PLs are relevant to stimulus processing, then the temporal relationships betweenindividual neuron latencies during UP-states should also be present in neuronal activity occuringduring stimulus processing. Several recent studies have found this, albeit only for high firing rate(>2Hz) neurons in rat somatosensory and auditory cortex where FWHM latency distributions forsome neurons could be up to 100ms wide (Luczak et al., 2007, 2009, 2013; Bermudez-Contreras etal., 2013) in contrast to visual cortex distributions where they are≤10ms. These studies hypothesizethat high firing rate neurons fire in packets which act as gating mechanisms to prepare higher-ordersensory cortex areas for the arrival of new low sensory cortex information (Luczak et al., 2015).The findings presented here, however, suggest a more complex picture where PL-based neuralcode may not be stable on longer time scales. This presents a challenge to the notion that high-order areas only have to learn PL-order of lower areas if such order changes over time. It wouldthus be important to investigate how any putative multi-neuron neural code based on PLs mightalso change over time.Evaluating the stability of a putative multi-neuron peak latency (PL)-based neural code firstrequires to computation of PL stability for each neuron and then visualizing PL stability for allneurons using dimensionality reduction. A 2-hour GCaMP6s mouse visual cortex recording wasrandomly chosen and 9 stable neurons (i.e. high PTP-voltage amplitude and stable firing rates) wereidentified (Figure 68A; note: none of the cat visual cortex recordings contained spontaneous activityperiods longer than about 45-minutes; see also Methods). PLs for each neuron were computed atevery minute of the recording using a 30-minute sliding window approach (see above) using eventsfrom the 2 LECs with the highest amplitudes (Figure 68B: colours represent the largest (red) andsecond-largest (blue) amplitude LECs). The relative PL time for neurons computed against the 2LECs reveal a similar but not identical latency order is present between the two LECs (Figure 68B).Conceptually, the relative PL order of the neurons at each point in time describes a putative firing-order neural code available for stimulus processing - in partial agreement with previous hypotheses(Luczak et al., 2015; the difference here is that all neurons’ relative spike is preserved, whereasprevious work suggests only high-firing rate neurons have preserved firing order). Most neuronsappear to have a stable latency order that changes minimally over many minutes (in part consistentwith previous findings in rat cortex).The next step is to convert the PLs for each neuron at each point in time into a 9-dimensional(i.e. 1 dimension per neuron) PL-based vector and visualize the trajectory of that vector over the143duration of the entire 2-hour recording (Figure 68C). The PL-space vector trajectory (using thefirst two components of PCA space) revealed that the two LECs defined different PL-based neuralcoding spaces across time. Perhaps more importantly, the highest amplitude LEC (Figure 68B,C:LEC #1 - red colour) showed a systematic - but gradual (i.e. non-random) - change over time(Figure 68A: lighter shading indicates earlier points in recording time).This finding suggests that any putative PL-order neural code would have to change along withthese - relative spiking order - relationships and a more systematic investigation would be warranted(e.g. more neurons, recordings and animals). Although mouse sensory cortex (i.e. visual, barreland auditory) data had been acquired specifically for this purpose, it was important to avoid thepotential issues discussed above in relation to Cre-dependent GCaMP6 mice (Steinmetz et al.,2017). Accordingly, the remainder of the analysis was carried out on cat visual cortex recordingsand additional non-GCaMP6 mouse recordings obtained from an external lab (see Methods).The approach implemented for tracking changes in multi-neuron PL state-space was to repeatthe analysis above (Figure 68). However, non-overlapping recording periods (spontaneous and stim-ulus evoked) were used to preserve the independence of the compared datasets. Additionally, assome of the comparisons were made across high-dimension distributions (e.g. 9 to 64 dimensions)more appropriate distance metrics were employed: the earth mover’s distance (EMD), also knownas Wasserstein metric (https://www.encyclopediaofmath.org/index.php/Wasserstein metric), which isa method for computing the distance between probability distributions; and hyper-angles betweenhigh-dimensional vectors. Both of these measures have been previously employed in neurosciencestudies: EMD has been previously used to measure distances in concept space in studies of con-sciousness (Hoel et al., 2013; Oizumi et al., 2014); and hyper-angle values were used to measure thedifference between firing rate space vectors in two-photon studies (Carrillo-Reid et al., 2015; Milleret al., 2014). An advantage of hyper-angle values is that they are less sensitive to changes thatoccur in a single dimension (e.g. a single neuron’s PLs) as compared to euclidean distances. Forexample, if in a 10 neuron recording it is found that 9 PL values are stable across two recordingepochs, but the 10th PL changes by a large amount, the euclidean distance will be biased by sucha change, whereas hyper-angles tend to be less sensitive and yield a value consistent with the factthat most neuron PLs were stable.These approaches were initially implemented in a 3.6hr cat visual cortex recording period con-taining 8 recording epochs - 3 of which were spontaneous activity (Figure 69; see Methods). The 8recording periods were concatenated and spike sorted together and the LFP events were clusteredinto LECs. The PLs relative to the highest amplitude LEC) of 23 stable neurons (i.e. neurons withPTP>80µV that fired throughout the entire 8-recording period) were computed in each recordingepoch. The PL order for the three spontaneous recording periods was visualized using PL orderdisplayed in the first recording period (Figure 69A; note: a fourth spontaneous recording periodwas present, but few neurons fired during that period and it was thus excluded from analysis).The PL display order over the three spontaneous recording epochs indicates a gradual change in144Figure 68: Peak latency-space analysis. A. Firing rate distributions for 9 neurons from arandomly selected mouse visual cortex recording (see main text; see also Methods). B. PLs forneurons in (A) reveal similar but not identical latency times relative to 2 different LECs (tracecolour represents firing rate of neuron as in (A)). C. PLs visualized in a PL-space (using PCA)reveal differences between latency times across different LECs (light shading: earlier in time; darkshading: later in time). D. PL-space plot for a single LEC from a different 300 minute long recordingin mouse visual cortex visualized using PCA reveals similar drifting codes over time. E. Same dataas (D) but shuffled in time yields a randomized distribution.145PL for some - but not all neurons. It is important to remember that the PLs were computedfrom completely separate recordings yet there is substantial similarity for many PLs even across2-3 hours of recording time. Qualitatively, this result is largely consistent with findings above thatshow PL drift is slow but can be larger in some neurons relative to others.Further tracking the PLs for each neuron across the three recording epochs shows that betweenthe last two recording epochs some neurons changed their PLs substantially (Figure 69B). Next,the ∆PL changes for each neuron were computed and plotted as histograms. Given 3 spontaneousrecording periods there are three possible pair-wise comparisons to be made (Figure 69C). Asexpected, the mean of the ∆PL distributions increased with greater time between recording epochs(Figure 69C, dashed lines indicate mean of distributions: 1.7ms, 5ms and 6.7ms for 53, 108 and 161minutes separation, respectively). Lastly, PLs were converted into 23-dimensional vectors and thedistance between each pair of vectors was computed using a cumulative ∆PL(Figure 69D) metric:the ∆PL values for each neuron across times were summed into a single representative value. Thescatter plot (and linear regression fit) also confirm that epochs that are farther apart also havelarger cumulative latency differences. In addition to the cumulative ∆PL metric (Figure 69D),the EMD between PL vectors at each epoch was computed with a nearly identical result to thecumulative latency metric (Figure 69E; see Methods).Next, a similar analysis of ∆PL changes was implemented over larger numbers of recordings todetermine whether similar trends would be observed (Figure 5.23). The first step was to identifypairs of spontaneous activity recordings that were <3hrs apart (i.e. to reduce potential errors fromspike sorting over longer periods of time). All cat visual cortex recordings (5 cats, 15 tracks, 410unique recording epochs) were examined specifically for periods of (i) synchronized cortical statescoupled with (ii) pairs of nearby (i.e. <200min) spontaneous recording epochs (Figure 48-Figure 50:spontaneous recordings annotated with black colour). As indicated at the beginning of this chapter,the cat visual cortex recordings were acquired in an unbiased fashion, i.e. not specifically aimedat addressing questions of synchronized state PL drift across time. Considering all available catvisual cortex recordings, 6 recording blocks in 6 tracks (i.e. multi-recording epochs) were identifiedin 2 cats that fit the requirements of containing both synchronized state and spontaneous activity(Figure 48 - recording IDs: C3.2, 3.3, 3.4 - isoflurane + N2O anesthetic; Figure 51 - recordingIDs: C5.2, 5.3, 5.4 - propofol + fentanyl anesthetic). Within each data block, recordings selectedcontained both spontaneous and stimulus evoked periods. The recordings in each data block wereconcatenated, UP-state transitions were clustered into LECs and single units were sorted. Ad-ditionally, stable neurons (i.e. neurons with PTP amplitudes >80µV and firing throughout therecording block) were identified (# of neurons: C3.2: 17; C3.3: 23; C3.4: 13; C5.2: 23; C5.3: 11;C5.4: 15; see also Methods). The results are visualized as scatter plots of PL similarity vs ∆timebetween pairs of recordings using both EMD (Figure 70A) and hyper-angle (Figure 70B,C) metrics.The data was plotted separately for propofol and isoflurane recordings. The metrics were computedfor pair-wise comparisons across: (i) two spontaneous recording epochs, (ii) a spontaneous record-146Figure 69: Additional methods for tracking latency drift A. PLs for 23 neurons across3 spontaneous recording epochs. B. individual neuron latencies at each epoch shown in (A). C.Distribution of ∆latencies for all neurons between pairs of spontaneous recording periods. D. Cu-mulative ∆latency metric capturing the sum latency changes across all neurons between recordingepochs. E. Same as (D) but differences between epochs were computed using an EMD metric onPL values in each epoch (see main text; see also Methods).ing epoch and a stimulus presentation epoch and (iii) two stimulus presentation epochs (Figure 70:black, magenta and orange, respectively). With one exception (discussed below), the results arelargely as expected: ∆PL is greater for recording epochs that are farther apart (Figure 70A,C: linesrepresent linear fit to each dataset). Both isoflurane and propofol recordings showed this trend ofdecreasing PL similarity with time across both EMD and inter-epoch hyper-angle metrics.The one exception was from stimulus recordings under isoflurane anesthesia where PL differencesbetween two epochs were larger on shorter time scales and more similar on longer time scales(Figure 70A, C: isoflurane plots, orange dots and traces). This trend was also present in probabilitydistribution comparisons (discussed below) and was likely not an artifact due to insufficient datapoints. The most likely explanation has to do with the dissimilarity of neighbouring stimulusrecordings. In particular, specific stimulus types (e.g. a natural scene or drifting grating visualstimulus) were often repeated during the recording blocks considered, but not immediately afterone another - thus resulting in stimulus epochs that were closer together being more likely tobe of different types as opposed to stimulus epochs that were further apart. For example, the147Figure 70: PL similarity across spontaneous and stimulus recordings. A. EMD values ofPL similarity between pairwise recordings in cat visual cortex recordings acquired under propofol(left) or isoflurane (right) anesthesia in two cats (C3.2, 3.3, 3.4 - isoflurane/N2O; C5.2, 5.3, 5.4 -propofol/fentanyl). The colours represent the types of recording periods used for the comparison(see inset). B. Hyperangle computation involves identifying the peaks of each latency of eachneuron (left), building a N-dimensional vector (where N is the number of neurons) and computingthe inter-epoch hyper-angle between the N-dimensional vectors (right; note: the vectors are onlyvisualized in 2 dimensions). C. Same as (A) but using a hyper-angle metric (see Methods).recording block from C5.2 (see Figure 51 black and blue alternating period) had four spontaneousrecording periods that were interspersed with two unique movie stimuli - but the unique moviestimuli alternated so that different movie epochs were closer in time similar movie epochs (i.e. thepattern was: movie #1, spontaneous, movie #2, spontaneous, movie #1, spontaneous, movie #2,spontaneous). However, this exception seems to have a dependency on anesthetic (i.e. present inisoflurane but not propofol based anesthesia) and should be further investigated using specificallydesigned experiments (i.e. longer spontaneous recordings with identical stimuli interspersed in bothcat and mouse visual cortex).Lastly, PL latency similarity over time was also computed using LEC triggered spiking dis-tributions (as opposed to only peaks of distributions, i.e. PLs in Figure 5.24). Briefly, spikingdistributions were computed for each neuron within each recording epoch considered above (i.e. allspikes occuring ±100ms from all UP-state transitions were pooled into a distribution). For example,in recording ID C5.2 which had 8 epochs (see Figure 51), for each of the 23 neurons recorded, theUP-state transition spiking distribution was computed in each of the 8 epochs. Thus, for the each148neuron, there were 8 distributions which reflected when that neuron spiked relative the UP-statetransitions in each of the 8 epochs. The 8 spiking distributions could then be compared against eachother using a KS test (see above; see also Methods) to determine how the distributions changedover time. This test for similarity across time is a more distribution-agnostic measure of similarityand is largely independent from the PL-order tests carried out above. For example, in contrast tothe PL based test above, if some neurons do not fire substantially during some recording epochsthe KS test is more likely to report statistically insignificant results - whereas peak latencies willstill be assigned in each epoch.The KS test was applied on combined isoflurane and propofol data sets (Figure 71). Forspontaneous-spontaneous recording comparisons the results were as expected: neuron firing dis-tributions became increasingly different with increasing time between the epochs being compared(Figure 71A-left: trends show p-values decrease with ∆time between epochs). This is further con-firmed by computing the ratio of statistically significant values (i.e. p-value <10-2) vs all computedvalues in each time bin. This computation is important because most pairwise comparisons fallabove p-value of 0.1 indicating they may not be significant. Accordingly, the ratio of statisticallysignificant p-value comparisons over the total number of comparisons provides a better measure oftrends considering only statistically significant comparisons (Note: p-values were corrected usingBenjamini-Hochberg; see Methods). There results show - as expected - differences appear in dis-tributions increase with separation time between epochs (Figure 71A-right). A similar result waspresent in the spontaneous-to-stimulus recording comparisons (Figure 71B-right; note, the linearfit was only carried out on the first three bins due as the last bin was from recording separation of>3 hrs which were not considered here).However, the stimulus-stimulus recording comparisons yielded the same odd-ball results resultas the PL similarity test across time (see above Figure 70): single neuron spiking distributions weremore likely to be different between closer stimulus recordings than those farther apart (Figure 71C-right).Lastly, single neuron PL drift over time was compared against neuron firing rate and putativecell type (e.g. pyramidal vs. inhibitory) using extracellular template width metrics previouslydescribed (Csicsvari et al., 1998; Bartho´ et al., 2004; Blanche, 2005; Sirota et al., 2008; Niell andStryker, 2008; Mizuseki et al., 2009; Sakata and Harris, 2009; Figure 72). There were no clearcorrelations between neuron firing rate or cell type for all three datasets considered suggesting thatall neuron types are subject to varying changes in latency across time.Overall, these results provide an additional, and partially independent, confirmation that spiking-order is present during individual recording epochs and that changes to order occur gradually overtime. This further supports that during synchronized states PL order is not just a result of averagingover random or epoch-dependent neuron firing order.149Figure 71: Trends in UP-state transition distributions across spontaneous and stim-ulus recordings. A. Left: P-value comparisons between LEC-triggered distributions of neuronsrecorded during two spontaneous neighbouring periods reveals that distribution similarity decreaseswith increased time between the recordings. Right: percentage of pair-wise comparisons that havep-values <0.01 increases with time between the spontaneous recordings suggesting differences aremore substantial for recordings farther apart. B. Same as (A) but for spontaneous vs stimulusrecordings (note linear fit on was done using first 3 datapoints only as the last datapoint reflected>3 hour recording separation not considered herein). C. Same as (A) but for two stimulus pe-riods. (Note: colour same as in Figure 70; all p values corrected for false discovery rate usingBenjamini-Hochberg, see also Methods).150Figure 72: Latency drift vs. firing rate and cell type - cat visual cortex.151Figure 72: (continued from previous page) A. Single neuron latency changes (y-axis) acrossspontaneous recording epochs (x-axis) as a function of firing rate do not show clear trends. B.Same as (A) but as a function of cell type (right scatter plot). C,D. Same as (A,B) but comparinglatency changes between one stimulus evoked and one spontaneous activity epoch. E,F. Same as(A,B) but comparing latency changes between two stimulus evoked epochs. (Note: all x-values arejittered 4 data points (mins) for better visualization).Spiking order is present outside UP-state transitionsThe work in this chapter was aimed at extending previous findings from rat somatosensory andauditory cortex where UP-state transition firing order of high firing rate neurons (>2Hz) wereshown to be relatively (i.e. within ≈10-20ms) preserved over recording periods (e.g. 1-2 hours;Luczak et al., 2007, 2009, 2013; Bermudez-Contreras et al., 2013). The findings above confirmthat virtually all neurons in visual cortex of cat and mouse have a prefered firing order duringLEC-defined UP-state transitions but that such an order changes gradually over periods of 30-120minutes. The analysis above has largely focused on tracking neuron firing order changes duringUP-state transitions across spontaneous or stimulus evoked recording periods. However, previouswork in rat somatosensory and auditory cortex had also focused on stimulus evoked spiking orderoutside UP-state transitions (see Figure 54B). The findings were that UP-state transition firingorder of high firing rate neurons (>2Hz) was also similar to the firing order during the first ≈100msof a stimulus presentation during both awake and anesthetized recordings (Luczak et al., 2007,2009, 2013; Bermudez-Contreras et al., 2013). In those studies the stimuli used were whiskerstimuli, natural sounds (e.g. cricket chirps) or pure tones.It is not obvious how to extend the stimulus-evoked analysis to visual cortex of cat and mouse.One issue is that the data presented here does not contain awake cat or mouse visual cortexrecordings (though see Fig 5.27 for a recording from an external lab). However, designing catexperiments where recordings are made during naturally sleep (or anesthetized) periods followedby awake periods is mostly unprecedented as getting cats to focus on visual targets over time isclose to impossible (note: very recent virtual reality and eye tracking tools may enable this inthe future). While mouse experiments with alternating anesthetized and awake periods have beenpreviously made (by other researchers) that data was not available (though see below for data fromanother lab).Perhaps more challenging, however, is determining how to define a temporally brief visualstimulus that would emulate the natural, but brief auditory and somatosensory stimuli used inprevious studies. The temporal precision of the onset of a somatosensory or auditory stimulus wascritical to determining precise neuronal firing order in such studies. A briefly flashed natural imagepresented during an ongoing natural scene movie or uniform (dark/grey) background is artificialand arguably has no correlate in the natural visual world. Perhaps a natural scene movie thatcontains long stationary periods (i.e. no visual motion) interspersed with brief motion periods may152#Figure 73: Method for detection of ordered synchrony. A. Detection of synchronouslyoccuring spikes for pairs of neurons involves detection of co-occurring spikes within a 25ms windowand binning the results into two bins reflecting percentage of co-occurring spikes fired by cell #1before cell #2 (+ bin) and percentage of co-occurring spikes fired by cell #2 before cell #1 (- bin).B. Histogram of binned order for co-occurring spikes reveals one neuron fired more often before theother neuron. Difference between bins is shown and is tested for pval<0.01 statistical significance(see main text; see also Methods). Note colour scheme indicates the percentage of co-occurringspikes fired by the first neuron relative to the second neuron.suffice - this may emulate the behaviour of a predatory animal (such as a cat) during huntingwhere the visual field is largely stable except for small areas where prey are moving. But it isnot certain whether this would be the correlate of a brief auditory chirp or whisker deflection orwhether recording from neurons whose RF was located exactly where the motion occurs would bethe equivalent of such brief auditory/barrel cortex stimuli.As such proposed recordings were not available for this chapter, an alternative method wasdeveloped that circumvents the challenges of designing novel experiments (and carrying them out)by focusing on spiking statistics (Figure 73-Figure 76). Specifically, if firing order among neuronsis preserved during both spontaneous and stimulus recording periods then such order should bedetectable in quasi synchronous pair-wise firing order of neurons on time-scales relevant for stimulusprocessing (Figure 73; see Methods). Thus, if such an ordered synchrony hypothesis holds thenneurons that spike within a 25ms window of each other (a window arguably relevant for stimulusrepresentation) should have a biased order that is detectable over longer recordings using simplestatistics (e.g. using a bionomial distribution test; Figure 73). The only assumption made here isthat the width of the window that is relevant to stimulus processing is 25ms. This value was takenfrom earlier findings showing that during an UP-state transition virtually all neurons recorded firedwith an ≈25ms window of each other (Figure 61B), but tests using 50ms window show similar (ifnot even better) results as those presented here. Applying the ordered synchrony method to arandomly chosen 60 minute awake mouse visual cortex recording split into 2 halves reveals that153Figure 74: Ordered synchrony matrix - awake mouse recording. A. Ordered synchronymatrix for 68 neurons from a 60-minute awake mouse visual cortex split into two 30-minute epochsreveals some neurons fire in relative order to other neurons (i.e. similar vertical/horizontal lines).Note: all pairwise comparison values are shown (i.e. even statistically insignificant values; see alsomain text). B. The stability of firing order for all pair-wise comparisons across time plotted as the% spikes fired by one cell before the other during the first 30 minute epoch (x-axis) vs the second30 minute epoch (y-axis) reveals some firing is preserved (blue line is linear fit). (continued on nextpage). 154Figure 74: (continued from previous page) Note: all pairwise comparisons are plotted. C. De-termination of pairwise order (irrespective of % spikes fired) across the 2 30-minute epochs revealsa substantial bias twoards firign order being preserved. D,E. Same as (B,C) except only for datawith p-values < 0.01 (see Figure 5.25).many neurons have substantial bias in the firing order across the two halves (Figure 74A; note:data provided by an exernal lab; see Methods). Comparing the relative spiking order between thefirst 30 minute (Figure 74B: x-axis) and the second 30 minute (Figure 74B: y-axis) reveals a biastowards preserved order (see linear fit). The preserved bias in firing order between the two epochsis statistically significant (Figure 74C - binomial test p value 3.8 x 10-25). Considering only thestatistically significant ordered firing relationships (i.e. firing order p value is < 0.01) reveals aperfectly preserved firing order between all such pairs of neurons (2.3 x 10-10; Figure 74D, E).The presence of ordered synchrony is indirectly supported by several two-photon imaging stud-ies which have identified synchronously activated ensembles of neurons in L2/3 of mouse visualcortex over 100-500µm regions (Cossart et al., 2003; Miller et al., 2014; Carrillo-Reid et al., 2015;Yuste, 2015). Given the lower temporal resolution of two-photon (i.e. calcium events generallyreflect bursting neural activity over hundreds of ms) synchronous ensemble activation observed intwo-photon calcium recordings may in fact be a correlate of ordered firing during UP-state transi-tions and stimulus onset.Most neurons exhibit ordered synchrony with at least one other neuronThe methodology described above was next applied to 2 sequential cat visual cortex recordings(one spontaneous recording followed by a natural scene movie stimulus) acquired during a syn-chronized cortical state (Figure 75; recording ID: C5.3). Because this was a synchronized staterecording, before computing the synchrony matrix all spikes occuring during UP-state transitionswere removed (i.e. spikes occuring within a±100ms window of an UP-state transition were masked).This was done to focus on order occuring outside of UP-state transitions - as order during UP-statetransitions had been previously explored at length (see previous sections).The recording time of the spontaneous recording was ≈16minutes and that of the followingnatural scene movie was ≈36minutes (Figure 75A). Computation of the ordered synchrony valuesfor all pairwise comparisons across the 23 recorded neurons revealed that during the spontaneousrecording period 11/23 neurons had statistically significant (p value <0.001) biased firing order withat least one other neuron (Figure 75B-left; note: longer spontaneous activity recordings would likelyyield more spikes and higher percentage of significant relationships due to nature of binomial test).During the following natural scene movie recording 21/23 neurons had a biased order (Figure 75B-right). Similar results were found across other recordings, with usually >50% of all neurons havingan ordered synchrony relationship with at least on other neuron. This is a remarkable result giventhat only 10-30 neurons were compared at a time in the recordings, suggesting that likely all neurons155Figure 75: Ordered synchrony preserved between spontaneous and stimulus evokedrecordings. A. Cat visual cortex synchronized state recording containing neighbouring sponta-neous and natural scene movie recordings. B. Ordered synchrony matrix for all 23 neurons recordedreveals similar spike order across many neurons (only data points with p value <0.001 are shown).C. Ordered synchrony during the spontaneous activity recording is abolished when jittering thespikes of all neurons by 25ms (see also Methods). D. Stability of of firing order for all pair-wisecomparisons during the recording block (C5.3 had 5 such pairwise epochs) for spontaneous vs.stimulus evoked periods reveals that order is substantially preserved between sequential epochs(blue line is linear fit; see also Figure 73B). E. Determination of pairwise order (irrespective of% spikes fired) across all pairwise epochs considered in (D) reveals almost perfect preservation offiring order (note only data p values <0.01 were shown and considered in D,E).156Figure 76: Similarity of ordered synchrony across spontaneous and stimulus evokedperiods A. Left: distribution of inter-epoch hyper-angles between ordered synchrony vectors ofneighbouring recordings in recording presented in Figure 68 reveals a peak in the 0-30 degree bin,i.e. spiking order was similar between two neighbouring epochs. Right: same as (left) but forshuffled vectors reveals largely orthogonal values (see also Methods). B. Left: same as (A-left) butfor all recordings presented in Figure 65 and Figure 66. Right: same as (A-right) but for data in(B).in cortex will have a biased ordered relationships with many other neurons (results not shown).Jittering the spikes of all neurons by as little as 25ms abolished most or all statistically significantordering observed in the original data (see Figure 75C). Carrying out the same analysis as above(see Figure 74), reveals that the statistically significant relationships are almost perfectly preserved(117 of 118 pairs) across the two recording epochs (Figure 75D,E).These last tests using awake mouse and anesthetized cat cortex (see above and Figure 74D,Eand Figure 75D,E) confirm that firing order is largely present and conserved over periods of 30minutes or less and that such order is lost when the temporal structure in the original (i.e. real)data is modified (e.g. Figure 5.28C).In the last example provided above, some of the rows and columns of the ordered synchronymatrix looked qualitatively similar across spontaneous and natural scene recordings (Figure 75B).Accordingly, an additional test of similarity between neighbouring spontaneous and stimulus evokedrecording periods was implemented using multi-dimensional metrics (Figure 5.29). The goal wasto evaluate the similarity of ordered synchrony of all neurons across short periods of time (i.e.≈60-minutes or less). The hypothesis was that ordered synchrony relationships would be preservedbetween neighbouring recordings relative to a shuffled condition. This would provide further (UP-157state transition independent) confirmation that spiking order is present in visual cortex.For this test two values were compared for each neuron: the N-dimensional vector representingthe spiking order of each neuron relative to all other neurons in the first vs. the second recording.For example, for a 23 neuron recording, the ordered synchrony vector for a neuron was computedduring a spontaneous recording and then compared against the vector for a stimulus recording.This amounts to comparing the same row from two synchrony matrices. For these comparisons ahyper-angle distance metric was implemented (see above; see also Methods).Computing the distribution of inter-epoch hyper-angles for the example provided above (see alsoFigure 70) revealed that synchrony order vectors across neighbouring epochs were more likely to besimilar to each other (Figure 5.29A - real data). Shuffling the vectors resulted in a distribution thatpeaked in the fourth bin (90-120 degrees differences) suggesting that, after shuffling, most vectorswere orthogonal to each other. (Note: this result is partially explained as a random rotation of asparse high-dimensional vector is more likely to result in a largely orthogonal new vector).This method was next implemented on all data blocks considered for the previous section(i.e. 6 multi-recording blocks from 2 cats; Figure 70, Figure 71). The goal was to determinehow any potential order during a spontaneous recording period compares during a subsequent (orpreceding) stimulus recording. Thus, recording pairs considered had one spontaneous and onestimulus recording. The resulting distributions revealed that in the real data there is a strong peakin the 50o histogram bin suggesting some differences were present between epochs, but overall thedistribution was different from a shuffled condition which peaked in the 90o bin (2-sample KS testpval = 2.3 x 10-26).Measuring spiking order using adaptive coincidence detectionIn the last two sections of this chapter two other broad approaches for evaluating firing order arebriefly discussed in relation to the work presented herein. As reviewed above, the search for timestructure in neuronal activity has lead to the development of several analytical methods: searchingfor synfire chains using triplet histograms (e.g. Abeles and Gerstein, 1988; Abeles, 1982a), identify-ing neuronal firing order during UP-state transitions (e.g. Luczak et al., 2007, 2015), identificationof synfire-chains using statistcal methods (e.g. Quaglio et al., 2017) and temporal-scale independentmethods for identifying repeating sequences (e.g. Russo and Durstewitz, 2017).An alternative approach to detecting temporal structure in time-trains (including both neuralspiking data and other naturally occurring phenomena) has been proposed which is based onadaptive coincidence detection (Kreuz et al., 2009, 2011, 2013) previously introduced as eventsynchronization (Quiroga et al., 2002). Adaptive coincidence detection is a method for measuringoverall (normalized) synchronization between two time series - including spike train rasters. Briefly,given two spike trains, the method matches each spike from one train with at most one spike in theother train and tracks order during such coincidental firing. The method considers only pairs ofspikes from two neurons that are closer to each other than to any other spike from the individual158neurons. For example, if one neuron fired one spike 10ms before a second neuron fired its spikeand neither neuron fired another spike (previously or after) for at least 10ms, then those two spikes(10ms apart) would be considered coincident and their inter-spike-interval (ISI) and order wouldbe retained for further analysis. In contrast, if one neuron fired a spike 10ms before another neuronfired two spikes 5ms apart, then all spikes would be eliminated from consideration. This is becausethe second neuron fired (temporarily) at a higher frequency (e.g. 200Hz) than the two neuronspike sequence (100Hz) and the second neuron is thus more related to its own spike train than tothe first neuron (see below for caveats to this definition of spiking relationship). The method isheuristic- and parameter-free since the ”local spike rates” provide the only parameters required forthe method.Using this ”adaptive” approach, three metrics were previously developed to capture differentproperties of pairs of spike trains: ISI-distance, SPIKE-distance and SPIKE-synchrony (Kreuz etal., 2015). These metrics are largely aimed at temporal-scale independent valuation of spike trainISI similarity, overall spike train similarity and spike train order. Briefly, SPIKE-synchronizationcaptures a (symmetric and normalized) measure of synchrony across all spikes from pairwise com-parisons where a coincidence in spiking is defined as above. SPIKE-distance is another normalizedmetric that estimates the dissimilarity between spike trains (somewhat similar to Victor-Purpuradistance Victor and Purpura, 1996) with low values (i.e. close to 0) indicating very similar spiketrains, and high values (i.e. closer to 1) indicating very different spike trains. Lastly, ISI-distancemeasures interspike interval information through the ratio of instantaneous firing rates of two spiketrains.The metrics discussed above were recently implemented in a freely available Python toolboxcalled SPIKY (Kreuz et al., 2015). SPIKY was implemented on three randomly selected recordingsin the datasets presented above (>1 hour in length) and metrics were computed for non-overlappingrecording epochs of ≈25-30 minutes (Figure 77). The goal was to qualitatively assess how themetrics captured firing order for different datasets and whether they could potentially quantifyorder changes over time. The results were not further analyzed (for reasons discussed below)but two important conclusions can be drawn from the qualitative plots (Figure 5.30). First, thecoincidence metrics indicate that non-overlapping neighbouring epochs have largely similar values.This suggests that the metrics capture similar values over time and do not reflect random propertiesof neuron trains as shuffling spike times (not shown) yields matrices that are substantially differentacross time. Second, and more relevant to work presented here, the matrices reveal small changesover time. Such changes may reflect the same neuronal firing order changes found in previoussections.While adaptive coincidence detection methods could be investigated further (e.g. tracking met-ric value changes over >30 minute periods using hyper-angle or other types of distance measures), itis unclear whether an adaptive coincidence mechanism is employed in cortex. Perhaps the greatestobstacle to further use of these methods is that many if not most spikes from pairs of neurons are159Figure 77: Local time-window metrics of neural synchrony. Adaptive coincidence metricsimplemented in 3 visual cortex recordings reveal similar, but slowly changing synchrony values overtime.160excluded from computation due to the strong exclusion criteria of the adaptive firing rate method(see above; see also Kreuz et al. (2009)). This would imply that many or most spikes in cortex(and other areas) do not form part of a locally implemented temporal neural code (though it ispossible, for example, that locally recorded neurons could be adaptively connected to other moredistant neurons that were not recorded). Such a definition of coincident firing excludes almostall spikes fired during a burst. For example, if one neuron fires one spike 10ms before a secondneuron fires a burst of spikes (with the first spikes in the burst <10ms apart) then not only arethe two neurons’ spikes not considered as coincident (even though they fired merely 10ms apart),but all the spikes occuring during the burst must be discarded (as they will not fulfil the requiredcoincidence criteria). The notion that most (tonic or burst) spiking is not relevant to neural codingis not a common opinion on temporal coding, and it is very difficult if not impossible to test: itwould require recording most or all neurons in an organism to determine the amount of adaptivecoincidence employed in cortex. Additionally, the metrics above were also not developed specificallyfor application to neural data and contain other assumptions (not discussed here) as well as yield-ing normalized and unit-less values which are challenging to directly interpret and compare withother measures of synchronization where time units (e.g. ms) are used. Future work on adaptivecoincidence detection may provide useful for analysis of neuronal spike trains with some possiblechanges (e.g. fixing coincidence windows, or implementing time-unit based metrics; personal com-munication with T. Kreuz) but given the time limitations of the current work and the outstandinginterpretation issues, it was not pursued further as of the time of writing.Measuring spiking order using pair-wise cross-correlogram order and theprinciple of additivityThe last approach discussed here for detecting neural firing order in cortex involves identifyingfiring sequences by fitting gaussians to 1ms bin cross-correlogram histograms (CCH) and convertingthe resulting 2-dimensional histograms to 1-dimensional linear sequences using the principle ofadditivity - i.e. assuming that pair-wise firing order is also present in multi-neuron sequences(Schneider and Nikolic, 2006; Nikolic´, 2007; Havenith et al., 2011). Perhaps the simplest way todescribe this method is to note that it is a solution to the problem of converting 2-dimensionalpair-wise order matrices generated in some examples above (e.g. matrices shown in Figures 5.27,5.28) into relative 1-dimensional firing order sequences between the recorded neurons.The approach for generating (1-dimensional) multi-neuron firing sequence from all pair-wiseorder (i.e. CCH) measures has been described in detail previously (Schneider and Nikolic, 2006;Nikolic´, 2007; Havenith et al., 2011) and here the steps are described briefly and applied (in simpli-fied form) to one example from a cat visual cortex recording during a drift bar stimulus presentation.As explained below, the steps in generating neuron sequences are complex and involve tracking oferrors, application of several heuristics and statistical solutions to fitting gaussian distributions tosparse data.161There are two stages involved each containing multiple steps: (i) computing gaussian fits forpair-wise CCH for all neurons and (ii) converting the CCH matrices to firing sequences (Figure 78).The first step requires removing stimulus-locked rate covariation from the neural responses usuallyby determining when a stimulus (e.g. a drift bar) enters and exits a neuron’s RF and excludingthose periods before and after from analysis, respectively. Next, 1ms-bin CCHs are computed for allneuron pairs, but for statistical reasons (i.e. fitting gaussians) only pairs of neurons with at least 3 -consecutive 1ms bins that contain >8 entries each are considered (Figure 78A). As not all stimulustypes evoke substantial spiking from all neurons (e.g. a non-preferred drift grating orientation)additional heuristics must be applied to exclude non-optimal stimuli from consideration. In aprevious study it was found that on average only 3.5 out of 8 possible drift grating orientationscould be used to generate sufficient spiking data (Havenith et al., 2011). Last, rather than fittinga gaussian to each CCH distribution, a correction must be implemented to address the ”lackingdistributivity of errors” (see Havenith et al., 2011 for in depth explanation). This is achieved byrandomly splitting each CCH distributions into 2 halves, fitting gaussians to each half and repeatingthe process 100 times on the randomly selected data. The final gaussian distribution is obtainedby averaging over the 100 gaussian fits.The second stage involves computing firing sequences from the CCH matrices derived above(Figure 78B). This is done by assigning each neuron a relative time position based on the averageof its CCH peaks (i.e. relative spiking times) with other neurons. For example, if, on average,neuron A fires 5ms (i.e. has a CCH peak) before neuron B and 10ms before neuron C, it’s relativetime position will be 7.5ms ((5ms+10ms)/2). Additionally, if neuron B fires on average 6ms beforeneuron C, its relative time position will be 0.5ms ((-5ms+6ms)/2). Neuron C’s relative time positionwill be -8ms ((-10ms-6ms)/2). Thus the 1D spiking sequence generated from the pairwise CCHswill be - A: 7.5ms, B: 0.5ms, C: -8ms (note: sum always should =0ms). There are a number ofcontrols and error computations that can be done to test the validity of the additivity postulate(for complete details, see Havenith et al., 2011). One type of error is called additivity error and itmeasures the error between the 1D spiking sequence time differences and the original CCH peaks.For the example above, this is computed in a few stages. First, time differences between 2D and1D time sequences are computed. The time differences for neuron A are A-B: 3ms; A-C: 5.5ms; forneuron B they are B-A: 2ms; B-C: 2.5ms; and for neuron C error they are C-A:5.5ms; C-B: 2.5ms.The individual errors are squared and summed for each neuron to yield a value QAdd. For exampleneuron A, QAdd = 32 + 5.52 = 39.25ms. Lastly, the error is normalized by the number of delayscomputed (for rationale, see Schneider and Nikolic, 2006):σadd =√QAdd2(n− 2)n2(8)While the errors can be substantial, previous studies concluded the average additivity errorswere <1ms (Havenith et al., 2011). One important note is that in the previous study of firing162Figure 78: Computing firing order using CCHs and the principle of additivity. A.Pairwise CCHs computed for 7 neurons during a drift-bar experiment reveal the presence of firingorder bias across the neurons. B. Using additivity, the pairwise CCHs are converted to a firingsequence. C. Firing sequences for the same stimulus can be different for recordings 7 hours apart.Adapted from Havenith et al., 2011 with permission.163Figure 79: Local time-window metrics of neural synchrony. A. Example CCHs withsignificant peaks in the ±25ms bin for neurons from a drift bar recording in cat V1 (C3.2). B. PLsfor the same neurons in (A) from a neighbouring synchronized state spontaneous activity recording.C. Comparison of UP-state latency difference (x-axis) vs. CCH peak (y-axis) reveals a correlationbetween firing times across the two methods (blue line is linear fit and CCH peaks were normalizedto sum to 0ms). D. CCH-based 1D firing sequence for 3 neurons (IDs: 2, 15, 18; x-axis) vs. UP-state transition based order (y-axis) also reveals a strong correlation between neuronal firing orderusing the two methods (note: CCH peaks and UP-state times were normalized to sum to 0ms; seealso main text).sequences generated from CCH order, changes in firing order were observed for the same stimuluspresented 7 hours apart (Figure 78C). While those results were not pursued further at the time ofthe publication, such firing sequence changes are consistent with the findings of UP-state transitiondrift over time.Given the findings of this chapter on UP-state generated neuron firing order for all neurons - the164best implementation of the additivity based approach would be to compare how UP-state transitionorder relates to stimulus driven order determined using CCHs. The ideal experimental paradigmshould thus involve a synchronous state recording period (e.g. anesthetized or SWS period) -during which UP-state based firing order can be determined - followed by a desynchronized state(e.g. awake, or NREM) period during which various stimuli can be presented and additivity basedorder can be computed. A desynchronized state is ideal because firing order detected duringsynchronized states using the CCH method may nonetheless reflect firing order during UP-statetransitions. This requires a spontaneous activity recording adjacent to a drift bar or drift gratingrecording during a synchronized state that changes to a desynchronized state during the stimuluspresentation. Measuring LEC based firing order (i.e. PLs) could then be compared with the CCHgaussian fitting approach described above. Given time and space limitations on the work presentedhere, this method was not fully implemented in this section (note: it will be pursued in additionalproposed work for publishing the work of this chapter). However, it should be noted that inprevious studies CCH approach was only implemented in anesthetized cortical recordings (withoutclear indication of cortical state measures) during preferred stimulus orientations as it requiredsubstantial amount of spiking for the neurons considered and it usually only identified 7 neurons(on average) per recording that could be converted into a firing sequence. In contrast, UP-statetransition order identified here can provide a preferred order for the vast majority of neurons evenduring spontaneous activity periods when spiking is very sparse.Across all available recordings, only one synchronized state recording example was found wherea spontaneous activity recording was next to a drift bar recording (C5.4). For that pair of record-ings, a simplified CCH-based firing sequence approach was implemented to determine if neuronalfiring order might be similar between the two recording periods (Figure 79). First, CCHs were com-puted during the drift bar experiment for all pair-wise neurons (23 neurons in total) and CCHs withsubstantial peaks were selected (Figure 79A; total unique neurons: 11; note: stimulus ON/OFFcovariation was not corrected for). Next, UP-state transitions latencies for the spontaneous activ-ity recording were computed for all neurons (Figure 79B; note: not all neurons fired during thespontaneous activity period to generate PLs). CCH-peak based order was then compared with theUP-state ∆latency order which revealed a correlation for several neurons with a linear fit that wasclose, but not identical to the x=y line (Figure 79C; note: CCH peaks sum was normalized to =0msas per Havenith et al., 2011). Lastly, 5 neurons were identified which had inter-neuron CCHs withclear peaks (IDs: 2, 11, 15, 18, 21) and their firing sequence was computed as per Havenith et al.,2011 and compared with UP-state firing order (Figure 79D; note: only 3 of the neurons had PLsthat were within ±25ms to match the CCH window; see also figure text). The CCH-based firingsequence order also shows strong correlation with UP-state latency based order.While some corrective steps were skipped in this analysis, this result does indicate some cor-relation between firing order across the two different methods. Overall, the additivity postulate ispromising and future investigations should consider it in complement to UP-state transition based165ordering as an additional way to corroborate the presence of spontaneous and stimulus driven spikeorder as well as confirming order drift over multi-hour recordings.DiscussionIn this chapter it was shown that during synchronized cortical states, multi-channel LFP recordingsin mouse and cat cortex contain transient large amplitude LFP events lasting ≈50-200ms that canbe clustered on the basis of their multi-channel extracellular waveforms into 1 to 4 distinct classestermed LECs. These large amplitude LFP events have been previously shown to be the correlatesof single neuron UP-state transitions and the methodology used provides a spiking independentdefinition of UP-state transition and can be applied to sparse firing cortical regions includingvisual cortex. The identified LECs and their CSD correlates revealed that UP-state transitionsfall into classes with distinctive laminar signatures and can be grouped within and across animals.LECs were also shown to have mesoscale correlates using VSD recordings. Additionally, almost allneurons showed narrow firing peaks and distributions near UP-state transitions. While both spikinglatencies and distributions were stable for most neurons on the order of <30-minutes, many neuronsshowed changes in their PLs (and distributions) over periods of 30-120 minutes. Lastly, it was shownusing both stimulus and spontaneous recordings, that even outside of UP-state transitions spikingorder is biased for many neuron pairs and that such biased order also appears to change with time.In this discussion section, the findings presented above are further explained relative to existingstudies and some opinions are provided for why temporal order may be present in visual cortex butappears to be drifting over time.LECs are a global, single-neuron independent, temporally precise correlate ofUP-State transitionsSynchronized state DOWN-to-UP state transitions (originally termed K-complexes, Amzica andSteriade, 1998a) have been studied for a few decades with the term UP-state referring to singleneurons membrane transitioning from a hyperpolarized (i.e. non-spiking) to a depolarized (i.e.often spiking) state. Additionally, the term UP-state also refers to the global correlate of UP-statetransitions where the vast majority (or possibly all) cortical neurons transition to depolarized states(Neske, 2016). This ambiguity in terminology is perhaps understandable as it is generally acceptedthat global UP-states do cause all single neurons to depolarize but not all neurons spike on everycycle (Volgushev et al., 2006; Chauvette et al., 2010). This makes it challenging to track UP-statesglobally solely by patching 1-2 neurons. In fact, patching 2-4 neighbouring (e.g. 100-200µm apart)cells shows that even such closely spaced neurons single can enter the UP-state cycle at differenttimes and in different orders (Ros et al., 2009) - consistent with the present results - with membranepotential dynamics during the transition period varying substantially in different cycles (Chauvetteet al., 2010; see Figure 53). As an alternative definition of global UP-state transitions, LECs have166advantages over single neuron patch clamp recordings. First, they take into account spatially broad(i.e. 100µm to 1000µm) LFP contributions from multiple sources (Buzsa´ki et al., 2012) from many(or all) cortical layers. Second, they are independent of any specific single neuron activity and canthus be used as single neuron-independent event triggers. Lastly, the temporal precision of LECevents can be as low as ≈5-15ms based on the stability of the FWMH of max-channel LFP troughor the width of single neuron PETH histograms. In fact, given the stochastic nature of UP-statetransitions (i.e. only some cells participate every cycle), using membrane potential dynamics evenfrom many single patched neurons to define a global UP-state transition time is not only impossibledue to current experimental limitations but must ultimately rely on averaging UP-state transitionon-times for individual neurons. Additionally, single neurons can take >50ms to transition to UP-states (see membrane depolarization profiles in Figure 53) and aligning and averaging over suchprofiles is unlikely to be more precise than an LFP based definition. When coupled with the higherprecision and stereotyped shapes of multi-channel LFP signals during UP-state transitions, LECscan provide a more precise, principled and non-circular definition of global UP-state which can beused to further investigate coding in cortex. While not addressed in our study, UP-state transitionsare known to mediate hippocampus to cortex interactions (Wilson et al., 1994; Sirota et al., 2003;Sirota and Buzsaki, 2005) and such studies will also benefit from temporally precise methods foridentifying UP-state transitions.Multiple LEC types suggests multiple sources of UP-state genesisAlthough it has has been previously shown that UP-state transitions have LFP correlates (Saleemet al., 2010; Chauvette et al., 2010), the method presented here is novel and the findings (e.g. mul-tiple classes of LFP events) are consistent with a growing body of work identifying stereotypy inLFP recordings: LFP recordings in rat hippocampus slices showed stereotyped shapes which couldbe clustered (Reichinnek et al., 2010) and LFP recordings in anesthetized macaque hippocampusalso exhibited LFP shape similarity (Ramirez-Villegas et al., 2015). In fact, using only MUA ac-tivity to define UP-state transitions, a recent study found two types of UP-state transitions inketamine/xylazine-anesthetized rats (Luczak and Bartho, 2012). The findings of multiple classesof UP-state transitions is enforced by CSD correlates which have very distinct laminar patterns -common across cortical areas and different animals - supporting the involvement of discrete corticalcircuits in UP-state initiation. These findings are consistent with suggestions such as the three car-dinal oscillator hypothesis that UP-states can be caused and sustained by potentially independentcortico-thalamic-cortico populations: synaptically-driven cortical populations involving neurons inL4 and L5/6, glutamatergic thalamo-cortical (TC) neurons from multiple nuclei which have intrin-sic oscillatory properties and GABAergic reticular thalamic neurons which also have oscillatoryproperties when coupled with external input (Crunelli and Hughes, 2010). The fact that UP-statetransitions are stereotyped and fall into 1 to 4 classes based on their LFP waveforms - as opposedto falling into a continuum of waveforms - also supports the multiple oscillator hypothesis.167UP-State transitions might engage discrete cortical avalanche circuitsThe findings of discrete CSD patterns correlating with UP-state transitions (i.e. LECs) is unex-pected as while UP-state transition extracellular templates (i.e. LFP averages) are similar in shapeacross multiple cortical layers they can yield CSD templates that are highly heterogeneous (e.g. seeFigure 56B for comparison between LFP template and CSD). This suggests that while UP-statestransitions can be initiated via broad multi-laminar inputs, once activated they recruit multiple(1-4) laminar-specific circuits as they evolve into global phenomena. LEC template troughs, inparticular, are generally the largest and most dynamic (i.e. fastest changing) components of theLFP during UP-state transitions and previous work has connected spontaneous ”negative LFPdeflections” to neuronal avalanches even in awake monkey recordings (Petermann et al., 2009).Cortical avalanche research originated in self-organized criticality studies which sought to char-acterize the dynamics of naturally occuring phenomena such as weather patterns and chemicalreactions (Bak et al., 1987). Self-organizing criticality is a ”property of dynamic systems that havea critical point as an attractor” (https://en.wikipedia.org/wiki/Self-organized criticality). What thismeans is that such systems tune themselves and always seek what are inherently unstable pointsbetween two (or more) states. In the context of neuronal activity, cortical avalanches are generallydescribed as sequences of synchronized neural activity bursts with both spatial extent (e.g. theamount of cortical tissue engaged) and duration (e.g. the length of time they are sustained for)being captured by power law distributions which often fall precisely at the self-organizing criticalityvalue (i.e. 1/f3/2 distributions). Cortical avalanches and power-law distributions have been shownto occur in neuronal cultures (Beggs and Plenz, 2003, 2004), cortical slices (Jimbo and Robinson,2000; Beggs and Plenz, 2003), anesthetized cats (Hahn et al., 2017), awake monkeys (Petermann etal., 2009; Hahn et al., 2017), human EEG (Pritchard, 1992) and other paradigms. UP-state tran-sitions specifically have been studied as phase-transitions and self-organizing criticality in severalmodeling studies (Mejias et al., 2010; Millman et al., 2010; Scarpetta and de Candia, 2014).The findings presented here suggest that UP-state transitions may not only engage distinctcircuits, but that they may be a type of default activity mode that manifests at a certain spatialscale (i.e. all cortical layers). It may be the case that the criticality identified in cortical recordingsin general, may also describe the multi-laminar, possibly multi-cortical area, nature of UP-statetransitions. Future work that focuses on the temporal and spatial dynamics of LFP and single neu-ron activity during UP-states transition may provide a further link between synchronized corticalstates and criticality as it is present in cortex.Spike timing codes might be ubiquitous - but difficult to track due to driftAn ongoing debate in neural coding is how information is represented by the firing of neuronswith much physiological evidence supporting that neuron firing-rates are important as they aremodulated by stimulus intensity (Adrian, 1926; Hubel, 1959; Shadlen and Newsome, 1994 and168many others; but see Hemmen and Sejnowski, 2006 for outstanding questions). Yet, precise neuronfiring times also encode information even to µs precision (Jeffress, 1948) with more recent findingssuggesting 1-2ms spiking precision (in vitro; Mainen and Sejnowski, 1995) and 10-20ms precisionin-vivo during natural scene viewing (cat V1; Fregnac and Zador, 2005). A very recent findingsuggests that information is encoded even in the precise spike timing of spikes within a single burstof thalamic neurons (Mease et al., 2017). Evidence of firing order between cells, i.e. of synfire chains(Abeles, 1982b) has been limited, but the studies of high firing rate neurons in rodent auditory andsomatosensory cortex - that this chapter sought to extend to visual cortex - suggest neurons fire ina generally preserved order during UP-state transitions as well as stimulus presentation (Luczak etal., 2007, 2009, 2013, 2015; Bermudez-Contreras et al., 2013).The findings presented here suggest that it is not just high-firing rate neurons in highly activeareas - but that all neurons in all mammalian sensory cortex including cat V1 and mouse visualcortex may fire in a transiently preserved order during both spontaneous and stimulus evokedperiods. This may mean that neurons spike together temporarily to represent specific informationat various times (see also below). Future work expanding the methodology presented here usinglarger datasets, including naturally sleeping animals and complex stimuli, may yield further insightinto how common temporal or sequence coding is in sensory cortex.LTP/LTD may underly temporal code change over timeThe results presented here indicating that relative firing order changes over time should not besurprising as synaptic processes underlying learning and metabolic activity are likely to constantlyengage long-term-potentiation/depression (LTP/LTD) mechanisms that have as their goal the mod-ification of connectivity between neurons. Importantly, spike-timing-dependent plasticity (STDP;Bi and Poo, 1998; Yao et al., 2004) is highly sensitive to the order of inputs and can make use ofmillisecond precise spiking order. Changes in connectivity driven by underlying processes that arestimulus-independent could thus manifest as changes in spiking order.Future experiments where simultaneous extracellular and two-photon data is acquired in chronicrecordings over several (e.g. 3-5) hours, and over multiple days with possible pharmacologicalinterventions limiting plasticity may further elucidate the relationship between latency drift andLTP/LTD. However, it is also possible that latency drift is related to sleep homeostasis mechanismsthat can only be properly characterized using future technologies where large scale synaptic imagingcan be captured and related to single neuron latency drift.The null-hypothesis should be that firing order changes over timeAlthough temporal order and drift have been shown using different methods, a few additional com-ments are warranted regarding the possibility that PL drift could result from poorly sorted neurons.First, isolating single neurons over multiple-hours of recording continues to be a challenging task169in experimental neuroscience. Accordingly, analysis in several sections in this chapter was limitedto using only large amplitude neurons that fired relatively stably over recording periods of <200minutes in length. While some neurons may have been incorrectly sorted, it is unlikely that mostneurons across multiple datasets contained sorting errors that would lead to the findings madeherein. Additionally, tracking changes in spiking distributions was challenging given that cat visualcortex datasets available were not designed with this goal in mind.While the findings were limited to datasets of ≤30 (stable) neurons which were largely selectedfor their stability, future recordings designed to specifically investigate latency drift (with potentiallyimproved hardware and acquisition paradigms) should allow for the recording of several dozens orhundreds of neurons simultaneously (e.g. neuropixels probes Jun et al., 2017b) across multipleregions within a sensory area. Given that the relatively small datasets considered here alreadyyield trends and statistically significant differences, it is likely that larger datasets will clarifythe results presented here rather than diminishing them. More importantly however, is that thenull-hypothesis for temporal order between spiking neurons should not be that temporal spikingrelationships - if they exist - stay the same over periods of hours or days. Rather, given STDPmechanisms that operate at the millisecond scale, the extraordinary complexity of cortical networkswhich appear to self-organize into scale free and small-world networks, the diverse input to singleneurons (up to 7,000 inputs per mouse visual cortex neuron), the inherent plasticity mechanismssuch as LTD/LTP and the requirement for learning - temporal relationships of single-neuronsare unlikely to be preserved for indefinite periods of time. Such temporal relationships must beinherently malleable if they are to underpin learning and plasticity mechanisms that are are requiredfor organism survival.Put another way - it would be much more surprising and interesting - if temporal relationshipsare present and they are static over periods of many hours or days. This would essentially meanthat a partial Rosetta Stone for understanding cortical coding may indeed be available from suchtemporal relationships. This was in part hypothesized by Luczak et al., 2015 in relation to high-firing rate neurons which show more temporal stability, however, the temporal variability of PLsstability in those studies was commonly ≥10-20ms and drift analysis was not likely not feasible(Luczak et al., 2007, 2009, 2013; Bermudez-Contreras et al., 2013).Non-stationary neural codes may reveal complex metastable attractor manifoldsIn two sections of this chapter it was shown that both single neuron peak latencies and spikingdistributions during UP-states change on the order of dozens of minutes (e.g. 30-120 minutes orlonger). An obvious question arises, namely, if stimulus information is encoded in the relative orderof neurons - but such order is constantly changing - how is it possible for downstream areas todecode this information?The simplest answer is that cortical decoders may also change simultaneously with the firing-order dependent encoders. For example, the same underlying mechanisms that lead information170being encoded in temporal but non-stationarity relationships could also underpin decoders that takeadvantage of non-stationarity. In fact, this is the exact hypothesis of a couple of recent theoreticalstudies namely that transient representation is possible in neural network models and is supportedby neuroscientific evidence (Rodny et al., 2017) and that self-reconfiguration of neural circuits canbe observed as a ”slow drift of network architecture and dynamics” (Kappel et al., 2017).Nonlinear systems concepts such as metastability and self-organization have been around for along time and where introduced to neuroscience over two decades ago (Kelso, 1995). In fact, the ideaof neural code drift is consistent with the notion of a ”dynamically changing attractor manifold”which has been explored by some theoreticians in network models (Friston, 1997). Previous workwas aimed at modeling optimal complexity and found that neural activity ”dynamics are modeled byneither an ensemble of separate attractors nor a simple low-dimensional attractor, but are consistentwith the attractor surface that ensues when many separate attractors are loosely coupled together”(Friston, 1997). Importantly, the work also suggested that an attractor manifold - as might beinstantiated by a temporal spiking order neural code - may only appear to vary due to limitationsin measurement.Recent studies of firing rate state-space manifolds in motor cortex (Gao and Ganguli, 2015)have suggested that recording from ever increasing number of neurons may not necessarily yieldadditional information as a large majority of firing-rate based variances are captured on low-dimensional manifolds using even <100 neurons. In other words, the amount of variance explainedby such low-dimensional manifolds increases asymptotically with increasing number of neuronsbeing recorded in motor cortex (and possibly other association cortical areas).However, the work presented here suggests a somewhat different picture for low sensory cortexwhere an underlying neuron firing order appears to change over time even in the absence of sensorystimulus or tasks (Figure 68C-E).Additionally, it is important to note that the results presented here were based only on verysmall numbers of cortical neurons (i.e. 10-30 neurons for state space analysis and <100 neuronsfor other analysis). A more complete picture, e.g. recording from thousands or tens of thousandsof visual cortex neurons in mice (or cat) may provide a much better description of potential typesof codes and underlying manifolds that could change over time.Future workDespite being present only during behavioural states such SWS, anesthesia or quiescent awakestates, UP-states have become an invaluable tool in investigations of neural coding in both awakeand anesthetized mammalian cortex. The main aim of this chapter was to improve and extend find-ings of single neuron order during UP-state transitions to the visual cortex of cat and mouse. Thefindings of multiple classes of UP-state transition type were confirmed across multiple cortical areasin two species, and using four different anesthetic preparations. These broad experimental datasetssuggest that LECs are a robust and ubiquitous phenomenon in mammalian cortex. The insight of171UP-state neuronal firing order first referenced almost two decades ago (Steriade and Amzica, 1998)has now lead to findings of preserved firing order of high-firing rate neurons (e.g. Luczak et al., 2007)and now can be extended to all spiking - in/outside of UP-states, including desynchronized stateswhich may have no inherent UP-state structure. Further experiments, using preferably chronicrecordings from multiple animals in different cortical areas may reveal distinct cortical correlatesof LECs. Specifically, habituation paradigms and recordings from awake, quiescent animals, andduring SWS may further reveal more nuanced neural codes and latency behaviours. An interestinggoal might be to determine whether PL order resets or cycles (e.g. based on 24-hour circadianrhythms) or whether it changes unidirectionally over time to potentially describe infinitely diversecoding strategies.172Conclusion“One could define the central goal of neuroscience as breaking the neural code -deciphering the relationships between spatiotemporal patterns of activity across groupsof neurons and the behavior of an animal or the mental state of a person.“Yuste and Bargmann, 2017Spontaneous neural activity, i.e. activity in the absence of stimulus, is arguably an understudiedfield of research in systems neuroscience which is dominated by reflexive experimental protocolsinvolving sensory stimuli and task paradigms. This thesis characterized several aspects of sponta-neous single neuron activity across multiple spatial and temporal scales. Visual cortex recordingsin cats and sensory cortex and thalamic recordings in mice using both extracellular electrophysi-ology and optical imaging provided extensive datasets which were analyzed using elementary (e.g.averaging) and more complex (e.g. state-space) analytical methods.Analyzing the spiking activity of dozens of single neurons recorded simultaneously requires thedevelopment and testing of spikesorting algorithms. In order to keep up with developing electrodetechnology, novel event (i.e. spike) detection and clustering algorithms must be developed andtested against datasets containing tens or hundreds or electrode channels and ground truth ofhundreds or thousands of neurons. The limited in vitro (i.e. cortical slice) datasets availablecontain only a few neurons (usually 1 neuron per recording) but provide the only available realtissue recordings that can be used to evaluate spikesorting algorithm performance (Fig 3.2-6).Simulated, i.e. in silico, datasets generated using the most advanced single neuron and networkmodels (currently available) can provide useful and realistic datasets. Sorting statistics from suchdatasets are very similar (i.e. statistically indistinguishable) from in vivo and in vitro data (Fig3.19). Sorting across simulated datasets reveals that while mature sorting suites can yield similarqualitative results, there are substantial differences (e.g. 50% or more neurons identified by oneoperator vs another) indicating operator skill may be more important than sorting suite algorithm.Sorting simulated datasets confirms that while higher density probes provide lower error ratessorting accuracy may saturate below a certain electrode density threshold (e.g. <20µm inter-channel spacing) with further increases in channel density not as useful for unit isolation. For thedata analyzed in this thesis, SpikeSorter (Swindale and Spacek, 2014; Swindale et al., 2017) providessorting results that are quantitatively as good as other sorting software suites. But many spikesorting challenges remain including sorting lower amplitude units, correcting for drift during bothacute and chronic recordings, improving sorting automation especially for hundreds of channels ofdata and developing better ground-truth datasets representing all brain areas and cortical states.Importantly, modeling of electrode layouts appears to suggest that narrow, single column electrodesmay yield similar sorting quality while being potentially less damaging to tissue. This may be173an important conclusion which should inform ongoing efforts at increasing electrode densities inextracellular probes.The physiological findings presented in this thesis relate largely to spontaneous neural activ-ity in mammalian cortex and thalamus and span multiple spatial scales: from the single neuronscale (e.g. 10-15µm somata) to the entire dorsal cortex of mouse (8-10mm across). Novel electro-physiological methods enabled the recording of activity of dozens of single neurons and correlatingwith mesoscale neuronal imaging methods that capture bihemispheric calcium activity (e.g. us-ing GCaMP6; Chapter 4) and membrane voltage (e.g. using VSDs; Chapters 4 and 5). Thesemethods enable the comparison - using correlation and averaging - of single neuron spikes (whichhave <1ms precision) to entire regions of cortex (recorded with 6.7ms - 33ms precision). The spike-triggered-mapping method used in calcium imaging supports a strong link across spatial scales evenduring spontaneous activity suggesting that spontaneously active neurons generally participate inmonosynaptically connected, functional networks. Subcortical neurons recorded from thalamic nu-clei exhibited additional variances in their relationship to dorsal cortex suggesting they are involvedin more complex, likely poly-synaptic functional networks. VSD imaging provided additional in-sight suggesting that the homogeneity present in the cortical neuron triggered GCaMP6 maps maynot be present on faster time scales and across membrane voltage profiles reported by VSDs. Thissupports future large-scale investigations of single-neuron VSD maps in visual (and other corticalareas) which may reveal spatio-temporal mesoscale activity patterns that span a comprehensiverepresentation-space of mono-synaptic connectivity between single neurons and dorsal cortex. Fu-ture spike-triggered mapping work should focus on a number of improvements in both experimentalacquisition and analysis including exploring cortical states (e.g. using synchrony index Saleem etal., 2010) and other cortical and subcortical areas. It would also be interestingly to use higher spa-tial resolution imaging such as two-photon microscopy to compute spike-triggered-maps which mayreveal specific ensembles correlating with the activity of distant neurons that could not otherwisebe captured (Yuste, 2015).Spontaneous single neuron activity also appears to have an inherent structure that may beimportant for understanding stimulus processing. Several previous studies found order duringintrinsic neuronal activity of high firing rate neurons in auditory and somatosensory areas of ratsoccuring during spontaneous activity (e.g. Luczak et al., 2007, 2015). Developing a novel methodto pursue this work further in visual cortex revealed a number of interesting findings including thatUP-state transitions appear to fall into distinct LFP-event-classes - LECs - which suggests multipletypes of UP-state transitions in cortex. Using this method, the latency of almost all (>90%)neurons can be tracked during UP-state transitions to provide a better and more complete analysisacross all sensory cortex. Multiple approaches further confirmed that over periods longer than 30minutes most neurons can change their relative UP-state latencies peaks as well as distributions.Importantly, these changes can also be observed in pair-wise synchronous firing order outside UP-states and even during desynchronized or awake states. These findings enable the investigation of174firing-order based neural coding in all areas of cortex and suggest a potentially much more complexscheme for temporal neural coding whereby a constantly changing order may represent complexinformation for downstream areas from primary sensory cortices.Current experimental limitations such as only being able to record from ≈100-300 neuronssimultaneously with millisecond precision will likely keep the neural coding debate active for manyyears to come. On the basis of association cortex recordings (e.g. monkey motor cortex) some haverecently argued that ”recording more neurons while repeating simple behaviours may not yieldricher datasets” (Gao and Ganguli, 2015) by showing that motor cortex firing-rate based neuronalactivity trajectories do not increase in complexity with ever increasing number of recorded neurons.Based on some of the studies reviewed and many findings made herein it is clear that this cannotbe the complete picture of neural coding - at least not in low sensory systems. It can be arguedthat until thousands (or hundreds of thousands of neurons) can be recored simultaneously withmillisecond precision the debate over neural coding will likely not be resolved. Until that time, it isbe important to be guided by data analyses that are generally agnostic regarding a more completetheory of cortical coding - especially in low sensory systems. Unraveling the story of mammaliancortex function seems to be much more complex with precise neuron inter-spiking relationshipslikely playing significant roles.175BibliographyAaron G, Yuste R (2006) Reverse optical probing (roping) of neocortical circuits.Synapse 60:437–440. → page 94Abdelfattah AS, Farhi SL, Zhao Y, Brinks D, Zou P, Ruangkittisakul A, Platisa J, Pieribone VA,Ballanyi K, Cohen AE, Campbell RE (2016) A bright and fast red fluorescent protein voltageindicator that reports neuronal activity in organotypic brain slices. Journal ofNeuroscience 36:2458–2472. → page 66, 97Abeles M (1982a) Local cortical circuits: An electrophysiological study. Springer-Verlag, Berlinand New York . → page 22, 136, 141, 158Abeles M (1982b) Local cortical circuits: An electrophysiological study. Springer, Berlin . →page 136, 169Abeles M (1991) Corticonics: Neural circuits of the cerebral cortex. Cambridge University Press,New York . → page 3, 109, 136Abeles M, Bergman H, Margalit E, Vaadia E (1993) Spatiotemporal firing patterns in the frontalcortex of behaving monkeys. J Neurophysiol 70:1629–1638. → page 136Abeles M, Gerstein GL (1988) Detecting spatiotemporal firing patterns among simultaneouslyrecorded single neurons. J Neurophysiol 60:909–924. → page 3, 136, 158Adrian ED (1926) The impulses produced by sensory nerve-endings: Part 1. J Physiol 61:49. →page 3, 109, 110, 168Adrian ED, Bronk DW (1928) The discharge of impulses in motor nerve fibres: Part I. Impulsesin single fibres of the phrenic nerve. J Physiol 66:81–101. → page 4Agmon A, Connors B (1989) Repetitive burst-firing neurons in the deep layers of mousesomatosensory cortex. Neuroscience Letters 99:137 – 141. → page 117Ahissar E, Arieli A (2001) Figuring space by time. Neuron 32:185–201. → page 3Amzica F, Steriade M (1995) Short- and long-range neuronal synchronization of the slow (¡ 1 hz)cortical oscillation. Journal of Neurophysiology 73:20–38. → page 117, 118, 136Amzica F, Steriade M (1997a) Cellular substrates and laminar profile of sleep k-complex.Neuroscience 82:671 – 686. → page 119Amzica F, Steriade M (1997b) The k-complex: its slow (¡1-hz) rhythmicity and relation to deltawaves. Neurology 49:952–9. → page 124Amzica F, Steriade M (1998a) Electrophysiological correlates of sleep delta waves.Electroencephalography and Clinical Neurophysiology 107:69 – 83. → page 4, 117, 119, 120, 128,166176Amzica F, Steriade M (1998b) Electrophysiological correlates of sleep delta waves.Electroencephalogr Clin Neurophysiol 107:69–83. → page 124Anastassiou C, Perin R, Buzsa´ki G, Markram H, Koch C (2015) Cell type- andactivity-dependent extracellular correlates of intracellular spiking. Journal ofNeurophysiology 114:608–623. → page 7, 14, 30, 31, 32, 39Anderson J, Lampl I, Reichova I, Carandini M, Ferster D (2000) Stimulus dependence of two-statefluctuations of membrane potential in cat visual cortex. Nat Neurosci 3:617–621. → page 118Arieli A, Shoham D, Hildesheim R, Grinvald A (1995) Coherent spatiotemporal patterns ofongoing activity revealed by real-time optical imaging coupled with single-unit recording in thecat visual cortex. Journal of Neurophysiology 73:2072–2093. → page 61, 94, 134Ascoli GA, Donohue DE, Halavi M (2007) Neuromorpho.org: A central resource for neuronalmorphologies. Journal of Neuroscience 27:9247–9251. → page 41Azouz R, Gray CM (2003) Adaptive coincidence detection and dynamic gain control in visualcortical neurons in vivo. Neuron 37:513–523. → page 3Bair W, Koch C (1996) Temporal precision of spike trains in extrastriate cortex of the behavingmacaque monkey. Neural Comput 8:1185–1202. → page 3Bak P, Tang C, Wiesenfeld K (1987) Self-organized criticality: An explanation of the 1/f noise.Phys. Rev. Lett. 59:381–384. → page 168Baker SN, Lemon RN (2003) Precise spatiotemporal repeating patterns in monkey primary andsupplementary motor areas occur at chance levels. Journal of Neurophysiology 84:1770–1780.→ page 136, 142Bar-Hillel A, Spiro A, Stark E (2006) Spike sorting: Bayesian clustering of non-stationary data. JNeurosci Meth 157:303–316. → page 28Barlow H (1961) Possible principles underlying the transformation of sensory messages. SensoryCommunication pp. 217–234. → page 2Bartho´ P, Hirase H, Monconduit L, Zugaro M, Harris KD, Buzsa´ki G (2004) Characterization ofneocortical principal cells and interneurons by network interactions and extracellular features.J Neurophysiol 92:600–608. → page 149Barth P, Freund TF, Acsdy L (2002) Selective gabaergic innervation of thalamic nuclei from zonaincerta. European Journal of Neuroscience 16:999–1014. → page 118Beggs JM, Plenz D (2003) Neuronal avalanches in neocortical circuits. JNeurosci 23:11167–11177. → page 168Beggs JM, Plenz D (2004) Neuronal avalanches are diverse and precise activity patterns that arestable for many hours in cortical slice cultures. J Neurosci 24:5216–5229. → page 168Beltramo R, D’Urso G, DalMaschio M, Farisello P, Bovetti S, Clovis Y, Lassi G, Tucci V,DePietriTonelli D, Fellin T (2013) Tbd. Nature Neuroscience 16:227–234. → page 117177Bere´nyi A, Somogyva´ri Z, Nagy AJ, Roux L, Long JD, Fujisawa S, Stark E, Leonardo A, HarrisTD, Buzsa´ki G (2014) Large-scale, high-density (up to 512 channels) recording of local circuitsin behaving animals. J Neurophysiol 111:1132–1149. → page 5Bermudez-Contreras E, Schjetnan A, Muhammad A, Bartho P, McNaughton B, Kolb B, GruberA, Luczak A (2013) Formation and reverberation of sequential neural activity patterns evokedby sensory stimulation are enhanced during cortical desynchronization. Neuron 79:555–566. →page 3, 4, 20, 21, 120, 121, 133, 136, 139, 140, 143, 152, 169, 170Bi Gq, Poo Mm (1998) Synaptic modifications in cultured hippocampal neurons: Dependence onspike timing, synaptic strength, and postsynaptic cell type. Journal ofNeuroscience 18:10464–10472. → page 3, 169Blanche TJ (2005) Large scale neuronal recording. Ph.D. diss., University of British Columbia,Vancouver, British Columbia, Canada. → page 149Blanche TJ, Spacek MA, Hetke JF, Swindale NV (2005a) Polytrodes: high-density siliconelectrode arrays for large-scale multiunit recording. J Neurophysiol 93:2987–3000. → page 5Blanche TJ, Spacek MA, Hetke JF, Swindale NV (2005b) Polytrodes: High-density siliconelectrode arrays for large-scale multiunit recording. Journal of Neurophysiology 93:2987–3000.→ page 10Blunck R, Starace DM, Correa AM, Bezanilla F (2004) Detecting rearrangements of shaker andnachbac in real-time with fluorescence spectroscopy in patch-clamped mammalian cells.Biophys J 86:3966–3980. → page 6Bohland JW, Wu C, Barbas H, Bokil H, Bota M, Breiter HC, Cline HT, Doyle JC, Freed PJ,Greenspan RJ, Haber SN, Hawrylycz M, Herrera DG, Hilgetag CC, Huang ZJ, Jones A, JonesEG, Karten HJ, Kleinfeld D, Ktter R, Lester HA, Lin JM, Mensh BD, Mikula S, Panksepp J,Price JL, Safdieh J, Saper CB, Schiff ND, Schmahmann JD, Stillman BW, Svoboda K,Swanson LW, Toga AW, Van Essen DC, Watson JD, Mitra PP (2009) A proposal for acoordinated effort for the determination of brainwide neuroanatomical connectivity in modelorganisms at a mesoscopic scale. PLOS Computational Biology 5:1–9. → page 60Brette R (2015) Philosophy of the spike: Rate-based vs. spike-based theories of the brain.Frontiers in Systems Neuroscience 9:151. → page 109, 110Broussard GJ, Liang R, Tian L (2014) Monitoring activity in neural circuits with geneticallyencoded indicators. Frontiers in Molecular Neuroscience 7:97. → page 5Brown T (1914) On the nature of the fundamental activity of the nervous centres; together withan analysis of the conditioning of rhythmic activity in progression, and a theory of theevolution of function in the nervous system. J Physiol 48:19–46. → page 2Bruno RM, Sakmann B (2006) Cortex is driven by weak but synchronously activethalamocortical synapses. Science 312:1622–1627. → page 66Burr DC, Ross J (1979) How does binocular delay give information about depth? VisionRes 19:523–532. → page 3178Busche M, Kekus M, Adelsberger H, Noda T, Forst H, Nelken I, Konnerth A (2015) Rescue oflong-range circuit dysfunction in alzheimer’s disease models. Nature Neuroscience 18:1623. →page 96Buzsa´ki G (2004) Large-scale recording of neuronal ensembles. Nat Neurosci 7:446–451. → page61Buzsa´ki G, Anastassiou CA, Koch C (2012) The origin of extracellular fields and currents —EEG, ECoG, LFP and spikes. Nat Rev Neurosci 13:407–420. → page 5, 28, 41, 72, 167Calabrese A, Paninski L (2011) Kalman filter mixture model for spike sorting of non-stationarydata. J Neurosci Meth 196:159–169. → page 28Carandini M, Shimaoka D, Rossi LF, Sato TK, Benucci A, Kno¨pfel T (2015) Imaging the awakevisual cortex with a genetically encoded voltage indicator. Journal of Neuroscience 35:53–63.→ page 60, 66Carrillo-Reid L, Miller JeK, Hamm JP, Jackson J, Yuste R (2015) Endogenous sequential corticalactivity evoked by visual stimuli. Journal of Neuroscience 35:8813–8828. → page 4, 144, 155Castro-Alamancos M (1997) Short-term plasticity in thalamocortical pathways: cellularmechanisms and functional role. Reviews in the neurosciences 8:95+. → page 77Chan A, Mohajerani M, LeDue J, Wang Y, Murphy T (2015) Mesoscale infraslow spontaneousmembrane potential fluctuations recapitulate high-frequency activity cortical motifs. NatureCommunications 6:7738. → page 12, 18, 60, 61, 68, 72, 96Chauvette S, Volgushev M, Timofeev I (2010) Origin of active states in local neocortical networksduring slow sleep oscillation. Cerebral Cortex 10:2660–2674. → page xiv, 117, 119, 121, 124,136, 166, 167Chemla S, Chavane F (2010) Voltage-sensitive dye imaging: Technique review and models.Journal of Physiology-Paris 104:40 – 50. → page 60Chen JL, Carta S, Soldado-Magraner J, Schneider BL, Helmchen F (2013a) Behaviour-dependentrecruitment of long-range projection neurons in somatosensory cortex. Nature 499:336–40. →page 60, 96Chen TW, Wardill TJ, Sun Y, Pulver SR, Renninger SL, Baohan A, Schreiter ER, Kerr RA,Orger MB, Jayaraman V, Looger LL, Svoboda K, Kim DS (2013b) Ultrasensitive fluorescentproteins for imaging neuronal activity. Nature 499:295–300. → page 6, 11, 12, 60, 61, 64, 66,96, 101Clark A (2013) Whatever next? predictive brains, situated agents, and the future of cognitivescience. Behavioral and Brain Sciences 36:181204. → page 2Clarke D, Sokoloff L (1999) Circulation and energy metabolism of the brain . → page 2Cohen LB, Salzberg BM, Davila HV, Ross WN, Landowne D, Waggoner AS, Wang CH (1974)Changes in axon fluorescence during activity: Molecular probes of membrane potential. TheJournal of Membrane Biology 19:1–36. → page 6, 102179Connors B, Gutnick M (1990) Intrinsic firing patterns of diverse neocortical neurons. Trends inNeurosciences 13:99 – 104. → page 22, 88Contreras D, Steriade M (1995) Cellular basis of eeg slow rhythms: a study of dynamiccorticothalamic relationships. Journal of Neuroscience 15:604–622. → page 118Cortelli, Gambetti, Montagna, Lugaresi (1999) Fatal familial insomnia: clinical features andmolecular genetics. Journal of Sleep Research 8:23–29. → page 119Cossart R, Aronov D, Yuste R (2003) Attractor dynamics of network up states in the neocortex.Nature 423:283–288. → page 4, 155Crunelli V, Hughes S (2010) The slow (¡1 hz) rhythm of non-rem sleep: a dialogue between threecardinal oscillators. Nature Neuroscience 13:9–17. → page 117, 125, 167Csicsvari J, Henze DA, Jamieson B, Harris KD, Sirota A, Bartho´ P, Wise KD, Buzsa´ki G (2003)Massively parallel recording of unit and local field potentials with silicon-based electrodes. JNeurophysiol 90:1314–1323. → page 5Csicsvari J, Hirase H, Czurko A, Buzsa´ki G (1998) Reliability and state dependence of pyramidalcell–interneuron synapses in the hippocampus: an ensemble approach in the behaving rat.Neuron 21:179–189. → page 149da Silva L (1991) Neural mechanisms underlying brain waves: from neural membranes tonetworks. Electroencephalogr Clin Neurophysiol. 79:81–93. → page 117David F, Schmiedt JT, Taylor HL, Orban G, Di Giovanni G, Uebele VN, Renger JJ, LambertRC, Leresche N, Crunelli V (2013) Essential thalamic contribution to slow waves of naturalsleep. J Neurosci 33:19599–19610. → page 118de Pasquale F, Della Penna S, Snyder AZ, Lewis C, Mantini D, Marzetti L, Belardinelli P,Ciancetta L, Pizzella V, Romani GL, Corbetta M (2010) Temporal dynamics of spontaneousmeg activity in brain networks. Proceedings of the National Academy ofSciences 107:6040–6045. → page 60DeNardo L, Berns D, DeLoach K, Luo L (2015) Connectivity of mouse somatosensory andprefrontal cortex examined with trans-synaptic tracing. Nature Neuroscience 18:1687–1697. →page 98Deneve S (2008) Bayesian spiking neurons i: Inference. Neural Computation 20:91–117. → page109Deschnes M, Takatoh J, Kurnikova A, Moore J, Demers M, Elbaz M, Furuta T, Wang F, KleinfeldD (2016) Inhibition, not excitation, drives rhythmic whisking. Neuron 90:374 – 387. → page 97Destexhe A, Contreras D, Steriade M (1999) Spatiotemporal analysis of local field potentials andunit discharges in cat cerebral cortex during natural wake and sleep states. JNeurosci 19:4595–4608. → page 117, 136Detari L, Rasmusson DD, Semba K (1997) Phasic relationship between the activity of basalforebrain neurons and cortical eeg in urethane-anesthetized rat. Brain Research 759:112 – 121.→ page 118180Douglas RJ, Koch C, Mahowald M, Martin KAC, Suarez HH (1995) Recurrent excitation inneocortical circuits. Science 269:981–985. → page 96Drake KL, Wise KD, Farraye J, Anderson DJ, BeMent SL (1988) Performance of planar multisitemicroprobes in recording extracellular single-unit intracortical activity. IEEE Trans BiomedEng 35:719–732. → page 4, 5Erisir A, VanHorn S, Sherman S (1997) Relative numbers of cortical and brainstem inputs to thelateral geniculatenucleus. Proc Natl Acad Sci 94:1517–1520. → page 118Erisken S, Vaiceliunaite A, Jurjut O, Fiorini M, Katzner S, Busse L (2014) Effects of locomotionextend throughout the mouse early visual system. Current Biology 24:2899–2907. → page 13Eschenko O, Magri C, Panzeri S, Sara SJ (2012) Noradrenergic neurons of the locus coeruleus arephase locked to cortical up-down states during sleep. Cerebral Cortex 22:426–435. → page 118Fee MS, Mitra PP, Kleinfeld D (1996) Automatic sorting of multiple unit neuronal signals in thepresence of anisotropic and non-gaussian variability. J Neurosci Meth 69:175–188. → page 28Ferezou I, Haiss F, Gentet LJ, Aronoff R, Weber B, Petersen CC (2007) Spatiotemporal dynamicsof cortical sensorimotor integration in behaving mice. Neuron 56:907 – 923. → page 60, 68Fox M, Raichle M (2007) Spontaneous fluctuations in brain activity observed with functionalmagnetic resonance imaging. Nature Reviews Neuroscience, 8:700–711. → page 1Fregnac Y, Zador M (2005) An intracellular view of time of cortical representation in v1 neuronsduring virtual oculomotor exploration of natural scenes. COSYNE . → page 3, 169Friston KJ (1997) Transients, metastability, and neuronal dynamics. NeuroImage 5:164 – 171. →page 171Fu Y, Tucciarone J, Espinosa J, Sheng N, Darcy D, Nicoll R, Huang Z, Stryker M (2014) Acortical circuit for gain control by behavioral state. Cell 156:1139 – 1152. → page 61Gabernet L, Jadhav SP, Feldman DE, Carandini M, Scanziani M (2005) Somatosensoryintegration controlled by dynamic thalamocortical feed-forward inhibition.Neuron 48:315 – 327. → page 61Galarreta M, Hestrin S (1998) Frequency-dependent synaptic depression and the balance ofexcitation and inhibition in the neocortex. Nature Neuroscience 1:587–594. → page 61Gao P, Ganguli S (2015) On simplicity and complexity in the brave new world of large-scaleneuroscience. Current Opinion in Neurobiology 32:148 – 155. → page 171, 175Garcia-Lazaro JA, Belliveau LA, Lesica NA (2013) Independent population coding of speech withsub-millisecond precision. J Neurosci 33:19362–19372. → page 3Gasthaus J, Wood F, Go¨ru¨r D, Teh YW (2009) Dependent Dirichlet process spike sorting In AdvNeural Inf Process Syst, Vol. 21, pp. 497–504. Curran Associates, Inc. → page 28Gerstein GL, Williams ER, Diesmann M, Grn S, Trengove C (2012) Detecting synfire chains inparallel spike data. Journal of Neuroscience Methods 206:54 – 64. → page 136, 142181Gil Z, Connors BW, Amitai Y (1997) Differential regulation of neocortical synapses byneuromodulators and activity. Neuron 19:679 – 686. → page 77Gold C, Henze DA, Koch C (2007) Using extracellular action potential recordings to constraincompartmental models. Journal of Computational Neuroscience 23:39–58. → page 3, 42, 43, 44Gold C, Henze DA, Koch C, Buzsa´ki G (2006) On the origin of the extracellular action potentialwaveform: A modeling study. Journal of Neurophysiology 95:3113–3128. → page 43, 44Gollisch T, Meister M (2008) Rapid neural coding in the retina with relative spike latencies.Science 319:1108–1111. → page 3Gong Y, Huang C, Li JZ, Grewe BF, Zhang Y, Eismann S, Schnitzer MJ (2015) High-speedrecording of neural spikes in awake mice and flies with a fluorescent voltage sensor.Science 350:1361–1366. → page 66Gratiy S, Arkhipov A, Feng D, Gouwens N, Cain N, Iyer R, Buice M, Mihalas S, Anastassiou C(2016) Performance driven neural network model representation: a relational approach. Societyfor Neuroscience, Annual Meeting, Poster 755.07 / NNN32. → page ivGratiy S, Billeh Y, Mitelut C, Dai K, Feng D, Gouwens D, Cain N, Anastassiou C, Arkhipov A(2017) Bionet: a python interface to neuron for large-scale network simulations. In progress .→ page iv, 5, 28, 44, 45, 202Gratiy S, Garner A, Cheng A, Durand S, Mizuseki K, Berg J, Sorensen S, Arkhipov A, Hines M,Shai A, Sunkin S, Phillips J, Zeng W, Reid R, Hawrylycz M, Koch C, Anastassiou C (2014)Simulating in vivo lfps and intracellular [ca2+] imaging data in mouse v1 using a detailedbiophysical model of the cortical column. Society for Neuroscience Annual Meeting p. 60.07. →page 45, 47Gratiy S, Mitelut C, Durand S, Denman D, Siegle J, Berg J, Sorensen S, Arkhipov A, Hines M,Shai A, Phillips J, Zeng H, Reid C, Hawrylycz M, Koch C, Anastassiou C (2015) Simulating lfpresponses in mouse v1 to sensory inputs using a large-scale, biophysically detailed multi-layercircuit model. Society for Neuroscience, Annual Meeting, Poster 598.21/K17. → page iv, 45, 50Gray CM, Ko¨nig P, Engel AK, Singer W (1989) Oscillatory responses in cat visual cortex exhibitinter-columnar synchronization which reflects global stimulus properties. Nature 338:334–337.→ page 3Gray CM, Maldonado PE, Wilson M, McNaughton B (1995) Tetrodes markedly improve thereliability and yield of multiple single-unit isolation from multi-unit recordings in cat striatecortex. J Neurosci Meth 63:43–54. → page 4, 28Green J (1958) A simple microelectrode for recording from the central nervous system.Nature 182:962. → page 4Grun S, Rotter S (2010) Analysis of parallel spike trains . → page 136Guggenmos DJ, Azin M, Barbay S, Mahnken JD, Dunham C, Mohseni P, Nudo RJ (2013)Restoration of function after brain damage using a neural prosthesis. Proceedings of theNational Academy of Sciences of the United States of America 110:21177–21182. → page 97182Guo Z, Li N, Huber D, Ophir E, Gutnisky D, Ting J, Feng G, Svoboda K (2014) Flow of corticalactivity underlying a tactile decision in mice. Neuron 81:179+. → page 61, 68, 96Hagen E, Ness T, Khosrowshahi A, Sorensen C, Fyhn M, Hafting T, Franke F, GT E (2015)Visapy: A python tool for biophysics-based generation of virtual spiking activity for evaluationof spike-sorting algorithms. Journal of Neuroscience Methods 245:182 – 204. → page 31, 206Hahn G, Ponce-Alvarez A, Monier C, Benvenuti G, Kumar A, Chavane F, Deco G, Frgnac Y(2017) Spontaneous cortical activity is transiently poised close to criticality. PLOSComputational Biology 13:1–29. → page 168Haider B, Duque A, Hasenstaub AR, McCormick DA (2006) Neocortical network activity in vivois generated through a dynamic balance of excitation and inhibition. J Neurosci 26:4535–4545.→ page 115Haider B, McCormick DA (2009) Rapid neocortical dynamics: Cellular and network mechanisms.Neuron 62:171–189. → page 119Hamel E, Grewe B, Parker J, Schnitzer M (2015) Cellular level brain imaging in behavingmammals: An engineering approach. Neuron 86:140 – 159. → page 60Harland D, Jackson R (2000) ’eight-legged cats’ and how they see - a review of recent research onjumping spiders (araneae: Salticidae). Cimbebasia 16:231–240. → page 1Harris KD, Csicsvari J, Hirase H, Dragoi G, Buzsa´ki G (2003) Organization of cell assemblies inthe hippocampus. Nature 424:552–6. → page 3Harris KD, Henze DA, Csicsvari J, Hirase H, Buzsa´ki G (2000) Accuracy of tetrode spikeseparation as determined by simultaneous intracellular and extracellular measurements. JNeurophysiol 84:401–414. → page 5, 28Harris KD, Henze DA, Hirase H, Leinekugel X, Dragoi G, Czurko´ A, Buzsa´ki G (2002) Spiketrain dynamics predicts theta-related phase precession in hippocampal pyramidal cells.Nature 417:738–741. → page 3Harris KD, Thiele A (2011) Cortical state and attention. Nat Rev Neurosci 12:509–523. → page115Harvey CD, Coen P, Tank DW (2012) Choice-specific sequences in parietal cortex during avirtual-navigation decision task. Nature 484:62–8. → page 60Hasenstaub A, Sachdev RNS, McCormick DA (2007) State changes rapidly modulate corticalneuronal responsiveness. J Neurosci 27:9607–9622. → page 117, 118Hatsopoulos N, Geman S, Amarasingham A, Bienenstock E (2003) At what time scale does thenervous system operate? Neurocomputing 52:25 – 29. → page 142Haupt D, Vanni M, Bolanos F, Mitelut C, LeDue J, Murphy T (2017) Mesoscale brain explorer, aflexible python-based image analysis and visualization tool. Neurophotonics 4:031210. → pageiv183Hausselt SE, Euler T, Detwiler PB, Denk W (2007) A dendrite-autonomous mechanism fordirection selectivity in retinal starburst amacrine cells. PLOS Biology pp. 1–20. → page 3Havenith MN, Yu S, Biederlack J, Chen NH, Singer W, Nikolic´ D (2011) Synchrony makesneurons fire in sequence, and stimulus properties determine who is ahead. JNeurosci 31:8570–8584. → page 3, 121, 138, 161, 162, 163, 165Hawrylycz M, Anastassiou C, Arkhipov A, Berg J, Buice M, Cain N, Gouwens NW, Gratiy S,Iyer R, Lee JH, Mihalas S, Mitelut C, Olsen S, Reid RC, Teeter C, de Vries S, Waters J, ZengH, Koch C, MindScope (2016) Inferring cortical function in the mouse visual system throughlarge-scale systems neuroscience. Proceedings of the National Academy ofSciences 113:7337–7344. → page iv, 5, 28, 41, 45Hay E, Hill S, Schrmann F, Markram H, Segev I (2011) Models of neocortical layer 5b pyramidalcells capturing a wide range of dendritic and perisomatic active properties. PLOSComputational Biology 7:1–18. → page 42, 45, 47, 202Hay E, Schu¨rmann F, Markram H, Segev I (2013) Preserving axosomatic spiking features despitediverse dendritic morphology. Journal of Neurophysiology 109:2972–2981. → page 42, 45, 47,202Hazan L, Zugaro M, Buzsa´ki G (2006) Klusters, NeuroScope, NDManager: a free software suitefor neurophysiological data processing and visualization. J Neurosci Meth 155:207–216. → page28Heil P (2004) First-spike latency of auditory neurons revisited. Current Opinion inNeurobiology 14:461 – 467. → page 3Hemmen J, Sejnowski T (2006) 23 problems in systems neuroscience. New York : OxfordUniversity Press, 2006. . → page 169Henze DA, Borhegyi Z, Csicsvari J, Mamiya A, Harris KD, Buzsa´ki G (2000) Intracellularfeatures predicted by extracellular recordings in the hippocampus in vivo. JNeurophysiol 84:390–400. → page 5, 24, 30Hines M (1986) Efficient computation of branched nerve equations. International Journal ofBio-Medical Computing 15:69–76. → page 5, 41Hines M, Carnevale N (1997) The neuron simulation environment. NeuralComputation 9:1179–1209. → page 5, 41Hodgkin AL, Huxley AF (1952) A quantitative description of membrane current and itsapplication to conduction and excitation in nerve. J Physiol 117:500–544. → page 41Hoel EP, Albantakis L, Tononi G (2013) Quantifying causal emergence shows that macro canbeat micro. Proceedings of the National Academy of Sciences 110:19790–19795. → page 144Holt GR, Koch C (1999) Electrical interactions via the extracellular potential near cell bodies.Journal of Computational Neuroscience 6:169–184. → page 43, 44, 202184Hooks BM, Mao T, Gutnisky DA, Yamawaki N, Svoboda K, Shepherd GMG (2013) Organizationof cortical and thalamic input to pyramidal neurons in mouse motor cortex. Journal ofNeuroscience 33:748–760. → page 61Hu H, Martina M, Jonas P (2009) Dendritic mechanisms underlying rapid synaptic activation offast-spiking hippocampal interneurons. Science 327:52–58. → page 45, 202Hubel DH (1957) Tungsten microelectrode for recording from single units. Science 125:549–550.→ page 4Hubel DH (1959) Single unit activity in striate cortex of unrestrained cats. JPhysiol 147:226–238. → page 3, 4, 168Hubel DH, Wiesel TN (1962) Receptive fields, binocular interaction and functional architecturein the cat’s visual cortex. J Physiol 160:106–154. → page 4Hulata E, Segev R, Ben-Jacob E (2002) A method for spike sorting and detection based onwavelet packets and Shannon’s mutual information. J Neurosci Meth 117:1–12. → page 28Hunnicutt B, Long B, Kusefoglu D, Gertz K, Zhong H, Mao T (2014) A comprehensivethalamocortical projection map at the mesoscopic level. Nature Neuroscience 17:1276. → page12, 68, 74, 98Iaccarino H, Singer A, Martorell A, Rudenko A, Gao F, Gillingham T, Mathys H, Seo J, KritskiyO, Abdurrob F, Adaikkan C, Canter R, Rueda R, Brown E, Boyden E, Tsai L (2016) Gammafrequency entrainment attenuates amyloid load and modifies microglia. Nature 540:230–235. →page 15Ikegaya Y, Aaron G, Cossart R, Aronov D, Lampl I, Ferster D, Yuste R (2004) Synfire chains andcortical songs: Temporal modules of cortical activity. Science 304:559–564. → page 3, 4, 136,140Izhikevich EM (2005) Polychronization: Computation with spikes. Neural Comput 18:245–282. →page 109, 136Jeffress L (1948) A place theory of sound localization. J Comp Physiol Psychol 41:35–39. → page3, 169Jia H, Rochefort NL, Chen X, Konnerth A (2010) Dendritic organization of sensory input tocortical neurons in vivo. Nature 464:1307–1312. → page 118Jimbo Y, Robinson H (2000) Propagation of spontaneous synchronized activity in cortical slicecultures recorded by planar electrode arrays. Bioelectrochemistry 51:107 – 115. → page 168Johansson RS, Birznieks I (2004) First spikes in ensembles of human tactile afferents codecomplex spatial fingertip events. Nat Neurosci 7:170–177. → page 3Jones E (1998) Viewpoint: the core and matrix of thalamic organization.Neuroscience 85:331 – 345. → page 118185Jun J, Mitelut C, Lai C, Gratiy S, Anastassiou C, Harris T (2017a) Real-time spike sortingplatform for high-density extracellular probes with ground-truth validation and drift correction.bioRxiv . → page iv, 7, 28, 31, 41, 45, 50, 51, 54, 57Jun J, Nicholas AS S J, Denman D, Bauza M, Barbarits B, Anastassiou C, Andrei A, Aydin C,Barbic M, Blanche T, Bonin V, Carandini M, Couto J, Dutta B, Gratiy S, Gutnisky D, HarrisK, Husser M, Karsh B, Koch C, Ledochowitsch P, Lee A, Mora C, Mitelut C, Musa S, O’KeefeJ, Okun M, Pachitariu M, Putzeys J, Rich P, Rossant C, Sun W, Svoboda K, Harris T (2017b)Neuropixels: Fully integrated silicon probes for high-density recording of neural activity. Nature(in review) . → page iv, 2, 5, 7, 28, 30, 41, 45, 46, 50, 170Kadir SN, Goodman DFM, Harris KD (2014) High-dimensional cluster analysis with the maskedem algorithm. Neural Computation 26:2379–2394. → page 27, 37, 48, 51Kahn I, Desai M, Knoblich U, Bernstein J, Henninger M, Graybiel AM, Boyden ES, Buckner RL,Moore CI (2011) Characterization of the functional mri response temporal linearity via opticalcontrol of neocortical pyramidal neurons. Journal of Neuroscience 31:15086–15091. → page 60Kappel D, Legenstein R, Habenschuss S, Hsieh M, Maass W (2017) Reward-based stochasticself-configuration of neural circuits. arXiv . → page 171Karni A, Tanne D, Rubenstein B, Askenasy J, Sagi D (1994) Dependence on rem sleep ofovernight improvement of a perceptual skill. Science 265:679–682. → page 119Katzner S, Nauhaus I, Benucci A, Bonin V, Ringach DL, Carandini M (2009) Local origin of fieldpotentials in visual cortex. Neuron 61:35–41. → page 5, 28Kayser C, Einha¨user W, Ko¨nig P (2003) Temporal correlations of orientations in natural scenes.Neurocomputing 52:117–123. → page 11Kelso J (1995) Dynamic patterns: The self-organization of brain and behavior . → page 171Kenet T, Bibitchkov D, Tsodyks M, Grinvald A, Arieli A (2003) Spontaneously emerging corticalrepresentations of visual attributes. Nature 425:954–956. → page 6, 58, 60Kim CK, Yang SJ, Pichamoorthy N, Young NP, Kauvar I, Jennings JH, Lerner TN, Berndt A,Lee SY, Ramakrishnan C, Davidson TJ, Inoue M, Bito H, Deisseroth K (2016) Simultaneousfast measurement of circuit dynamics at multiple sites across the mammalian brain. NatureMethods 13:325–328. → page 60, 97Kleinfeld D, Delaney KR, Fee MS, Flores JA, Tank DW, Gelperin A (1994) Dynamics ofpropagating waves in the olfactory network of a terrestrial mollusk: an electrical and opticalstudy. Journal of Neurophysiology 72:1402–1419. → page 60Kozai TDY, Marzullo TC, Hooi F, Langhals NB, Majewska AK, Brown EB, Kipke DR (2010)Reduction of neurovascular damage resulting from microelectrode insertion into the cerebralcortex using in vivo two-photon mapping. Journal of Neural Engineering 7:46011. → page 53Kreuz T, Chicharro D, Houghton C, Andrzejak RG, Mormann F (2013) Monitoring spike trainsynchrony. J Neurophysiol 109:1457–1472. → page 158186Kreuz T, Chicharro D, Andrzejak RG, Haas JS, Abarbanel HD (2009) Measuring multiple spiketrain synchrony. Journal of Neuroscience Methods 183:287 – 299. → page 158, 161Kreuz T, Chicharro D, Greschner M, Andrzejak RG (2011) Time-resolved and time-scale adaptivemeasures of spike train synchrony. Journal of Neuroscience Methods 195:92 – 106. → page 158Kreuz T, Mulansky M, Bozanic N (2015) Spiky: a graphical user interface for monitoring spiketrain synchrony. Journal of Neurophysiology 113:3432–3445. → page 159Kuffler SW (1946) The relation of electric potential changes to contracture in skeletal muscle.Journal of Neurophysiology 9:367–377. → page 4Kuffler SW (1953) Discharge patterns and functional organization of mammalian retina. Journalof Neurophysiology 16:37–68. → page 4Latham PE, Roth A, Hausser M, London M (2006) Requiem for the spike? Soc NeurosciAbstr 36:432.12. → page 3Le Bon-Jego M, Yuste R (2007) Persistently active, pacemaker-like neurons in neocortex.Frontiers in Neuroscience 1:123–129. → page 117Lein ES, Hawrylycz MJ, Ao N, Ayres M, Bensinger A, Bernard A, Boe AF, Boguski MS,Brockway KS, Byrnes EJ, Chen L, Chen L, Chen TM, Chi Chin M, Chong J, Crook BE,Czaplinska A, Dang CN, Datta S, Dee NR, Desaki AL, Desta T, Diep E, Dolbeare TA, DonelanMJ, Dong HW, Dougherty JG, Duncan BJ, Ebbert AJ, Eichele G, Estin LK, Faber C, FacerBA, Fields R, Fischer SR, Fliss TP, Frensley C, Gates SN, Glattfelder KJ, Halverson KR, HartMR, Hohmann JG, Howell MP, Jeung DP, Johnson RA, Karr PT, Kawal R, Kidney JM,Knapik RH, Kuan CL, Lake JH, Laramee AR, Larsen KD, Lau C, Lemon TA, Liang AJ, LiuY, Luong LT, Michaels J, Morgan JJ, Morgan RJ, Mortrud MT, Mosqueda NF, Ng LL, Ng R,Orta GJ, Overly CC, Pak TH, Parry SE, Pathak SD, Pearson OC, Puchalski RB, Riley ZL,Rockett HR, Rowland SA, Royall JJ, Ruiz MJ, Sarno NR, Schaffnit K, Shapovalova NV,Sivisay T, Slaughterbeck CR, Smith SC, Smith KA, Smith BI, Sodt AJ, Stewart NN, StumpfKR, Sunkin SM, Sutram M, Tam A, Teemer CD, Thaller C, Thompson CL, Varnam LR, ViselA, Whitlock RM, Wohnoutka PE, Wolkey CK, Wong VY, Wood M, Yaylaoglu MB, Young RC,Youngstrom BL, Feng Yuan X, Zhang B, Zwingman TA, Jones AR (2007) Genome-wide atlasof gene expression in the adult mouse brain. Nature 445:168–176. → page 14, 59Lemieux M, Chen JY, Lonjers P, Bazhenov M, Timofeev I (2014) The impact of corticaldeafferentation on the neocortical slow oscillation. J Neurosci 34:5689–5703. → page 118Lewicki M (1994) Bayesian modeling and classification of neural signals. NeuralComput 6:1005–1030. → page 28Lewicki M (1998) A review of methods for spike sorting: the detection and classification of neuralaction potentials. Network 9:R53–R78. → page 28Li CY, Poo MM, Dan Y (2009) Burst spiking of a single cortical neuron modifies global brainstate. Science 324:643–646. → page 19187Lim D, Mohajerani M, LeDue J, Boyd J, Chen S, Murphy T (2012) In vivo large-scale corticalmapping using channelrhodopsin-2 stimulation in transgenic mice reveals asymmetric andreciprocal relationships between cortical areas. Frontiers in Neural Circuits 6:11. → page 97Lin M, Schnitzer M (2016) Genetically encoded indicators of neuronal activity. NatureNeuroscience 19:1142–1153. → page 5Linden JF, Schreiner CE (2003) Columnar transformations in auditory cortex? a comparison tovisual and somatosensory cortices. Cerebral Cortex 13:83. → page 96Litke AM, Bezayiff N, Chichilnisky EJ, Cunningham W, Dabrowski W, Grillo AA, Grivich M,Grybos P, Hottowy P, Kachiguine S, Kalmar RS, Mathieson K, Petrusca D, Rahman M, SherA (2004) What does the eye tell the brain?: Development of a system for the large-scalerecording of retinal output activity. IEEE Trans Nucl Sci 51:1434–1440. → page 28Llinas R (2001) I of the vortex: From neurons to self . → page 2Logothetis NK, Eschenko O, Murayama Y, Augath M, Steudel T, Evrard HC, Besserve M,Oeltermann A (2012) Hippocampal-cortical interaction during periods of subcortical silence.Nature 491:547–53. → page 60, 95London M, Roth A, Beeren L, Ha¨usser M, Latham PE (2010) Sensitivity to perturbations in vivoimplies high noise and suggests rate coding in cortex. Nature 466:123–127. → page 3Lopez CM, Putzeys J, Raducanu BC, Ballini M, Wang S, Andrei A, Rochus V, Vandebriel R,Severi S, Hoof CV, Musa S, Helleputte NV, Yazicioglu RF, Mitra S (2017) A neural probe withup to 966 electrodes and up to 384 configurable channels in 0.13 µm soi cmos. IEEETransactions on Biomedical Circuits and Systems 11:510–522. → page 5Luczak A, Bartho P, Harris KD (2013) Gating of sensory input by spontaneous cortical activity.J Neurosci 33:1684–1695. → page 3, 4, 20, 110, 120, 121, 128, 132, 133, 134, 136, 139, 140, 143,152, 169, 170Luczak A, Bartho´ P, Marguet SL, Buzsa´ki G, Harris KD (2007) Sequential structure ofneocortical spontaneous activity in vivo. PNAS 104:347. → page 3, 4, 20, 21, 22, 110, 120, 121,122, 128, 132, 133, 134, 136, 139, 140, 141, 143, 152, 158, 169, 170, 172, 174Luczak A, McNaughton B, Harris K (2015) Packet-based communication in the cortex. NatureReviews Neuroscience 16:745–755. → page 3, 4, 109, 110, 120, 121, 122, 132, 143, 158, 169, 170,174Luczak A, Bartho P, Harris K (2009) Spontaneous events outline the realm of possible sensoryresponses in neocortical populations. Neuron 413-425:413–425. → page 3, 4, 20, 21, 110, 120,121, 122, 128, 132, 133, 134, 136, 139, 140, 143, 152, 169, 170Luczak A, Bartho P (2012) Consistent sequential activity across diverse forms of up states underketamine anesthesia. European Journal of Neuroscience 36:2830–2838. → page 20, 110, 167Ma Y, Shaik MA, Kim SH, Kozberg MG, Thibodeaux DN, Zhao HT, Yu H, Hillman EMC (2016)Wide-field optical mapping of neural activity and brain haemodynamics: considerations andnovel approaches. Philosophical Transactions of the Royal Society of London B: BiologicalSciences 371. → page 15, 59, 88, 90, 108, 210188Mackevicius EL, Best MD, Saal HP, Bensmaia SJ (2012) Millisecond precision spike timingshapes tactile perception. J Neurosci 32:15309–15317. → page 3MacLean JN, Watson BO, Aaron GB, Yuste R (2005) Internal dynamics determine the corticalresponse to thalamic stimulation. Neuron 48:811–823. → page 117Madisen L, Garner A, Shimaoka D, Chuong A, Klapoetke N, Li L, vanderBourg A, Niino Y, EgolfL, Monetti C, Gu H, Mills M, Cheng A, Tasic B, Nguyen T, Sunkin S, Benucci A, Nagy A,Miyawaki A, Helmchen F, Empson R, Knpfel T, Boyden E, Reid R, Carandini M, Zeng H(2015) Transgenic mice for intersectional targeting of neural sensors and effectors with highspecificity and performance. Neuron 85:942 – 958. → page 6, 12, 60, 61, 63, 96Madisen L, Zwingman TA, Sunkin SM, Oh SW, Zariwala HA, Gu H, Ng LL, Palmiter RD,Hawrylycz MJ, Jones AR, Lein ES, Zeng H (2010) A robust and high-throughput cre reportingand characterization system for the whole mouse brain. Nat Neurosci 13:133–140. → page 6Mainen ZF, Sejnowski TJ (1995) Reliability of spike timing in neocortical neurons.Science 268:1503–1506. → page 3, 169Mank M, Santos AF, Direnberger S, Mrsic-Flogel TD, Hofer SB, Stein V, Hendel T, Reiff DF,Levelt C, Borst A, Bonhoeffer T, Hu¨bener M, Griesbeck O (2008) A genetically encodedcalcium indicator for chronic in vivo two-photon imaging. Nat Methods 5:805–811. → page 5Manns ID, Alonso A, Jones BE (2000) Discharge properties of juxtacellularly labeled andimmunohistochemically identified cholinergic basal forebrain neurons recorded in associationwith the electroencephalogram in anesthetized rats. Journal of Neuroscience 20:1505–1518. →page 118Markram H, Sean H, Schurman F et al. (2015) Reconstruction and simulation of neocorticalmicrocircuitry. Cell 163:456–492. → page 5, 141Marr D, Poggio T (1976) From understanding computation to understanding neural circuitry.Artificial Intelligence Laboratory, MIT . → page 1Mazor O, Laurent G (2005) Transient dynamics versus fixed points in odor representations bylocust antennal lobe projection neurons. Neuron 48:661–673. → page 52McCormick D, Feeser H (1990) Functional implications of burst firing and single spike activity inlateral geniculate relay neurons. Neuroscience 39:103–113. → page 97McCormick D, Shu Y, Hasenstabu A (2004) Balanced recurrent excitation and inhibition in localcortical networks. Excitatory-Inhibitory Balance: Synapses, Circuits, Systems pp. 113–122. →page 115, 119McCormick D, Shu Y, Hasenstaub A, SanchezVives M, Badoual M, Bal T (2003) Persistentcortical activity: mechanisms of generation and effects on neuronal excitability. CerebCortex 13:1219–31. → page 119McCormick D, Yuste R (2006) Up states and cortical dynamics. Microcircuits: the Interfacebetween Neurons and Global Brain Function pp. 327–346. → page 115, 119189McCormick DA, McGinley MJ, Salkoff DB (2015) Brain state dependent activity in the cortexand thalamus. Current Opinion in Neurobiology 31:133 – 140. → page 61, 96McCulloch W, Pitts W (1943) A logical calculus of the ideas immanent in nervous activity. BullMath Biol. 52:99–115. → page 5, 41McDougal R, Morse T, Carnevale T, Marenco L R, Migliore M, Miller P, Shepherd G, Hines M(2017) Twenty years of modeldb and beyond: building essential modeling tools for the future ofneuroscience. Journal of Computational Neuroscience 42:1–10. → page 42Mease R, Kuner T, Fairhall AL, Groh A (2017) Multiplexed spike coding and adaptation in thethalamus. Soc Neurosci Abstr 19:1130–1140. → page 169Mejias JF, Kappen HJ, Torres JJ (2010) Irregular dynamics in up and down cortical states.PLOS ONE 5:1–13. → page 168Meliza CD, Dan Y (2006) Receptive-field modification in rat visual cortex induced by pairedvisual stimulation and single-cell spiking. Neuron 49:183–189. → page 3Mena-Segovia J, Sims HM, Magill PJ, Bolam JP (2008) Cholinergic brainstem neurons modulatecortical gamma activity during slow oscillations. J Physiol 586:2947–2960. → page 118Menda G, Shamble P, Nitzany E, Golden J, Hoy R (2014) Visual perception in the brain of ajumping spider. Current Biology 24:2580 – 2585. → page 1Migliore M, Cannia C, Lytton W, Markram H, Hines M (2006) Parallel network simulations withneuron. Journal of Computational Neuroscience 21:119. → page 5, 41Miller JeK, Ayzenshtat I, Carrillo-Reid L, Yuste R (2014) Visual stimuli recruit intrinsicallygenerated cortical ensembles. Proceedings of the National Academy ofSciences 111:E4053–E4061. → page 4, 144, 155Miller LM, Escab MA, Read HL, Schreiner CE (2001) Functional convergence of responseproperties in the auditory thalamocortical system. Neuron 32:151 – 160. → page 96Millman D, Mihalas S, Kirkwood A, Niebur E (2010) Self-organized criticality occurs innon-conservative neuronal networks during /‘up/’ states. Nat Phys 6:801–805. → page 168Mitelut C, Gratiy S, Denman D, Siegle J, Durand S, Godfrey K, Lee C, Reid C, Hawrylycz M,Koch C, , Swindale N, Anastassiou C (2015) Standardizing spike sorting: an in vitro, in silicoand in vivo study to develop quantitative metrics for sorting extracellularly recorded spikingactivity. Society for Neuroscience Annual Meeting D.04.d. → page iv, 7, 28, 40, 41, 43, 44, 45,49, 50, 51, 52, 53Mitelut C, Gratiy S, Perin R, Durand S, Mizuseki K, Godfrey K, Lee C, Blanche T, Swindale N,Reid C, Hawrylycza M, Markram H, Koch C, Anastassiou C (2014) On spike detection and thedevelopment of quantitative meaures for spike clustering using ”ground truth” data: Acomputational and slice electrophysiology study. Society for Neuroscience Annual Meeting 60.→ page iv, 7, 28, 32, 35, 38, 41, 45, 47, 48190Mitelut C, Spacek M, Allen C, Murphy T, Swindale N (2015) Local field potential events insynchronized state in cat v1 cluster and correlate with single-unit activity. Brain States:Characterization and Neuromodulation by Deep Brain Stimulation . → page ivMitelut C, Luo A, Silasi G, Sekino Y, Boyd J, Bolanos F, Swindale N, Murphy T (2017a)Movement initiation in gcamp6 mice is preceded by stereotyped, multi-second dorsal cortexdynamics. Neurofutures . → page ivMitelut C, Luo A, Silasi G, Sekino Y, Boyd J, Bolanos F, Swindale N, Murphy T (2017b)Movement initiation in gcamp6 mice is preceded by stereotyped, multi-second dorsal cortexdynamics. Psychiatry Research Day, University of British Columbia . → page ivMitelut C, Xiao D, Vanni M, Chan A, LeDue J, Xie Y, Chen A, Swindale N, Murphy T (2016)Mapping cortical mesoscopic networks of single spiking cortical or sub-cortical neurons.Neurofutures . → page iv, 59, 98, 101Mizuseki KS A, Pastalkova E, Buzsa´ki G (2009) Theta oscillations provide temporal windows forlocal circuit computation in the entorhinal-hippocampal loop. Neuron 64:267–280. → page 149Mohajerani M, Chan A, Mohsenvand M, LeDue L, Liu R, McVea D, Boyd J, Wang Y, Murphy T,Reimers M (2013) Spontaneous cortical activity alternates between motifs defined by regionalaxonal projections. Nature Neuroscience 16:10. → page 11, 12, 14, 16, 18, 23, 60, 61, 64, 66, 68,72, 74, 76, 96, 97Mohajerani M, McVea D, Fingas M, Murphy T (2010) Mirrored bilateral slow-wave corticalactivity within local circuits revealed by fast bihemispheric voltage-sensitive dye imaging inanesthetized and awake mice. Journal of Neuroscience 30:3745–3751. → page 16, 61Moore JD, Deschenes M, Furuta T, Huber D, Smear MC, Demers M, Kleinfeld D (2013)Hierarchy of orofacial rhythms revealed through whisking and breathing. Nature 497:205–10. →page 97Moruzzi G (1965) The Functional Significance of Sleep with Particular Regard to the BrainMechanisms Underlying Consciousness, pp. 345–388. → page 119Murphy T, Boyd J, Bolanos F, Vanni M, Silasi G, Haupt D, LeDue J (2016) High-throughputautomated home-cage mesoscopic functional imaging of mouse cortex. NatureCommunications 7:11611. → page 88, 90Nakai J, Ohkura M, Imoto K (2001) A high signal-to-noise ca2+ probe composed of a singlegreen fluorescent protein. Nature Biotechnology 19.2:137–141. → page 6Nauhaus I, Busse L, Carandini M, Ringach D (2009) Stimulus contrast modulates functionalconnectivity in visual cortex. Nature Neuroscience 12:70+. → page 94Neske GT (2016) The slow oscillation in cortical and thalamic networks: Mechanisms andfunctions. Frontiers in Neural Circuits 9:88. → page 110, 115, 120, 166Neto JP, Lopes G, Fraza˜o J, Nogueira J, Lacerda P, Baia˜o P, Aarts A, Andrei A, Musa S,Fortunato E, Barquinha P, Kampff AR (2016) Validating silicon polytrodes with pairedjuxtacellular recordings: method and dataset. Journal of Neurophysiology 116:892–903. → page29, 30, 57, 206191Nguyen DP, Frank LM, Brown EN (2003) An application of reversible-jump markov chain montecarlo to spike classification of multi-unit extracellular recordings. Network 14:61–82. → page 28Niell CM, Stryker MP (2008) Highly selective receptive fields in mouse visual cortex. JNeurosci 28:7520–7536. → page 149Nikolic´ D (2007) Non-parametric detection of temporal order across pairwise measurements oftime delays. Journal of Computational Neuroscience 22:5–19. → page 161Norenberg A, Hu H, Vida I, Bartos M, Jonas P (2010) Distinct nonuniform cable propertiesoptimize rapid and efficient activation of fast-spiking gabaergic interneurons. Proceedings of theNational Academy of Sciences 107:894–899. → page 45, 202Nowak LG, Sanchez-Vives MV, McCormick DA (2005) Role of synaptic and intrinsic membraneproperties in short-term receptive field dynamics in cat area 17. Journal ofNeuroscience 25:1866–1880. → page 88O’Connor D, Hires S, Guo Z, Li N, Yu J, Sun Q, Huber D, Svoboda K (2013) Neural codingduring active somatosensation revealed using illusory touch. Nature Neuroscience 16:958+. →page 97Oh SW, Harris JA, Ng L, Winslow B, Cain N, Mihalas S, Wang Q, Lau C, Kuan L, Henry AM,Mortrud MT, Ouellette B, Nguyen TN, Sorensen SA, Slaughterbeck CR, Wakeman W, Li Y,Feng D, Ho A, Nicholas E, Hirokawa KE, Bohn P, Joines KM, Peng H, Hawrylycz MJ, PhillipsJW, Hohmann JG, Wohnoutka P, Gerfen CR, Koch C, Bernard A, Dang C, Jones AR, Zeng H(2014) A mesoscale connectome of the mouse brain. Nature 508:207–14. → page 23, 59, 68, 74,76, 97, 98Oizumi M, Albantakis L, Tononi G (2014) From the phenomenology to the mechanisms ofconsciousness: Integrated information theory 3.0. PLOS Computational Biology 10. → page 144Okun M, Steinmetz NA, Cossell L, Iacaruso MF, Ko H, Barth P, Moore T, Hofer SB,Mrsic-Flogel T, Carandini M, Harris KD (2015) Diverse coupling of neurons to populations insensory cortex. Nature 521:511–515L. → page 96Oram MW, Wiener MC, Lestienne R, Richmond BJ (1999) Stochastic nature of precisely timedspike patterns in visual system neuronal responses. Journal of Neurophysiology 81:3021–3033.→ page 3, 142Pachitariu M, Steinmetz N, Kadir S, Carandini M, Harris KD (2016) Kilosort: realtimespike-sorting for extracellular electrophysiology with hundreds of channels. bioRxiv . → page27, 28, 51Pachitariu M, Stringer C, Dipoppa M, Schro¨der S, Rossi LF, Dalgleish H, Carandini M, HarrisKD (2017) Suite2p: beyond 10,000 neurons with standard two-photon microscopy. bioRxiv . →page 2Pape HC, McCormick D (1995) Electrophysiological and pharmacological properties ofinterneurons in the cat dorsal lateral geniculate nucleus. Neuroscience 68:1105 – 1125. → page22192Paz J, Davidson T, Frechette E, Delord B, Parada I, Peng K, Huguenard J (2013) Closed-loopoptogenetic control of thalamus as a tool for interrupting seizures after cortical injury. NatureNeuroscience 16:64+. → page 96, 97Perin R, Markram H (2013) A computer-assisted multi-electrode patch-clamp system. J VisExp p. 50630. → page 31Peterka D, Takahashi H, Yuste R (2010) Imaging voltage in neurons. Neuron 69:9–21. → page 5,6, 102Petermann T, Thiagarajan TC, Lebedev MA, Nicolelis MAL, Chialvo DR, Plenz D (2009)Spontaneous cortical activity in awake monkeys composed of neuronal avalanches. Proceedingsof the National Academy of Sciences 106:15921–15926. → page 168Petersen CCH, Hahn TTG, Mehta M, Grinvald A, Sakmann B (2003) Interaction of sensoryresponses with spontaneous depolarization in layer 2/3 barrel cortex. PNAS 100:13638–13643.→ page 118Petrus E, Isaiah A, Jones A, Li D, Wang H, Lee H, Kanold P (2014) Crossmodal induction ofthalamocortical potentiation leads to enhanced information processing in the auditory cortex.Neuron 81:664–673. → page 61Plonsey R (1974) The active fiber in a volume conductor. IEEE Transactions on BiomedicalEngineering BME-21:371–381. → page 43Pnevmatikakis E, Soudry D, Gao Y, Machado T, Merel J, Pfau D, Reardon T, Mu Y, LacefieldC, Yang W, Ahrens M, Bruno R, Jessell T, Peterka D, Yuste R, Paninski L (2016)Simultaneous denoising, deconvolution, and demixing of calcium imaging data.Neuron 89:285–299. → page 23, 64, 66Poskanzer KE, Yuste R (2016) Astrocytes regulate cortical state switching in vivo. Proceedings ofthe National Academy of Sciences 113:E2675–E2684. → page 117Pouzat C, Delescluse M, Viot P, Diebolt J (2004) Improved spike-sorting by modeling firingstatistics and burst-dependent spike amplitude attenuation: a markov chain monte carloapproach. J Neurophysiol 91:2910–2928. → page 28Prete FR (2004) Complex worlds from simpler nervous systems. MIT Press . → page 1Pritchard WS (1992) The brain in fractal time: 1/f-like power spectrum scaling of the humanelectroencephalogram. International Journal of Neuroscience 66:119–129. → page 168Quaglio P, Yegenoglu A, Torre E, Endres D, Grun S (2017) Detection and evaluation ofspatio-temporal spike patterns in massively parallel spike train data with spade. Frontiers inComputational Neuroscience 11. → page 136, 158Quian Quiroga R, Nadasdy Z, Ben-Shaul Y (2004) Unsupervised spike detection and sorting withwavelets and superparamagnetic clustering. Neural Comput 16:1661–1687. → page 18, 28Quiroga Q, Kreuz T, Grassberger P (2002) Event synchronization: A simple and fast method tomeasure synchronicity and time delay patterns. Phys. Rev. E 66:041904. → page 158193Raichle M (2010) Two views of brain function. Trends in cognitive sciences 14:180–190. → page1, 3Rall W (1964) Theoretical significance of dendritic trees for neuronal input-output relations, in r.f. reiss (ed.), neural theory and modeling p. 7397. → page 5, 41Ramirez-Villegas JF, Logothetis NK, Besserve M (2015) Diversity of sharp-waveripple lfpsignatures reveals differentiated brain-wide dynamical events. Proceedings of the NationalAcademy of Sciences 112:E6379–E6387. → page 119, 167Rao R, Ballard D (1999) Predictive coding in the visual cortex: a functional interpretation ofsome extra-classical receptive-field effects. Nature Neuroscience 2:79–87. → page 2Reichinnek S, Ku¨nsting T, Draguhn A, Both M (2010) Field potential signature of distinctmulticellular activity patterns in the mouse hippocampus. Journal ofNeuroscience 30:15441–15449. → page 119, 167Reid RC, Victor JD, Shapley RM (1997) The use of m-sequences in the analysis of visualneurons: Linear receptive field properties. Visual Neurosci 14:1015–1027. → page 11Renshaw B, Forbes A, Morison BR (1940) Activity of isocortex and hippocampus: electricalstudies with microelectrodes. J Neurophysiol 3:74–105. → page 4Rickgauer J, Deisseroth K, Tank D (2014) Simultaneous cellular-resolution optical perturbationand imaging of place cell firing fields. Nature Neuroscience 17:1816–1824. → page 97Rigas P, Castro-Alamancos MA (2007) Thalamocortical up states: Differential effects of intrinsicand extrinsic cortical inputs on persistent activity. Journal of Neuroscience 27:4261–4272. →page 117, 118Rodny JJ, Shea TM, Kello CT (2017) Transient localist representations in critical branchingnetworks. Language, Cognition and Neuroscience 32:330–341. → page 171Rolls ET, Treves A (2011) The neuronal encoding of information in the brain. Progress inNeurobiology 95:448 – 490. → page 109, 110Ros H, Sachdev RNS, Yu Y, Sˇestan N, McCormick DA (2009) Neocortical networks entrainneuronal circuits in cerebellar cortex. Journal of Neuroscience 29:10309–10320. → page 166Russo E, Durstewitz D (2017) Cell assemblies at multiple time scales with arbitrary lagconstellations. eLife 6:e19428. → page 136, 158Sakata S, Harris KD (2009) Laminar structure of spontaneous and sensory-evoked populationactivity in auditory cortex. Neuron 64:404–418. → page 117, 149Sakmann B, Neher E (1984) Patch clamp techniques for studying ionic channels in excitablemembranes. Annual Review of Physiology 46:455–472. → page 29, 115Salami M, Itami C, Tsumoto T, Kimura F (2003) Change of conduction velocity by regionalmyelination yields constant latency irrespective of distance between thalamus and cortex.PNAS 100:6174–6179. → page 3194Saleem AB, Chadderton P, Apergis-Schoute J, Harris KD, Schultz SR (2010) Methods forpredicting cortical UP and DOWN states from the phase of deep layer local field potentials. JComput Neurosci 29:49–62. → page 19, 107, 119, 167, 174Sanchez-Vives M, Massimini M, Mattia M (2017) Shaping the default activity pattern of thecortical network. Neuron 94:993–1001. → page 110, 115, 120Sanchez-Vives M, McCormick D (2000) Cellular and network mechanisms of rhythmic recurrentactivity in neocortex. Nature Neuroscience 3:1027–1034. → page 115, 117, 118Scarpetta S, de Candia A (2014) Alternation of up and down states at a dynamicalphase-transition of a neural network with spatiotemporal attractors. Frontiers in SystemsNeuroscience 8:88. → page 168Schneider G, Nikolic D (2006) Detection and assessment of near-zero delays in neuronal spikingactivity. Journal of Neuroscience Methods 152:97 – 106. → page 138, 161, 162Schrader S, Gru¨n S, Diesmann M, Gerstein GL (2008) Detecting synfire chain activity usingmassively parallel spike train recording. J Neurophysiol 100:2165–2176. → page 136Schuetze SM (1983) The discovery of the action potential. Trends in Neurosciences 6:164–168. →page 4Schweimer JV, Mallet N, Sharp T, Ungless MA (2011) Spike-timing relationship ofneurochemically-identified dorsal raphe neurons during cortical slow oscillations.Neuroscience 196:115–123. → page 118Serruya M, Hatsopoulos N, Paninski L, Fellows M, Donoghue J (2020) Brain-machine interface:Instant neural control of a movement signal. Soc Neurosci Abstr 416:141–142. → page 97Shadlen MN, Newsome WT (1994) Noise, neural codes and cortical organization. Curr OpinNeurobiol 4:569–579. → page 3, 110, 168Shadlen MN, Newsome WT (1998) The variable discharge of cortical neurons: Implications forconnectivity, computation, and information coding. J Neurosci 18:3870–3896. → page 3Shapley R, Reid RC, Soodak R (1991) Spatiotemporal receptive fields and direction selectivity InLandy MS, Movshon JA, editors, Computational Models of Visual Processing, pp. 109–118.MIT Press, Cambridge, MA. → page 11Sherman S, Guillery R (1996) Functional organization of thalamocortical relays. Journal ofNeurophysiology 76:1367–1395. → page 97Sherman S, Guillery R (2006) Exploring the thalamus and its role in cortical function, 2ndedition . → page 23, 77Sheroziya M, Timofeev I (2014) Global intracellular slow-wave dynamics of the thalamocorticalsystem. Journal of Neuroscience 34:8875–8893. → page 61, 118Sherrington C (1906) The integrative action of the nervous system. Yale University Press . →page 2195Shoham D, Glaser DE, Arieli A, Kenet T, Wijnbergen C, Toledo Y, Hildesheim R, Grinvald A(1999) Imaging cortical dynamics at high spatial and temporal resolution with novel bluevoltage-sensitive dyes. Neuron 24:791 – 802. → page 16, 66Siapas AG, Wilson MA (1998) Coordinated interactions between hippocampal ripples andcortical spindles