Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Evolutionary dynamics of ovarian cancer microenvironments and tumour cells Zhang, Allen Wenyu 2019

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

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
24-ubc_2019_september_zhang_allen.pdf [ 43.09MB ]
Metadata
JSON: 24-1.0379237.json
JSON-LD: 24-1.0379237-ld.json
RDF/XML (Pretty): 24-1.0379237-rdf.xml
RDF/JSON: 24-1.0379237-rdf.json
Turtle: 24-1.0379237-turtle.txt
N-Triples: 24-1.0379237-rdf-ntriples.txt
Original Record: 24-1.0379237-source.json
Full Text
24-1.0379237-fulltext.txt
Citation
24-1.0379237.ris

Full Text

Evolutionary dynamics of ovariancancer microenvironments and tumourcellsbyAllen Wenyu ZhangA THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate and Postdoctoral Studies(Bioinformatics)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)May 2019c© Allen Wenyu Zhang 2019The following individuals certify that they have read, and recommend to the Faculty of Graduateand Postdoctoral Studies for acceptance, the dissertation entitled:Evolutionary dynamics of ovarian cancer microenvironments and tumour cellssubmitted by Allen Wenyu Zhang in partial fulfillment for the requirements forthe degree of Doctor of Philosophyin BioinformaticsExamining Committee:Sohrab Shah, Pathology and Laboratory MedicineCo-supervisorWyeth Wasserman, Medical GeneticsCo-supervisorBrad Nelson, Medical GeneticsSupervisory Committee MemberWan Lam, Pathology and Laboratory MedicineUniversity ExaminerGabriela Cohen Freue, StatisticsUniversity ExaminerAdditional Supervisory Committee Members:Martin Hirst, Microbiology and ImmunologySupervisory Committee MemberRaymond Ng, Computer ScienceSupervisory Committee MemberDaniel Renouf, MedicineSupervisory Committee MemberiiAbstractHigh-grade serous ovarian cancer (HGSC) is the most common and lethal histotype ofepithelial ovarian cancer. Often presenting as multi-site disease, HGSC exhibits extensivemalignant clonal diversity with widespread but non-random patterns of disease dissemination.The proclivity of HGSC toward clonally heterogeneous disease is thought to underlie theprevalence of treatment-resistant disease. Yet, the factors that influence the spatial distributionof cancer clones in HGSC remain largely uncharacterized. Hypothesizing that distinct peritonealniches formed by microenvironmental cell types shape the observed patterns of clonal dynamicsin HGSC, the primary aim of this thesis was to understand how microenvironmental factorsinfluence malignant cell evolutionary dynamics.To establish the experimental substrate for this thesis, I led the construction of a cohortof 148 tumour samples from 41 HGSC cases (Chapter 2). In addition to coordinating clinicalcase identification, I oversaw and learned how to create patient-derived xenograft models andconduct single cell experiments from patient tumours. Leveraging this resource, I exploredwhether local immune microenvironment factors shape tumor progression properties at theinterface of tumor-infiltrating lymphocytes and cancer cells (Chapter 3). Through multi-regionstudy with whole-genome sequencing, immunohistochemistry, image analysis, gene expressionprofiling, and T- and B-cell receptor sequencing, I identified three immunologic subtypes acrosssamples associated with patterns of malignant clonal diversity. These findings were consistentwith immunological pruning of tumor clones. Finally, in order to explore the non-lymphocyticcomponents of the tumour microenvironment, I developed an automated approach to celltype identification from single cell RNA-seq data that eliminates the manual work involved intraditional workflows reliant on post-hoc expert annotation (Chapter 4). I demonstrated howthis method performs superiorly to state-of-the-art workflows for cell type identification andapplied the method to profile the HGSC microenvironment.Collectively, this work highlights multiple interfaces of evolutionary interplay between malig-nant and non-malignant cells in the HGSC microenvironment, identifying novel mechanismsby which tumour cells escape from immune recognition. These results will inform the inter-pretation of results from immunotherapy clinical trials and set the stage for comprehensivemicroenvironment profiling in large HGSC cohorts and other cancers.iiiLay SummaryOvarian cancer is one of the leading causes of death from cancer in the developed world. Over50% of patients with the most common type of ovarian cancer, high-grade serous ovarian cancer,die within 5 years of diagnosis. While most patients get better initially with treatment, the diseaseeventually becomes resistant. These cancers often contain multiple distinct subpopulations ofcancer cells. Treatment that works on one cancer cell population may not on others, allowingthe cancer to survive. Tumours also contain non-cancerous cell types, including cells from theimmune system. Some of these non-cancerous cell types are linked to how long patients survive.The goal of this work is to understand how these cell types affect cancer cell growth. In doingso, we may be able to change the way non-cancerous cells interact with cancer cells to treatovarian cancer.ivPrefaceUnder the guidance of my co-supervisors, Dr. Sohrab P. Shah and Dr. Wyeth W. Wasserman,I was involved in the conception and design of the work presented in this thesis. I was responsiblefor the experimental research, data analysis, interpretation, and presentation of the work. Inaddition, I learned how to perform the wet lab experiments for some components of Chapter 2,including sample processing and patient-derived xenograft (PDX) construction. This workwould not have been possible without the generous help of many close clinical and researchcollaborators, acknowledged below.Chapter 2 is unpublished work describing the accrual of the largest currently publishedcohort of multi-site high-grade serous ovarian cancer (148 tumour samples from 41 patients). Iled this project under the oversight of gynecologic oncology surgeon Dr. Jessica McAlpine andDr. Sohrab Shah, coordinating a multidisciplinary team of clinical and research staff includingsurgeons, fellows, clinical research assistants, animal research staff, and technologists to identify,recruit, and collect from probable cases of multi-site HGSC. Dr. Jessica McAlpine and variousclinical fellows were responsible for earmarking HGSC cases. I forwarded cases and directedsample collection for multiple experiments, including PDX construction, single cell processing,and bulk sequencing. Additionally, I worked with Jamie Lim, Dr. Ciara O’Flanagan, and Dr.Peter Eirew from BC Cancer to design the experimental protocol for single cell RNA-sequencing.Gayle Jagpal and Chriselle Mariz Serna from Vancouver General Hospital (VGH) prioritizedresearch cases on surgical slates. Various surgeons at VGH and University of British Columbia(UBC) Hospital performed the surgeries. Khaye Rufin and Stephanie Lam from BC Cancer wereresponsible for obtaining consent and coordinating peripheral blood collection. Dr. Lien Hoangfrom VGH was responsible for pathologic evaluation. Nancy Ferguson and Martin Avancena fromVGH were responsible for clinical processing of samples from the operating room. Jamie Limand Clara Salamanca from BC Cancer additionally processed samples for research purposes andsequencing. Jamie Lim and Dr. Ciara O’Flanagan performed single cell RNA-seq experiments.Teresa Ruiz de Algara and So Ra Lee from BC Cancer transplanted patient tumours into PDXmouse models, and taught me how to do so. I performed the bioinformatics data analysis inthis chapter.Chapter 3 is a modified version of material published in “Zhang, AW et al. Interfaces ofMalignant and Immunologic Clonal Dynamics in Ovarian Cancer. Cell (2018).” [1] This projectvwas led by myself and co-supervised by Dr. Sohrab Shah and Dr. Brad Nelson, Director ofthe Deeley Research Centre at BC Cancer. This research leverages the experimental resourcecreated in Chapter 3. I was primarily responsible for the bioinformatics analysis, includingthe identification of immune infiltration patterns, the associative analysis between malignantclonal diversity and these patterns, processing and analysis of T-cell receptor and B-cell receptorsequencing data, neoantigen calling and integrative analysis, HLA loss-of-heterozygosity analysis,and integrative analysis of mutational signatures and immune patterns. Additionally, I formulatedthe hierarchical Bayesian probabilistic model for inferring subclonal HLA loss-of-heterozygosity,and helped formulate the clonal inference model with Dr. Andrew McPherson. Dr. PhineasHamilton performed cell type identification from H&E (hematoxylin & eosin) images, and Iperformed the hotspot identification and integrative analysis with immune patterns. Dr. KatyMilne, Sonya Laan, Stacey LeDoux, and Dr. David Kroeger from the Deeley Research Centreperformed the immunohistochemistry experiments. I was responsible for generating all of thefigures and tables associated with the paper. Together, Dr. Sohrab Shah, Dr. Brad Nelson,Dr. Robert Holt, and I were responsible for conceiving the project, designing the experiments,interpreting the results, and writing the manuscript. For the full list of contributors and theirimportant contributions, please refer to [1].Chapter 4 is a modified version of material that is under peer review and was preprinted inBioRxiv [2]: “Zhang, AW et al. Probabilistic cell type assignment of single-cell transcriptomicdata reveals spatiotemporal microenvironment dynamics in human cancers”. I led this projectunder the co-supervision of Dr. Kieran Campbell and Dr. Sohrab Shah. Together with Dr.Kieran Campbell and Dr. Sohrab Shah, I helped formulate the model and interpret the results.I performed the majority of the bioinformatics data analysis in the paper, including the modelimplementation, simulation analysis, Bayesian model fitting with pymc3, analysis of externaldatasets, and analysis of high-grade serous ovarian cancer and follicular lymphoma single cellRNA-seq data. I led the accrual of the high-grade serous ovarian cancer cohort for this chapter,as described in Chapter 2. Jamie Lim and Dr. Ciara O’Flanagan processed ovarian samplesfor single cell RNA-sequencing. Dr. Sohrab Shah, Dr. Christian Steidl (BC Cancer), Dr.Anja Mottok (BC Cancer), and Dr. Clementine Sarkozy (BC Cancer) identified the follicularlymphoma cases. Elizabeth Chavez (BC Cancer) performed the experimental work for singlecell RNA-seq of the follicular lymphoma and reactive lymph node samples. Matt Wiens, PascaleWalters, and Tim Chan, co-operative education students at BC Cancer I helped supervise,helped build infrastructure and perform other components of analysis related to the paper.I wrote the manuscript along with Dr. Kieran Campbell and Dr. Sohrab Shah. All otherco-authors assisted in data collection, generation, and/or interpretation of the results.viEthical approval for the content in Chapters 2, 3, and 4 was obtained from the University ofBritish Columbia (UBC) Research Ethics Board (ethics numbers H08-01411, H14-02304, andH18-01090).viiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiiList of Supplementary Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviList of Symbols and Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviiAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xviiiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xx1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 High-grade serous ovarian cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.1 Epidemiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.2 Pathophysiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2 Intra- and inter-tumoural heterogeneity in HGSC . . . . . . . . . . . . . . . . . . 41.3 The tumour microenvironment in HGSC . . . . . . . . . . . . . . . . . . . . . . . 51.3.1 The immune microenvironment . . . . . . . . . . . . . . . . . . . . . . . . 61.3.2 Other microenvironmental factors . . . . . . . . . . . . . . . . . . . . . . 91.4 Emerging approaches to study tumour heterogeneity and the microenvironment . 101.4.1 Phylogenetic approaches for reconstructing tumour evolution . . . . . . . 101.4.2 T and B cell receptor sequencing . . . . . . . . . . . . . . . . . . . . . . . 131.4.3 Single cell methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161.5 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21viii2 Collection and processing of multi-site HGSC samples for high-throughputsequencing, PDX creation, and single cell experiments . . . . . . . . . . . . . 232.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.2.1 Summary of accrual process . . . . . . . . . . . . . . . . . . . . . . . . . . 242.2.2 Patient cohort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.2.3 Collection of surgical specimens and peripheral blood . . . . . . . . . . . 272.2.4 Sample preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.2.5 Patient-derived xenograft creation . . . . . . . . . . . . . . . . . . . . . . 282.2.6 Whole-genome sequencing of patient tumours . . . . . . . . . . . . . . . . 282.2.7 Single cell RNA-seq pilot project . . . . . . . . . . . . . . . . . . . . . . . 282.2.8 Single cell dissociation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.2.9 Viability sorting and assessment . . . . . . . . . . . . . . . . . . . . . . . 342.2.10 Single cell RNA-seq library preparation and quality control . . . . . . . . 342.2.11 Sequencing of single cell RNA-seq libraries . . . . . . . . . . . . . . . . . 342.2.12 Data curation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.3.1 Accrual of 41 HGSC cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.3.2 Construction of patient-derived xenograft models . . . . . . . . . . . . . . 392.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533 The evolutionary interface between tumour-infiltrating lymphocytes andcancer cells in multi-site HGSC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.2.1 Experimental Model and Subject Details . . . . . . . . . . . . . . . . . . 573.2.2 Method Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583.2.3 Quantification and Statistical Analysis . . . . . . . . . . . . . . . . . . . . 613.2.4 General statistical methods . . . . . . . . . . . . . . . . . . . . . . . . . . 803.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813.3.1 High-Resolution Multi-site Profiling of Immune and Malignant Populationsin the HGSC Tumor Microenvironment . . . . . . . . . . . . . . . . . . . 813.3.2 Tumor-Infiltrating Lymphocyte Subtypes Reveal Extensive IntrapatientVariation in Immune Responses across Peritoneal Sites . . . . . . . . . . . 85ix3.3.3 Evidence for Purifying Malignant Clonal Selection at Tumor Sites withHigh Epithelial Lymphocyte Infiltration . . . . . . . . . . . . . . . . . . . 893.3.4 T Cell, but Not B Cell, Clonotypes Show Evidence of Tumor Clone Tracking 953.3.5 Mutation Signatures Prognostically Associate with Patient-Level Immuno-logic Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1074 Probabilistic cell type assignment of single-cell transcriptomic data revealsspatiotemporal microenvironmental dynamics in human cancers . . . . . . . 1104.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1104.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1124.2.1 The CellAssign model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1124.2.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1154.2.3 Koh et al. dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1204.2.4 High-grade serous ovarian cancer . . . . . . . . . . . . . . . . . . . . . . . 1214.2.5 Follicular lymphoma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1224.2.6 Reactive lymph node data . . . . . . . . . . . . . . . . . . . . . . . . . . . 1254.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1264.3.1 Automated assignment of cell types with CellAssign . . . . . . . . . . . . 1264.3.2 Performance of CellAssign relative to state-of-the-art unsupervised andsupervised classification methods . . . . . . . . . . . . . . . . . . . . . . . 1284.3.3 Profiling the malignant and nonmalignant composition of high-grade serousovarian cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1364.3.4 Stromal subpopulations in the ovarian cancer microenvironment . . . . . 1384.3.5 Dissecting the lymphocyte composition of follicular lymphoma . . . . . . 1414.3.6 CellAssign uncovers compositional and phenotypic changes in the follicularlymphoma microenvironment . . . . . . . . . . . . . . . . . . . . . . . . . 1494.3.7 Malignant cell dynamics associated with early progression and transformation1524.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1575 Conclusions and Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . 1595.1 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1615.1.1 Understanding the molecular basis of immunologic infiltration patterns inHGSC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1615.1.2 Deciphering the mechanisms of treatment resistance in HGSC . . . . . . . 1625.1.3 Deconvolution of the HGSC microenvironment . . . . . . . . . . . . . . . 163x5.1.4 In situ profiling of the tumour microenvironment . . . . . . . . . . . . . . 1645.1.5 Guiding precision immunotherapies for HGSC . . . . . . . . . . . . . . . . 1655.2 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204Appendix A Chapter 3 Supplementary Materials . . . . . . . . . . . . . . . . . . 205Appendix B Chapter 4 Supplementary Materials . . . . . . . . . . . . . . . . . . 206xiList of Tables2.1 Identifiers of samples used for 10x Chromium library preparation and/or se-quencing. The tissue dissociation protocols and types of 10x Chromium librarypreparation are listed for each sample. . . . . . . . . . . . . . . . . . . . . . . . . 342.2 Patient identifiers, histotype as determined by final pathologic evaluation, andthe number of tumour samples collected per case. The number of tumour samplestransplanted (to create PDX models) is also shown. . . . . . . . . . . . . . . . . . 372.3 Sample and patient identifiers of HGSC samples used for whole genome sequencing. 392.4 Inventory of ovarian PDXs created from patient primary tumours. The strain ofmouse used (NSG or NRG) and whether or not a macroscopically visible tumourwas grown and harvested from each model (1 = grown, 0 = not grown) are indicated. 492.5 Summary statistics for PDX collection by strain for HGSC tumours. The numberpatients and samples with at least 1 grown PDX, along with the engraftmentrate (by model, sample, and patient) for models collected ≥ 1 year ago are shown. 513.1 Studied Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 844.1 HGSC samples profiled by single cell RNA-seq. Raw and filtered correspond toraw and preprocessed cell counts, respectively. . . . . . . . . . . . . . . . . . . . . 136xiiList of Figures1.1 Fallopian tube origin of HGSC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2 Intratumoural heterogeneity and treatment resistance. . . . . . . . . . . . . . . . 41.3 Cell types in the tumour microenvironment. . . . . . . . . . . . . . . . . . . . . . 61.4 T- and B-cell-mediated mechanisms of antitumour immunity. . . . . . . . . . . . 71.5 Next-generation sequencing enables clonal inference based on allelic fractions. . . 111.6 V(D)J recombination and TCR/BCR structure. . . . . . . . . . . . . . . . . . . . 141.7 Single cell WGS library preparation by direct library preparation (DLP) andclonal analysis of CNVs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171.8 Bulk RNA-seq vs. single cell RNA-seq. . . . . . . . . . . . . . . . . . . . . . . . . 191.9 Thesis outline. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.1 Overall HGS project pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.2 Single cell RNA-sequencing workflow. . . . . . . . . . . . . . . . . . . . . . . . . 302.3 HGSC sample accrual. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372.4 HGSC PDX accrual. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502.5 Engraftment rates of HGSC tumours as a function of time. . . . . . . . . . . . . 522.6 Surgery to euthanasia times for models by growth. . . . . . . . . . . . . . . . . . 533.1 Multisite Profiling of HGSC Reveals Three Distinct TIL Subtypes with ExtensiveIntrapatient Variation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 823.2 Schematic Diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853.3 TIL Densities in Multisite HGSC. . . . . . . . . . . . . . . . . . . . . . . . . . . . 863.4 Differences in Cancer-Lymphocyte Hotspot Colocalization within Tumor Epithe-lium between TIL Subtypes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 883.5 Relationship between Malignant Clone Composition and BCR Clonotype Reper-toires. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 903.6 Patterns of Clonal Complexity, Relationship to TIL Subtypes, and Expression ofInhibitory Immune Checkpoint Molecules. . . . . . . . . . . . . . . . . . . . . . . 923.7 Low ITH, Neoantigen Depletion, and Subclonal HLA Loss of Heterozygosity inSamples with High Epithelial TIL. . . . . . . . . . . . . . . . . . . . . . . . . . . 93xiii3.8 TCR/BCR Repertoire Diversity and Within-Patient Similarity, and Relationshipto TIL Profiles. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 963.9 Correlations between TIL Densities/Subtypes and TCR/BCR-Seq Data. . . . . . 973.10 Relationships between Malignant Clone Composition and TCR Clonotype Reper-toire. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1003.11 Mutation Signatures Inferred from MMCTM. . . . . . . . . . . . . . . . . . . . . 1023.12 Mutational Subtypes Prognostically Associate with Immune Patterns in HGSC. . 1053.13 Differentially Expressed Pathways between Mutational Subtypes (for HRD versusFBI, TD versus FBI, and HRD versus TD Comparisons) in OV-AU Cases. . . . . 1064.1 Fitting single cell RNA-seq simulation models to real data . . . . . . . . . . . . . 1164.2 Fitting single cell RNA-seq simulation models to real data . . . . . . . . . . . . . 1174.3 Schematic description of CellAssign. . . . . . . . . . . . . . . . . . . . . . . . . . 1274.4 Benchmarking runtime speed of CellAssign. . . . . . . . . . . . . . . . . . . . . . 1294.5 Performance of CellAssign on simulated data. . . . . . . . . . . . . . . . . . . . . 1304.6 Simulation performance across a range of proportions of differentially expressedgenes, using differential expression parameters derived from B and CD8+ T cells. 1324.7 Simulation performance across a range of proportions of randomly flipped entriesin the binary marker gene matrix, using differential expression parameters derivedfrom comparing naive CD8+ and naive CD4+ T cells. . . . . . . . . . . . . . . . 1334.8 Performance of clustering methods on FACS-purified H7 human embryonic stemcells in various stages of differentiation. . . . . . . . . . . . . . . . . . . . . . . . 1354.9 Composition of the HGSC microenvironment. . . . . . . . . . . . . . . . . . . . . 1374.10 Proportions and probabilities of cell type assignments. . . . . . . . . . . . . . . . 1384.11 Stromal subpopulations in the HGSC microenvironment. . . . . . . . . . . . . . . 1404.12 CellAssign infers the composition of the follicular lymphoma microenvironment. . 1424.13 Expression of select marker genes in follicular lymphoma single cell expression data.1434.14 Temporal changes in nonmalignant cells in the follicular lymphoma microenviron-ment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1454.15 Expression of κ and λ light chain constant region genes in nonmalignant B cells. 1464.16 Expression of selected marker genes in scvis embedding of reactive lymph nodedata. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1474.17 Differential expression results for malignant vs. nonmalignant B cells in FL1018and FL2001. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148xiv4.18 Pathway enrichment of significantly upregulated genes in CD8+ T cells at T2 vs.T1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1504.19 Differentially expressed genes between T follicular helper and other CD4 T cellsin recurrence and diagnostic samples. . . . . . . . . . . . . . . . . . . . . . . . . . 1514.20 Temporal changes in malignant cells in the follicular lymphoma microenvironment.1534.21 Differentially expressed genes between malignant B cells from T2 vs. T1. . . . . 1554.22 Pathway enrichment of significantly downregulated genes in malignant B cells atT2 vs. T1 in FL1018. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156xvList of Supplementary MaterialsSupplementary Table A.1. Primers for deep amplicon sequencing.Supplementary Table A.2. TIL densities, TIL subtypes, molecular subtypes, epithelialcolocalization measures from histologic image analysis, somatic SNV and rearrangement counts,ITH measures, and TCR and BCR repertoire diversity.Supplementary Table A.3. Neoepitope tableSupplementary Table A.4. HLA LOH tableSupplementary Table A.5. Mutation signature proportions and cluster assignments formultisite HGSC, OV-AU, and [3] cohorts.Supplementary Table A.6. OV-AU differential gene expression analysis tableSupplementary Table A.7. TCGA sample information, foldback-HLAMP, and cytotoxicityclustersSupplementary Table B.1. Performance measures on simulated data.Supplementary Table B.2. Marker gene matrices.Supplementary Table B.3. Pathway enrichment results.xviList of Symbols and AbbreviationsBCR B-cell receptorBH Benjamini-HochbergCNV Copy number variantDLP Direct library preparationDNA Deoxyribonucleic acidFACS Fluorescence-activated cell sortingFBI Foldback inversionFDR False discovery rateH&E Hematoxylin & eosinHLA Human leukocyte antigenHGSC High-grade serous ovarian cancerHRD Homologous recombination deficiencyITH Intratumoural heterogeneityLOH Loss of heterozygosityMCMC Markov chain Monte CarloMHC Major histocompatibility complexPARP Poly ADP-ribose polymerasePDX Patient-derived xenograftRNA Ribonucleic acidscRNA-seq Single-cell RNA sequencingSNV Single-nucleotide variantSV Structural variantTCR T-cell receptorTIL Tumour-infiltrating lymphocyteTME Tumour microenvironmentWGS Whole-genome sequencingxviiAcknowledgementsI would like to extend my deepest gratitude to my supervisors Dr. Sohrab Shah and Dr.Wyeth Wasserman. I consider myself extremely fortunate to have been co-mentored by two ofthe world’s leading experts in genome analysis who both took an acute interest in my training.Thank you both for providing me with numerous incredible research and career opportunities,and for not holding back despite my shorter timeline and obligations to clinical studies in theMD/PhD program. I was truly privileged to have had experience coordinating projects, designingexperiments, performing wet lab techniques, and interacting with stellar cancer researchers atinternational workshops in your labs. Most of all, thank you for believing in and supportingme in my endeavours, even when they did not directly align with your own interests. I amforever indebted for these opportunities that have allowed me to grow, as both a scientist andan individual, over the last few years.My MD/PhD has brought me close to many talented scientists who I have had the privilegeof getting to know. Firstly, I would like to thank my committee members Dr. Brad Nelson,Dr. Daniel Renouf, Dr. Raymond Ng, and Dr. Martin Hirst for their advice, feedback, andenthusiasm on my research. I am especially grateful to Dr. Brad Nelson for going aboveand beyond to support me as a collaborator and mentor throughout my training, especiallyduring the beginning of my graduate studies. I would also like to recognize the many incredibleindividuals from the Shah, Nelson, Holt, Aparicio, and Huntsman labs who were always willingto help me — you have made the research in this thesis possible. Specifically, I would like tothank my research and clinical collaborators Dr. Andrew McPherson, Dr. Phineas Hamilton,Dr. Alex Miranda Rodriguez, Dr. Katy Milne, Dr. Jessica McAlpine, and Dr. David Huntsman,who helped make Chapter 3 possible. Thank you to Dr. Kieran Campbell for supervising andco-leading Chapter 4 with me, and Dr. Ciara O’Flanagan and Dr. Samuel Aparicio for theirhelp in advising and executing the project. Special thanks to Jamie Lim for supporting much ofthe work in this thesis through her diligent experimental work and enthusiasm to innovate. Also,thank you to the research staff and fellows who I had the pleasure of working with, including Dr.Camila de Souza and Dr. Daniel Lai. Big thanks to Cynthia Berry, Yessie Werner, Jad Maanaki,Dora Pak, Carolyn Lui, and Sogol Tahmasbi for their incredible administrative support. Thankyou to Dr. Anthony Mathelier and Dr. Fong Chun Chan who provided invaluable guidanceduring the early phases of my MD/PhD, and to Dr. Torsten Nielsen, Dr. Lynn Raymond, andxviiithe UBC MD/PhD program. And of course, none of this work would have been possible ormeaningful without the patients and their families that support us.I am grateful to the Canadian Institutes of Health Research, BC Children’s Hospital, UBCMD/PhD Program, BC Cancer Foundation, and Canadian Cancer Society for generous financialsupport for Chapter 2, Chapter 3, and Chapter 4. Additionally, thanks to the Vanier CanadaGraduate Scholarship Program, the Michael Smith Foreign Study Supplement Program, theCanada Graduate Scholarship-Masters Program, the UBC Four Year Fellowship Program, theUBC Faculty of Medicine, the UBC Faculty of Science, the UBC Bioinformatics Program, andthe Canadian Conference for Ovarian Cancer Research for providing scholarships and travelawards for my studies.Finally, I would like to thank my family and friends for their unwavering support andpatience. To my mother and father — thank you for always looking out for me, putting me first,and understanding when I needed time to work. None of what I have accomplished today andwill achieve would be possible without the bold sacrifice you made to move to Canada. To mygrandfather and grandmother — thank you for raising and loving me. And to my close friends —thank you for making the effort to stay connected and pulling me away from my books once inawhile.xixDedicationTo my parents, grandfather, late grandmother, and loving family.xxChapter 1IntroductionCancer is a disease of the genome and one of the leading causes of death in Canada [4]. Mostcancers develop from a single cell that undergoes successive cycles of expansion, diversification,and pruning [5]. Thus, cancers are non-homogeneous mixtures of genomically and phenotypicallydistinct populations of tumour cells called clones. The process by which these clones expand,shrink, and diversify over time is known as clonal evolution [5]. The phenotypic diversity ofcancer cells generated by clonal evolution explains some cases of resistance, where pre-existing,treatment-resistant clones survive initial therapeutic assault and expand to give rise to tumourcells present at relapse [6].Importantly, these processes do not occur in isolation. Malignant cells occupy niches sharedby non-malignant cells, such as lymphocytes, macrophages, fibroblasts, adipocytes, and pericytes[7]. These cells can have tumour-promoting or inhibitory functions that alter the evolutionarytrajectories of tumour cells. For example, tumour-infiltrating lymphocytes (TILs) can respondto tumour-associated antigens, mounting anti-tumour immune responses [8]. On the otherhand, cancer-associated fibroblasts can promote metastasis, angiogenesis, and tumour cellproliferation through extracellular matrix remodelling and cytokine secretion [9]. Together,the dynamic network of interactions formed by malignant and non-malignant cells forms thetumour microenvironment (TME). Like tumour cells, the composition and properties of themicroenvironment can change over carcinogenesis, progression, and treatment [7]. Understandinghow the microenvironment shapes the evolutionary histories of tumours will aid in predictinghow tumours will respond to treatment.I investigate these phenomena in high-grade serous ovarian cancer (HGSC), a subtypeof ovarian cancer characterized by rampant dissemination throughout the peritoneal cavity,forming an ideal substrate for investigating how tumour cells evolve in diverse microenvironmentalcontexts. In Section 1.1, I review the epidemiology and pathophysiology of HGSC. In Section 1.2,I discuss the current understanding of clonal diversity in HGSC. Section 1.3 provides a briefoverview of the microenvironmental cell types most relevant to HGSC, and Section 1.4 reviewscontemporary experimental and computational approaches to study tumour evolution and the1microenvironment. Finally, Section 1.5 outlines the central aims of this thesis – to characterizethe TME of HGSC and understand how non-malignant cells impact malignant evolutionarydynamics.1.1 High-grade serous ovarian cancer1.1.1 EpidemiologyIn North America, ovarian cancer is the leading cause of death from gynecological malignanciesand the fifth most common cause of cancer death [10]. The most common histopathologicalsubtype of ovarian cancer, high-grade serous ovarian cancer, makes up 70% of diagnoses and hasa 5-year survival rate of under 50% [10]. HGSC patients often present with peritoneal spreadto invasive foci in the omentum, small bowel, and other organs. The current standard-of-caretreatment, primary debulking surgery followed by combination platinum-taxane chemotherapy,is effective at treating the primary tumour, but the disease almost always recurs (80%) [11].Despite extensive research into developing new screening and therapeutic strategies, patientsurvival has not improved substantially over the last 3 decades [12].HGSC is typically sporadic, with approximately 10-20% of cases being hereditary [13].Most of these correspond to germline BRCA1 or BRCA2 variants, which are associated withsuperior response to platinum chemotherapy due to the impaired ability of BRCA-mutatedtumours to repair platinum-induced DNA damage through homologous recombination (HR)[14, 15]. The recent introduction of poly ADP-ribose polymerase (PARP) inhibitors for HGSC,a class of drugs that exploits synthetic lethality by impairing compensatory DNA repairpathways, has demonstrated improved outcomes especially among HR-deficient (HRD) cases[16, 17]. Nevertheless, outcomes remain bleak, especially for cases without HRD which constituteapproximately half of HGSC [3].1.1.2 PathophysiologyTraditionally, HGSC was thought to originate from the ovaries in cortical inclusion cysts ofovarian surface epithelium [18]. While some studies still support an ovarian origin of HGSC, mostnew evidence points toward the epithelium of the distal fallopian tube as the anatomic origin ofHGSC in the majority of cases (Figure 1.1) [19, 20]. Serous tubal intraepithelial carcinoma(STIC) lesions found on the fallopian tubes of BRCA carriers share genomic features of HGSC[21]. Furthermore, phylogenetic analysis of mutations present in STICs and HGSC tumoursfrom the same patients has established these lesions as the precursors to most occurrences of2HGSC [21].Figure 1.1: Fallopian tube origin hypothesis of HGSC pathogenesis. HGSC is thought to bederived from fallopian tube epithelial cells that acquire TP53 mutations. These cells may eventuallydevelop into histologically detectable STIC lesions. STICs seed the ovary and possibly also metastaticlesions elsewhere. Used with permission from https://www.cancer.gov/news-events/cancer-currents-blog/2017/ovarian-cancer-fallopian-tube-origins (Carolyn Hruban).Almost all HGSC tumours harbour mutations in TP53, which occur as an early event indisease progression [14, 22]. Approximately 18% of cases have somatic BRCA1 or BRCA2variants, which are largely exclusive with CCNE1 amplification (20%); together, when combinedwith germline variants and epigenetic silencing, approximately half of all HGSC cases have HRD[14]. The prevalence of TP53 mutations and HRD leads to incompetence in DNA repair. Thus,widespread chromosomal aberrations, aneuploidy, and severely disrupted karyotypes are definingcharacteristics of HGSC. Other genes affected by recurrent mutations and copy number events3include MYC (> 20%), PIK3CA (17%), NF1 (12%), RB1 (11%), KRAS (11%), PTEN (7%),and CDK12 (3%) [14].Recently, integrated genomic analyses of single-nucleotide (SNVs) and structural variants(SVs) derived from whole-genome sequencing studies of HGSC have revealed 2 major genomicsubtypes of HGSC: homologous recombination-deficient (HRD) and foldback inversion-associated(FBI) [3, 23]. HRD cases, comprising approximately 50% of HGSC, are primarily defined bythe presence of the HRD-associated SNV signature along with short deletions and tandemduplications. BRCA1 and BRCA2 -mutated tumours are included in this subgroup and aredefined by distinct SV signatures (BRCA1 by short tandem duplications and BRCA2 by shortdeletions) [3]. FBI tumours make up most of the remainder of cases (40%) and harbour foldbackinversions – duplicated sequences that face away from a breakpoint [3]. Foldback inversionsare thought to arise through successive breakage-fusion-bridge cycles, co-occur with CCNE1amplification, and are associated with poor survival in HGSC [24]. A third tandem duplicatorsubgroup, associated with CDK12 mutations, accounts for a final, minor fraction of cases (10%)[25] and has been linked to the worst outcomes in a recent preprint [26].1.2 Intra- and inter-tumoural heterogeneity in HGSCThe clonal evolution theory of cancer cell populations posits that tumours arise from a singleorigin and diversify through acquisition of genomic alterations over time [5]. Ultimately, thisprocess gives rise to genotypically distinct populations of cells, called clones, with correspondingphenotypes (Figure 1.2). Computational methods permit identification of clones based onsomatic variants from bulk and single-cell DNA sequencing data (Section 1.4.1).Figure 1.2: Pre-existing heterogeneity in a tumour increases the likelihood of treatmentresistance due to the presence of a resistant clonal population. The red, resistant popula-tion expands following treatment due to a selective bottleneck. Used with permission fromhttps://www.nature.com/articles/nrclinonc.2017.166.4Previous studies in HGSC have revealed a significant degree of clonal diversity in treatment-na¨ıve tumours. On average, only 50% of mutations are shared across all samples from a tumour[27, 28]. Through detailed Bayesian probabilistic phylogenetic reconstruction of clonal lineages,our group has recently shown that tumours can harbour clones from divergent evolutionarylineages with markedly different mutational profiles [29]. Clonal diversity also exists betweentumours from an individual patient, and there is significant cellular migration between metastaticsites [29]. Maximum parsimony reconstruction of clonal dissemination patterns in HGSC isconsistent with monoclonal and polyclonal seeding from a single diverse site, typically an ovaryor fallopian tube, in most cases [29]. However, recent organoid studies have established thatmulticellular aggregates (MCAs) are significantly more successful than single cells at invadingovarian mesothelium [30, 31]. These MCAs can contain phenotypically and morphologicallydistinct populations of cells [30]. Metastatic foci in mouse models implanted with phenotypicallymixed tumour cells also contained phenotypically mixed populations [31]. Thus, successfulmetastatic spread likely involves multicellular aggregates rather than single cells, hinting atthe possibility that polyclonal seeding and reseeding followed by pruning may be a commonoccurrence.Contrary to the assumption that the intraperitoneal space allows for indiscriminate admixtureof tumour cells, we have observed restricted clonal mixing in the majority of HGSC patients [29].As such, local factors, such as the tumour microenvironment, may be involved in patterningclonal seeding and establishment. Understanding how these factors affect clonal migration mayoffer crucial insights into developing strategies to contend with the burden of metastatic diseasein HGSC.1.3 The tumour microenvironment in HGSCSolid tumours are ecosystems populated by a milieu of malignant and non-malignant celltypes, including tumour cells, fibroblasts, immune cells, endothelial cells, and adipocytes(Figure 1.3) [7]. Collectively, these cell types and the interactions that occur between themcompose the tumour microenvironment (TME). Tumour cells can shape the composition of theTME, sustaining growth and proliferation while evading immune-mediated elimination [32, 33].Likewise, the TME can impose extrinsic pressures such as hypoxia, altering the metabolicprocesses of tumour cells and contributing to the development of ’cancer hallmark’ traits [32].Inflammatory mediators in the TME can contribute to tumourigenesis through the pro-growthactivity of cytokines released during inflammation [32]. These reciprocal interactions betweentumour cells and the rest of the TME represent molecular targets that can potentially be5exploited by therapeutics.Figure 1.3: Repertoire of immune and non-immune cell types in the tumour microenvironment.Used with permission from http://tcr.amegroups.com/article/view/1549/2264.1.3.1 The immune microenvironmentPopulations of immune cells in the TME include cytotoxic T cells, helper T cells, regulatoryT cells, B cells, NK cells, macrophages, and granulocytes (Figure 1.3) [32]. In HGSC, theimmune microenvironment is dominated by T cells, B cells, and macrophages [1, 34, 35].As in virtually every solid cancer type, cytotoxic (CD8+) tumour-infiltrating lymphocytes(TILs) have been associated with increased survival in HGSC [36–38]. Ordinarily, T cellsrecognize and respond to aberrant peptides presented by the Major Histocompatibility Complex(MHC/HLA). Recent studies have reported that CD8+ TILs can recognize somatically mutatedpeptides in HGSC, suggesting that TIL mount anti-tumour responses in part through neoantigenrecognition [39]. Survival is more strongly associated with intraepithelial rather than stromalCD8+ TIL, implying that spatial localization of TIL is important for anti-tumour immunity[40]. Immunohistochemistical studies have identified that these intraepithelial CD8+ T cells6preferentially express CD103, an integrin subunit involved in epithelial localization of normalintraepithelial lymphocytes, as well as activation and cytolytic markers [40], supporting theinterpretation that intraepithelial CD8+ T cells are involved in anti-tumour immunity.However, CD8+ T cells do not function independently. In addition to CD8+ T cells,the presence of additional TIL types – B cells and plasma cells – is associated with superioroutcomes in HGSC [38]. T and B cells spatially co-localized in tertiary lymphoid structures andother lymphoid aggregates may support interactions between these cell types in the context ofanti-tumour immunity (Figure 1.4) [41]. B and plasma cells may be involved in anti-tumourimmunity through autoantibody production, direct cytotoxicity, Th1/Th2 polarization, andantigen presentation (Figure 1.4) [42]. However, the relative contribution of each of thesemechanisms in HGSC is poorly understood. On the other hand, regulatory CD4+CD25+FOXP3+T cells curtail T effector functions by inhibiting type 1 cytokine (IFNγ, IL-2) production and Tcell proliferation [34, 43, 44].Figure 1.4: T cell- and B cell-mediated mechanisms of antitumour immunity. CD8+ T cells exertdirect cytotoxic effects on tumour cells through granzyme and perforin secretion. CD4+ T cellscan license dendritic cells to induce activation of CD8+ T cells, and activate B cells. Used withpermission from Brad Nelson AACR 2017.During acute infection, T cells differentiate into effector populations to mount antigen-specificresponses. Following antigen clearance, these effector populations shrink to small memory T7cell pools capable of rapid reactivation upon future encounter with the same antigen. In thetumour microenvironment, however, prolonged exposure of effector T cells to antigen-dependingsignalling can lead to the development of an “exhausted” phenotype characterized by progressiveimpairment of effector activity [45]. In HGSC, CD8+ T cell infiltration is linked to expressionof exhaustion markers PD-1 and CTLA4, and CD8+CD103+ TIL are associated with PD-1positivity by immunohistochemistry [46, 47]. Nevertheless, CD8+PD-1+ TIL appear to retainsome degree of effector functionality and are associated with superior patient survival [47].The use of immune checkpoint inhibitors — monoclonal antibodies that block receptor-ligandinteractions implicated in T cell exhaustion — to mobilize exhausted T cells has been associatedwith superior clinical trajectories compared to standard-of-care therapy in melanoma and sometypes of lung and colorectal cancer [8, 48]. However, despite the prevalence of PD-1+ andCTLA4+ TIL in HGSC, response rates to checkpoint inhibitor blockade in HGSC have beendisappointing [49, 50]. As such, our understanding of immune checkpoints in HGSC remainsincomplete.Thus far, most studies of anti-tumour immunity have focused on the adaptive immune system.However, the observation that some patients show T cell priming to tumour-associated antigensprior to treatment (“spontaneous” T cell priming) has driven inquiry into understanding theinnate immune mechanisms that ultimately give rise to intratumoural T cell infiltration. Themajor cell types involved in innate immunity include NK cells, macrophages, and dendritic cells[51]. NK cells are lymphocytes that express inhibitory receptors that bind to HLA complexes,and are thus thought to selectively target tumour cells that lack HLA expression. Consequently,tumour cells that downregulate HLA evade T cell recognition are vulnerable to NK-mediatedcytotoxicity through granzyme and perforin release [52]. Macrophages, a type of leukocyte inthe monocyte lineage, engulf abnormal cells or cellular debris through phagocytosis. The twomain phenotypic subtypes of macrophages are pro-inflammatory, phagocytic M1 macrophagesand tissue repair-associated M2 macrophages. In the context of the tumour microenvironment,M1 macrophages arise earlier in tumourigenesis, tend to promote inflammation and generallyoppose tumour progression, whereas M2 macrophages dominate tumours at the time of diagnosisand are generally immunosuppressive and pro-tumourigenic [51]. However, the inflammatoryresponse associated with M1 macrophages may also induce carcinogenesis [53]. In HGSC,M2-associated expression signatures have been linked to a stromal reorganization phenotype andinferior outcomes [54], whereas M1-associated genes including type I interferons are associatedwith superior patient survival [55]. Tumour transplantation into type I IFNR(-/-) mice resultedin reduced T cell responses against tumour antigens due to deficiencies in CD8+ T cell primingby antigen presenting cells (APCs), demonstrating that type I interferon signalling is necessary8for antitumour immunity. Studies to investigate factors upstream of type I interferon signallingin the tumour microenvironment have highlighted the STING (stimulator of DNA sensing genes)DNA sensing pathway [56]. STING is activated by the presence of cytosolic DNA – typicallyfrom intracellular pathogens, such as viruses and parasites – inducing innate immunity throughtype I interferon production. In tumours, STING pathway activation has been associated withthe presence of tumour-derived DNA in APCs, and STING-deficient mice show defective T cellpriming [57]. Thus, STING-dependent sensing of tumour DNA may lead to type I interferonproduction and T cell priming in cancers. However, this mechanism has yet to be shown in thecontext of HGSC specifically.1.3.2 Other microenvironmental factorsMajor non-immune cellular populations in the HGSC microenvironment include fibroblasts andendothelial cells. Fibroblasts are the major non-immune cellular component of the tumourstroma, involved in wound healing and responsible for extracellular matrix (ECM) depositionand maintenance through collagen and matrix metalloproteinase (MMP) production. In cancers,fibroblasts can be co-opted to facilitate tumour progression, transitioning to a myofibroblasticphenotype characteristic of cancer-associated fibroblasts (CAFs) [58]. CAFs are thought topromote tumour progression through a number of mechanisms including pro-growth signalling,vascular stabilization, and ECM remodelling [58]. Mechanistic studies have revealed a linkbetween fibroblasts and platinum resistance in HGSC, demonstrating that fibroblasts candiminish intranuclear platinum accumulation in cancer cells in vitro [59].Angiogenesis refers to the formation of new vasculature from existing blood vessels. Tumourcells rapidly proliferate, requiring the complementary development of vasculature to sustainincreasing nutrient and oxygen demand. When angiogenesis cannot keep up with tumour growth,tumour regions that receive insufficient perfusion eventually become necrotic [60].The inner lining of normal blood vessels is composed of a monolayer of squamous cells, calledendothelial cells, that restrict influx and efflux between the vascular lumen and surrounding tissue.In contrast, tumour vasculature often displays an altered phenotype characterized by endothelialcell disorganization and abnormal branching [61]. The defective endothelium is associated withincreased vascular permeability, faciltating nutrient extravasation and hematogenous metastasis[60]. The tumour endothelium can also act as a physical barrier to immune cell infiltration,preventing circulating lymphocytes from reaching certain tumour regions to mount anti-tumourimmune responses [61]. Ultimately, the formation of tumour-associated endothelium is thoughtto be mediated by cancer cells. Tumour cells can alter normal endothelium by secreting pro-angiogenic and vasodilatory factors, such as vascular endothelial growth factor (VEGF), and9imposing biomechanical strain by encroaching on the vasculature itself [61]. Additionally, somegroups have observed evidence of vasculogenic mimicry, where tumour cells are capable oftrans-differentiating into endothelial-like cells to form “pseudo-vasculature” [62].Cancer cell growth and metabolism can also be influenced by other microenvironmental factors.Lack of oxygenation – hypoxia – in poorly perfused regions of the tumour microenvironmentleads to various tumour cell adaptations including pro-angiogenic signalling and an increasedreliance on anaerobic respiration to maintain adenosine triphosphate (ATP) production [63].Anaerobic respiration through glycolysis results in extracellular accumulation of tumour-derivedlactate, impairing the ability of intratumoural T cells to perform glycolysis within hypoxicmicroenvironments [64]. Thus, tumour hypoxia can impair TIL activity. Moreover, lactate caninduce angiogenesis and promote tumour cell migration [65]. Correspondingly, the expression ofhypoxia-associated factors is associated with decreased overall survival and chemoresistance inHGSC [66].1.4 Emerging approaches to study tumour heterogeneity andthe microenvironment1.4.1 Phylogenetic approaches for reconstructing tumour evolutionGenomic heterogeneity in cancers can be assayed with bulk or single cell sequencing. Bulkgenomic sequencing generates sequencing reads from DNA derived from thousands to billions ofcells that can be mapped onto a reference genome to identify germline and somatic variants(Figure 1.5) [67]. Currently, the most widely-used methods for bulk DNA sequencing are shotgunsequencing technologies, which generate millions of overlapping reads, providing a coarse viewof genomic heterogeneity through analysis of the sequencing depth and allelic fraction of eachvariant (Figure 1.5) [67]. Recently developed single cell sequencing technologies utilize cellularbarcodes that allow each read to be mapped back to its source cell so constituent genotypesof mixed cellular populations can be directly profiled [68]. However, these technologies werenot available widely or at scale until recently. Furthermore, key tradeoffs between sequencingdepth for variant calling and coverage uniformity for copy number profiling in single cell whole-genome sequencing, discussed in Section 1.4.3.1, confound complete reconstruction of cell-levelgenotypes. Meanwhile, bulk whole-genome sequencing, which yields aggregate measurements ofvariant abundance for all sampled cells, has been successfully employed to characterize tumourevolution and profile driver mutations in many cancer types [1, 27, 29, 69, 70]. Evaluatinggenetic heterogeneity from bulk sequencing data demands statistical methods that can robustly10deconvolute clonal genotypes, abundances, and phylogenetic relationships in the presence ofcontaminating normal cells and aneuploidy.Figure 1.5: Whole-genome sequencing of heterogeneous cellular populations generates allelic readcounts which are proportional to mutational prevalence (accounting for copy number and cellularity).Mutational prevalence values can be deconvolved, e.g. with PyClone [71], to yield clonal genotypes andprevalences. Used with permission from https://www.nejm.org/doi/full/10.1056/NEJMra1204892..Reproduced with permission from [72], Copyright Massachusetts Medical Society.Somatic variant calling from whole-genome sequencing data yields single-nucleotide variants(SNVs), short insertions and deletions (indels), copy number profiles (CNVs), and structuralvariant calls (SVs). In the majority of the discussion below, I focus on SNVs as these aretypically the most common class of somatic variants. For each SNV, some reads map to thealternative (somatically mutated) allele, while the remainder map to the germline allele. Thefraction of reads mapped to the alternative allele (henceforth referred to as the variant allelefraction, or VAF) can be used to compute the total cellular prevalence of clones harbouring thevariant after accounting for the fraction of contaminating normal cells and allelic copy number11(Figure 1.5). Well-established methods such as PyClone [71] and SciClone [73] exploit allelicread counts and copy number profiles with Bayesian mixture models to identify mutationalclusters: groups of SNVs present at similar cellular prevalence from shared clonal membership.However, clones present at similar proportions within a tumour can confound these models, dueto significant overlap between clone-specific SNV clusters [71]. Data for multiple samples withshared clonal populations – for instance, from multiple metastatic sites within a patient or fromdifferent timepoints – provides greater resolution for resolving these scenarios [29, 71].In order to construct a clonal phylogeny from SNV clusters, cluster acquisition events mustbe temporally ordered. Two computational models that tackle this problem are CITUP [74]and LiCHEE [75]. Fundamentally, these methods rely on a set of common assumptions: (1) thetotal cellular prevalence of SNV clusters from distinct lineages cannot exceed the prevalence oftheir most recent common ancestor (pigeonhole principle), (2) each SNV is acquired only once(infinite sites assumption) and cannot be lost, and (3) descendant SNV clusters must have lowercellular prevalence than their immediate ancestors. Other tools, such as PhyloWGS [76], performmutation clustering and phylogeny inference simultaneously by leveraging distributions of treesto describe clonal mixing proportions. Intratumoural heterogeneity can be quantified followingclonal decomposition. Heterogeneity can be expressed in terms of the relative proportionsof clonal populations (mixture entropy) and in terms of the genotypic divergence betweenco-existing clones (phenotypic divergence) [29].Each of the computational methods for mutation cluster inference and phylogeny reconstruc-tion described above relies on key assumptions to reduce the search space of valid solutions.However, these assumptions may be invalid under certain realistic regimes. SciClone ignoresnon-diploid and copy number neutral loss-of-heterozygosity (LOH) regions of the genome [73],rendering it unsuitable for highly aneuploid cancer types such as HGSC. While PyClone usesinformation from copy number-altered regions, it does not allow for multiple variant clonalpopulations with different copy number profiles at any given locus [71]. Variants falling in regionsof subclonal (non-integer) copy number may violate this assumption. Moreover, a small numberof loci likely violate the infinite sites assumption imposed by phylogeny inference methods.Single cell sequencing, described in Section 1.4.3, can help resolve clonal substructure in theseregimes. For instance, a recent method called ddClone [77] jointly leverages bulk and single cellvariant information to infer clonal subpopulation abundances based on genotypes informed bysingle cell sequencing.121.4.2 T and B cell receptor sequencingLymphocyte differentiation from common lymphoid progenitors in the bone marrow proceedsthrough a ordered series of events marked by the acquisition of certain cell surface proteins (e.g.CD3, CD4, CD8, CD19, CD20) and antigen-specific receptor molecules [78, 79]. The eventualantigenic specificity of each lymphocyte is determined by the structure of its antigen-specificreceptor – the T cell receptor (TCR) or B cell receptor (BCR) in T or B cells, respectively(Figure 1.6) [80]. To contend with the enormous space of potential antigens, a diverse repertoireof T and B cell receptor sequences are generated by somatic rearrangement of constituent germlinevariable (V), diversity (D), and joining (J) gene segments (Figure 1.6).13BCR TCRDNA: V(D)J recombinationFigure 1.6: Top: Structure of a membrane-bound B cell receptor (BCR) and an α:β T cell receptor(TCR). The outer portions of the BCR/TCR, primarily composed of variable (V) chain sequence,are directly involved in epitope binding. Bottom: Depiction of the combinatorial diversity generatedfrom V(D)J recombination and somatic hypermutation (for BCRs). Used with permission fromhttps://www.cell.com/trends/immunology/fulltext/S1471-4906(14)00155-0.In germline DNA, multiple V, D, and J variants are present (Figure 1.6). During somaticVDJ rearrangement, a D gene variant is first joined to a J gene, followed by V gene addition tothe resulting D-J fragment. This process is faciltated by V(D)J recombinase, a collection ofenzymes including RAG1, RAG2, TdT, and Artemis that bind to recombination signal sequencesflanking V, D, and J genes [80]. For some TCR/BCR subunits, recombination occurs directlybetween V and J genes without the D segment. In α:β T cells – the dominant subpopulationof T cells – the TCR is composed of one VJ α subunit and one VDJ β subunit [80]. Similarly,14BCRs are composed of heavy (VDJ) and light (VJ) chains. Thus, the combinatorial space of V,D, and J germline genes forms the basis for T and B cell receptor diversity. Additional sequencediversity arises from terminal deletion and insertion events that occur at the ends of V, D, andJ sequences during recombination (Figure 1.6) [80]. These junctional sequences, together withthe end of the V and beginning of the J segments (and in TCR-β/BCR-heavy chains, the entireD gene) comprise the hypervariable portion of the TCR/BCR referred to as the CDR3 [80]. InB cells, somatic hypermutation introduces additional variants, primarily SNVs, concentrated inthe CDR3 region (Figure 1.6) [81]. In total, the estimated sequence diversity of TCRs andBCRs generated through V(D)J recombination exceeds 1015 [82].Due to the immense sequence diversity created through V(D)J recombination, the probabilityof two clonally unrelated T or B cells sharing the same receptor sequence is highly unlikely.Thus, TCR/BCR sequencing enables identification and quantification of clonally related familiesof T and B cells [83]. Expanded clonal families are thought to correspond to T and B cellsthat have been stimulated by antigen to proliferate. TCR/BCR repertoire profiling can beperformed using either genomic DNA or RNA templates as starting material [83, 84]. GenomicDNA-based protocols enable direct quantification of lymphocyte abundance, as each T/B cellonly produces one productive TCR/BCR species. However, genomic DNA protocols require V-and J-gene-specific primers that can result in PCR bias [84], and are prone to off-target primingof non-rearranged VDJ genes [85]. RNA-based protocols enrich for and amplify TCR/BCR-derived RNA or cDNA using sequence-specific primers to V and/or C genes [84]. Clone-specificread counts derived from RNA-based TCR/BCR sequencing roughly correspond to clonotypeabundance, but are affected by variability in TCR/BCR expression levels [85]. Nevertheless,RNA-based methods generally capture more receptor sequence diversity than DNA-basedprotocols. Furthermore, RNA-based methods can utilize 5’ rapid amplification of cDNA ends(RACE) PCR from constant region sequences, minimizing primer bias [84].Following data generation, the readouts of TCR/BCR sequencing are processed by TCR/BCRclonotype calling methods to reconstruct the T/B cell clonotype repertoire. Several pipelineshave been developed for TCR clonotype calling, including MiXCR [86], LymAnalyzer [87], andIMSEQ [88]. These methods rely on a similar framework: (1) initial alignment of input sequencereads to germline V, D, and J segments, (2) assignment of mapped sequence reads into clonesaccording to sequence identity, usually by the CDR3 region, and (3) correction of sequenceerrors by merging clones with high sequence similarity. BCR clonotype calling can be similarlyperformed, but somatic hypermutation complicates clone assignment as BCR clonotypes fromthe same clonal family may harbour CDR3 sequences with multiple nucleotide mismatches.Thus, relaxing the similarity threshold for clonal merging may be more appropriate for calling15BCR clonotypes [86]. Alternatively, a recently developed approach for BCR repertoire inferenceuses Hidden Markov Models to describe the generative process of V(D)J recombination andsomatic hypermutation [89, 90]. However, this approach is currently slow for large datasets(execution time of hours for datasets with > 105 reads).Most TCR/BCR sequencing methods target the TCRβ and BCR-heavy chains, as thesecontain D genes and thus have greater diversity than TCRα and BCR-light chains. However,T cells with the same TCRβ chain can harbour different TCRα chains. A recently developedassay from Adaptive Biotechnologies, pairSEQ, allows for paired sequencing of TCRα and βchains from the same individual cells [91]. Single cell TCR/BCR sequencing also permits directassessment of α:β pairing.1.4.3 Single cell methods1.4.3.1 Single cell DNA sequencingSingle cell DNA sequencing aims to bring the readouts of bulk genome sequencing – SNVs,CNVs, and SVs – to the cellular level. In the context of cancer genomics, single cell DNAsequencing offers distinct advantages over bulk sequencing in identifying rare clonal populationsin heterogeneous tumours and resolving phylogenetically divergent clonal mixtures to understandthe mechanistic bases of tumour progression (Figure 1.7). Analysis of single cell DNA sequencingdata has provided insights into patterns of intratumoural genomic heterogeneity in patient-derivedxenograft models [29, 92], spatial invasion of breast cancer clones [93], and chemoresistance intriple-negative breast cancer [94].16Figure 1.7: a) Depiction of single cell WGS library preparation by DLP. Single cells are isolatedin individual wells, and lysed. The resulting DNA is tagmented prior to amplification to allow forcomputational identifiation of PCR duplicates and thus accurate recovery of copy number variants.b) Single cells can be clustered into clones with similar copy number profiles; the resulting clonalconsensus copy number profiles can be used to build a phylogenetic tree. Used with permission fromhttps://www.nature.com/articles/nmeth.4140/figures/introduction/1.To date, no singular single cell DNA sequencing technology has been widely adopted acrosscancer genomics. Generally, single cell DNA sequencing technologies aim to optimize coveragedepth, breadth, uniformity, and accuracy to faithfully recapitulate single cell genotypes. However,virtually all single cell DNA sequencing technologies exhibit key tradeoffs in one or more ofthese areas. Most single cell DNA sequencing workflows can be divided into 3 steps: (1) cellisolation, (2) DNA amplification, and (3) amplicon sequencing and interpretation (Figure 1.7)[95]. Differences in how DNA amplification is performed primarily underlie the key tradeoffsassociated with single cell DNA sequencing technologies. Isothermal amplification methodssuch as multiple displacement amplification (MDA) are highly sensitive and generate highcoverage depth, allowing for detailed interrogation of SNVs and indels at the cost of coverageuniformity [95]. Consequently, isothermal methods are unsuitable for CNV assessment. Onthe other hand, degenerate oligonucleotide primed PCR (DOP-PCR) employs thermocyclingto recover amplification products with lower coverage bias but inferior depth [96]. Furtherimprovements in coverage uniformity were provided by direct library preparation (DLP), ananolitre-volume protocol carried out in a microfluidic device requiring no pre-amplification,designed for single cell WGS (Figure 1.7) [97]. DLP has been successfully employed to recoverwhole genome copy number profiles of cell lines and patient-derived xenograft samples [98].“Pseudobulk” aggregation of single cell readouts from DLP faithfully recapitulates bulk whole17genome sequencing SNV profiles at similar coverage depth [97]. DLP can distinguish clonalpopulations of cancer cells defined by distinct copy number profiles, which can be leveraged toreconstruct pseudobulk SNV, CNV, and SV profiles at the clone level (Figure 1.7). A recentlydeveloped commercial assay from 10x Genomics also recovers whole genome CNV profiles at thesingle cell level [99], but the performance of this method has yet to be rigorously validated.Single cell DNA sequencing can also be targeted to particular regions of the genome throughtarget-specific amplification or capture, enabling superior coverage depth and breadth in targetedregions at the cost of uniformity [95]. This enables long-range variant phasing, which can beemployed to validate clonal genotypes proposed by statistical deconvolution of bulk genomes[77, 92].1.4.3.2 Single cell transcriptomicsTranscriptomics provides measurements of cellular phenotype through quantification of relativeRNA abundance. Contemporary bulk transcriptome technologies such as microarrays andRNA sequencing (RNA-seq) have enabled quantitation of gene expression in tumour samples,establishing prognostically relevant transcriptomic subtypes and microenvironmental propertiesof many cancers [100–103]. However, the aggregate measurements provided by bulk transcrip-tomics are affected by cell type composition, making direct interrogation of malignant, immune,and stromal phenotypes difficult (Figure 1.8). Single cell RNA-sequencing generates wholetranscriptomes at the single cell level, enabling direct assessment of individual cellular pheno-types, tissue composition, gene regulation, and cell state evolution throughout development anddifferentiation (Figure 1.8).18Figure 1.8: When applied to heterogeneous cellular populations, single cell RNA-sequencing cansimultaneously recover single cell transcriptomes and cell type proportions in a nearly unbiasedmanner. In contrast, bulk RNA-seq recovers average expression profiles. Deconvolution methods canrecover cell type proportions, but these usually require prior information on cell type expressionprofiles [104]. Image provided by 10x Genomics from https://community.10xgenomics.com/t5/10x-Blog/Single-Cell-RNA-Seq-An-Introductory-Overview-and-Tools-for/ba-p/547.Over the past decade, single cell mRNA-seq technologies have matured from digital tran-scriptomic assessment of a single cell to droplet-based technologies capable of profiling thousandsof cells per run [105]. Most single cell RNA-seq protocols can be divided into 4 main steps: (1)initial sample preparation, (2) single cell capture, (3) nucleic acid extraction and amplification,and (4) sequencing of amplified products; primarily differing from one another in (2) and (3).Fluorescence-activated cell sorting (FACS), microdroplet and microfluidic technologies enablehigh-throughput capture of hundreds to millions of single cells, but require dissociated tissuesamples as input [105]. The enzymatic treatments used in tissue dissociation can introducephenotypic changes marked by upregulation of immediate early genes (IEGs) such as FOSand JUN [106, 107]. Laser-capture microdissection (LCM) and micropipetting enable singlecell capture from intact tissue specimens, but require manual isolation of single cells [105]. Arecently developed method, SPLiT-seq, uses combinatorial barcoding to bypass cell capturealtogether [108]. Following single cell capture and cell lysis, mRNA can be amplified by poly-Tpriming and second strand synthesis or 5’ template switching synthesis. Template switchingamplification, employed in the SmartSeq and STRT-Seq protocols, enables full-length transcriptcoverage with reduced bias while other methods suffer from 3’ bias due to incomplete reversetranscription [109]. Recently, 10x Genomics has released a droplet-based commercial platform19for single cell RNA-seq that quantifies the abundance of 3’ transcript fragments for thousandsof cells per sample [110].Imperfections in cell capture, reverse transcription, and amplification present unique technicalchallenges for single cell RNA-seq data analysis. During cell capture, multiple or dead cells maybe collected in place of a single viable cell [105]. Cell lysis introduces ambient RNA that cancontaminate libraries generated from live cells [110]. Due to Poisson sampling, many transcriptspecies may not be reverse transcribed prior to amplification, leading to transcript dropoutunremedied by increasing sequencing depth. Amplification bias can also lead to dropout forsimilar reasons [105]. Moreover, single cell RNA-seq data generated from different centers,reagent pools, or reaction chips is not directly comparable due to sizeable batch effects [111].These batch effects can overwhelm or obscure subtle phenotypic differences between similar celltypes or states.As such, proper quality control is a critical step in single cell RNA-seq analysis. Rupturedcells are first removed by identifying libraries enriched for mitochondrial transcripts, indicatingloss of cytoplasmic transcripts due to increased cell membrane permeability [112]. Doublets canbe identified experimentally by imaging in plate- or well-based protocols or cell hashing withbarcoded antibodies [113], and computationally by expression of mutually exclusive cell typemarkers. A recently developed tool, SoupX, removes signal from ambient RNA [114]. Somemodels for single cell RNA-seq data employ negative binomial distributions with zero-inflationto model transcript counts subject to dropout [115]. Alternatively, imputation approaches suchas MAGIC attempt to correct for dropout by using information from similar cells [116]. Manydifferent batch correction methods have been employed in single cell RNA-seq analysis. Batcheffect correction by linear regression [117] can improve concordance between datasets generatedfrom similar input material across different centers, but can also introduce biases when workingwith samples with different cellular composition. In these scenarios, more sophisticated batcheffect correction methods that adjust transcript expression based on shared cellular populationsacross batches can be employed [118–120].The readouts of single cell RNA-seq can be used to understand tissue composition, develop-mental trajectories, and gene networks. Single cell RNA-sequencing allows for virtually unbiasedassessment of cell types by dimensionality reduction and subsequent unsupervised clustering.This approach has been employed in various tissue types and organisms to quantify known andnovel cell types [110, 121–123]. In the cancer context, these methods have been used to profile thephenotypic subsets of immune and stromal cells in the microenvironment and their relationshipswith patient survival [124–126]. Algorithms that model continuous transitions between cellstates – pseudotime algorithms [127–130] – have been developed to delineate phenotypic changes20that occur during cell differentiation from stem cells to terminally differentiated cells [131]. Thewealth of individual measurements provided by contemporary single cell RNA-seq methodshas enabled regulatory network reconstruction at unprecedented scale [132]. The networksidentified by these algorithms could be feasibly used to improve orthogonally collected single cellRNA-seq data through imputation [133]. In summary, single cell RNA-seq enables simultaneousprofiling of both cancer cell and microenvironment phenotypes to study cancer-microenvironmentinterplay.1.5 Problem statementDespite extensive effort being made to identify new therapeutic targets for HGSC, long-termoutcomes have remained bleak for many patients. Crucially, most patients present with advancedstage disease characterized by within- and between-site heterogeneity [27, 29]. This genomicheterogeneity provides considerable substrate for selection to act on, and is thus thought to leadto the development of resistance. However, the observation that tumour-infiltrating lymphocyteabundance is associated with superior outcomes [36] hints at the tantalizing possibility thatintratumoural immune infiltration may be able to contend with genomic heterogeneity.Apart from T cells, the microenvironmental properties of HGSC and their associations withgenomic and clinical features remain poorly understood. Initial efforts have been made tounderstand the role of B cells and plasma cells in HGSC [38, 41], but their antigenic targetsand interacting partners in the HGSC microenvironment are unclear [42]. Even less is knownabout non-immune cell types such as fibroblasts and endothelial cells. Transcriptome-basedsubtyping of HGSC by independent groups identified 4 prognostically distinct subgroups largelydistinguished by immune and stromal markers [14, 101], implying that stromal cell types alsoinfluence disease progression in HGSC.In this thesis, I set out to investigate the evolutionary interplay between malignant andnon-malignant cells in HGSC. Many HGSC cases present with disseminated disease, providing anincredible opportunity to study tumour evolution across distinct peritoneal microenvironments.In Chapter 2, I describe the assembly of a collection of 148 tumours from 41 patients, the largestmulti-site HGSC cohort to date I am aware of (Figure 1.9). I co-ordinated an integrated teamof clinical and research personnel to identify and notify the team of potential HGSC cases forcollection on a weekly basis, and helped devise methods for single cell processing and datacuration from these samples. I outline the steps from clinical case identification to sampleprocessing that generated the experimental substrate used in the following chapters. In addition,these samples will be used in future work involving drug testing of patient-derived xenograft21models and single cell whole-genome sequencing. While patient-derived xenografts are poormodels for directly studying tumour-microenvironment interaction, they enable drug sensitivitytesting of clones from different microenvironmental contexts. In Chapter 3, I interrogate theinterface of lymphocytic and malignant evolutionary dynamics in HGSC, described in a Cellpublication (Figure 1.9) [1]. I was responsible for the bulk of computational work described inthis chapter, including most of the integrative analysis, clonal inference, human leukocyte antigenloss-of-heterozygosity analysis, gene expression, TCR/BCR-seq analysis, and large portions ofthe histologic image analysis. The last section of this thesis extends the work of Chapter 3 toother non-malignant cell types using single cell RNA-sequencing (Figure 1.9). I describe aprobabilistic method for identifying known cell types from single cell RNA-seq data in Chapter 4,demonstrating its utility on simulated data. Finally, I apply this approach to single cell RNA-seqdata from spatial samplings collected in Chapter 2 to comprehensively characterize the HGSCmicroenvironment. I led the work desribed in this chapter, helping formulate the model andconduct most of the analysis on simulated and real data.Chapter 2 - Accrual of (multi-site)HGSC cases Chapter 3 - Analysis of the evolutionaryinterfaces between the immunemicroenvironment and cancerclones Fresh tissue samplesPatient-derivedxenograftsFrozen andpreserved tissuesamplesImmune profilingBulk whole-genomesequencingChapter 4 - Profiling of immune and non-immune cell types in the HGSCmicroenvironment at single cellresolution Future work - Mechanisms of drug resistancerelated to microenvironmentalproperties Single cell RNA-seq Single cell WGSFigure 1.9: Outline of relationships between thesis chapters. White boxes (single cell WGS andfuture work) correspond to elements that were not performed as part of this thesis.22Chapter 2Collection and processing ofmulti-site HGSC samples forhigh-throughput sequencing, PDXcreation, and single cell experiments2.1 IntroductionThe high prevalence of multi-site disease and well-described site-to-site genomic, transcriptomic,and microenvironmental variation in HGSC necessitates study of multiple tumour foci fromthe same patient to understand disease pathogenesis. However, only a handful of groups haveattempted to conduct multi-site studies of HGSC, and these studies have been restricted to smallcohort sizes [27–29]. These initial studies have exemplified the degree of inter-site heterogeneity inHGSC and raised questions on how this heterogeneity affects prognostically relevant associationsbetween genomic features and the tumour microenvironment [29]. Systematic collection ofmulti-site HGSC cases at scale will be required to decipher the evolutionary mechanisms bywhich HGSC tumours develop treatment resistance and thwart immunologic surveillance in vivo.Patient-derived xenograft (PDX) models are laboratory mice transplanted, usually subcuta-neously or subcapsularly, with human tumour cells. Under the assumption that these modelsfaithfully recapitulate the phenotypic properties of their source tumours, PDXs serve as malleablesystems for studying tumour evolution and drug response. Most PDXs are constructed fromimmunodeficient mice to prevent transplant rejection, but newer methods enable establishmentof ‘humanized’ PDX models that contain human-like immune systems. Thus far, PDXs thatrecapitulate genomic properties of their source tumours have been established for several cancertypes including ovarian cancer [134], breast cancer [135], and B cell lymphomas [136].Despite extensive profiling of clonal heterogeneity in HGSC [27–29], the genomic and23transcriptomic properties of clones associated with treatment resistance remain unknown.Identifying the hallmarks of clones associated with treatment resistance and dissemination mayprovide critical insights into predicting response and personalizing therapeutic regimens forHGSC patients. One of the aims of this chapter will be to build PDXs for each tumour in acohort of multi-site HGSC patients in order to study tumour evolution in response to treatmentpressure. These PDXs will serve as an ideal substrate for interrogating the relative fitness ofclonal genotypes and the reproducibility of clonal dynamics between clones derived from differenttumour microenvironments in response to external selection pressure.In order to supplement the tumour cell-focused view of tumour evolution provided byPDX modeling, another aim of this chapter is to create experimental substrates and methodsfor profiling the tumour microenvironment of HGSC. To date, most studies of the HGSCmicroenvironment utilize histologic image analysis for cell type quantification or bulk geneexpression profiling for phenotypic analysis [14, 36, 38, 41, 101]. However, routine histologicimage analysis and immunohistochemistry can only capture a limited number of cell types, anddeconvolving cell type proportions and transcriptomes from bulk gene expression profiles isdifficult. Single cell RNA sequencing, with technologies such as SMART-Seq [109], Drop-Seq[121], and 10x Chromium [110], enables simultaneous capture of cell type abundances andtranscriptomes, but its use for studying solid tumours, especially ovarian cancers, is limited.Most single cell RNA-seq experiments thus far have utilized material from peripheral bloodor mouse models, which yield high quality data due to the minimal extent of manipulationrequired to obtain viable single cell suspensions. In the context of profiling gross HGSC tissuespecimens, methods for preparing single cell suspensions and libraries must be optimized tominimize technical effects on microenvironmental composition and phenotypes [106, 107].With the goal of profiling the pre-treatment microenvironment and clonal dynamics of HGSCin response to treatment, we systematically collected a cohort of multi-site HGSC cases. In thischapter, I outline the process of sample accrual, from case identification to sample processing andPDX construction, that served as the basis for the work described in Chapter 3 and Chapter 4.2.2 Materials and Methods2.2.1 Summary of accrual processSurgeons and senior surgical residents identified potential HGSC cases at local hospitals, includingVancouver General Hospital (VGH) and the University of British Columbia (UBC) Hospital.Consents and peripheral blood were obtained for each patient prior to surgery, and cases were24prioritized if possible as first or second on surgical slates.Surgeons sent debulking specimens for initial processing by hospital research assistants andmedical technologists. Following initial pathologic assessment to confirm diagnosis of HGSC,specimens were transferred to the BC Cancer Research Center (BCCRC) for further laboratorywork involving preparation and preservation of material for bulk sequencing, PDX establishment,and single cell RNA sequencing. Final pathologic assessment was performed by a trainedpathologist at VGH and non-HGSC cases were retroactively removed from the study.The overall accrual pipeline is shown in Figure 2.1. Each of these steps are describedfurther in the sections below.25Case identification - HGSC - multi-site (for PDX) Surgery - booking: 1st/2nd onslate Clinical tissueprocessing - initial pathologicalevaluation Research tissueprocessing - mechanical dissociation - aliquoting Single cell RNA-sequencing workflowPatient-derivedxenograft creationBulk whole-genomesequencingBulk RNA experiments - transcriptome - TCR/BCR-seq Single cell whole-genome sequencing Viably frozenFlash frozenFreshData curation - Storage of caseinformation in database BC CancerVGH/UBCClinical tissue banking Peripheral bloodcollection FormalinRemainderHistologyFigure 2.1: Clinical and research pipelines for processing high-grade serous ovarian cancer cases.Steps in red are executed at the hospital (either Vancouver General Hospital or the University ofBritish Columbia Hospital) by clinical personnel; steps in red are carried out at BC Cancer byresearch personnel.262.2.2 Patient cohortPre-operatively, patients were screened based on the following criteria: (1) clinical suspicionof HGSC based on history, imaging and blood work, (2) no prior treatment, i.e. chemo- orradiotherapy, and (3) patient consent. All cases – except for those dedicated for single cellRNA-seq pilot experiments – were additionally required to have at least 2 tumour foci fromanatomically distinct masses or distal regions of a single mass that could be collected. Recurrencespecimens were obtained from patients that presented for a subsequent operation related totheir disease.2.2.3 Collection of surgical specimens and peripheral bloodPatient consent was obtained prior to specimen collection and banking and documented atVGH or UBC Hospital Laboratory (Research Ethics Board numbers H08-01411 and H18-01090).Specimens of consented patients were placed into cold media and brought to the clinicallaboratory by the messenger porter. Following this, each specimen was assigned a uniqueresearch identifier and processed as per VGH/UBC Anatomical Pathology specimen handlingprocedures (Figure 2.1).Each case was initially assessed to determine whether or not the disease was HGSC andif sufficient material was available for research purposes. Specimens for cases where sufficientmaterial was available from multiple tumour foci (or a single tumour focus for single cell RNA-seqexperiments) were considered eligible. For cases with multiple sites, each site was sent outindividually upon collection to minimize delay to sample processing.Peripheral blood was separately collected in purple/pink top (plasma and buffy coat) andred top (serum) tubes (Figure 2.1). Blood components were spun down and transferred intolabelled cryovials, snap frozen in liquid nitrogen and stored in the -80◦C freezer.2.2.4 Sample preparationEach specimen was assigned a unique anonymous research identifier linked to a case identifier(Section 2.2.12). Specimens were placed in a Petri dish and measured. One millilitre cryovialscorresponding to each aliquot type to be created (formalin, flash frozen, transplant, viablefrozen, and remainder) were prepared (Figure 2.1). A 1mm piece was first cut and placed inthe formalin cryovial containing 1mL formalin. The remaining pieces were chopped finely on acell culture dish and used to create the remaining aliquots.A small portion of the finely chopped tissue was aliquoted into a stomacher bag containing1mL of media, while the remaining tissue was set on ice for single cell dissociation (and single27cell RNA-seq). The stomacher machine was run for 60 seconds at normal settings to furtherdissociate the sample. One hundred microlitres of the supernatant was added to the transplantvial, with the rest aliquoted to the viable frozen vial. The remaining chunks of tissue in thestomacher bag were added to the remainder vial. Following this, the transplant and viable frozenvials were spun down and 1mL of freezing media was added to each vial. Transplant, viablefrozen, and remainder vials were placed in a Mr. Frosty freezing container (Thermo Scientific)to be gradually frozen overnight, and transferred to the -80◦C freezer the next day. Flash frozenvials were placed into a cryobox and stored directly in the -80◦C freezer. Formalin vials weresent for embedding and hematoxylin and eosin (H&E) staining.2.2.5 Patient-derived xenograft creationFor each specimen, the aliquot set aside for xenografting was used for PDX construction. Whenpossible, PDXs were created from freshly processed aliquots; otherwise, aliquots set aside forxenografting were viably frozen for transplantation at a later date.Each transplantation vial was divided into 4 equally-sized aliquots of approximately 250microlitres each in Eppendorf tubes. Aliquots were spun down for 5 minutes at 1200rpm. Pelletswere resuspended in 200 microlitres of 50% Matrigel and kept on ice until transplantation.Each aliquot was subcutaneously injected into a NOD.Cg-PrkdcscidIl2rgtm1Wjl/SzJ (Nod-Scid-gamma, NSG) or NOD.Cg-Rag1tm1MomIl2rgtm1Wjl/SzJ (Nod-Rag-gamma, NRG) mouseaged 5-12 weeks with a 21-gauge needle. Mice were placed in cages (up to 4 mice per cage,identified by ear punching) and monitored weekly initially and more frequently as humane orexperimental endpoints were reached. Following euthanasia, mice were biopsied and tumourswere collected and frozen.2.2.6 Whole-genome sequencing of patient tumoursDNA was extracted from flash frozen tumour sample aliquots using the Qiagen Blood and TissueExtraction Kit. DNA samples were submitted for sequencing at the BC Genome Sciences Centre(BCGSC). For all tumor and corresponding normal (blood) samples, sequencing was performedusing Illumina HiSeq2500 whole genome shotgun v4 chemistry with paired-end 125bp reads.Samples were sequenced to an average of 96X coverage [1].2.2.7 Single cell RNA-seq pilot projectFor all suspected HGSC cases (single and multi-site), a portion of each specimen was set asidefor single cell suspension creation and subsequent library preparation with the 10x Genomics283’ or 5’ gene expression kits [110]. Various sample dissociation times, enzyme mixture, andviability assessment methods were piloted (Table 2.2). The workflow is shown in Figure 2.2.29Fresh tissue sample - weighing gentleMACS 37Cdissociation - Miltenyi enzyme mixes gentleMACS 6Cdissociation - Bacillus lichenformisprotease Post-dissociation wash - viability assessment Viability enrichment - Miltenyi Dead CellRemoval Kit 10x Chromium librarypreparation Sequencing - NextSeq 500 - R1 26bp, R2 58bp CellRanger Viability >= 75%?NoYesQuality control, analysis - Dimensionality reduction - Batch correction - CellAssign Figure 2.2: Single cell RNA-sequencing workflow from fresh tissue samples to analysis.302.2.8 Single cell dissociationFor each specimen, the portion set aside for single cell RNA-seq was used to prepare a singlecell suspension.2.2.8.1 37C proteaseAfter weighing in a cell culture dish, tissue was transferred into a gentleMACS C tube andpipetted up and down using a wide bore pipette tip. GentleMACS programs h tumour 01,h tumour 02, and h tumour 03 were run, with samples incubated for 30 minutes at 37◦C undercontinuous rotation using the MACSmix Tube Rotator between programs. Miltenyi Biotecenzymes H (200µl), R (100µl) and A (25µl) were used for dissociation. Following dissociation,cells were assessed for viability using the cell counter (5µl cells + 5µl trypan blue) under amicroscope. Cells were then pelleted by centrifugation for 5 minutes at 4◦C, resuspended infreezing media, placed in Mr. Frosty overnight, and frozen at -80◦C.2.2.8.2 6C proteaseAfter weighing in a cell culture dish, tissue was transferred into a gentleMACS C tube, and onemillilitre of 10 mg/mL Bacillus lichenformis protease (Creative Enzymes NATE-0633; henceforthreferred to as 6C protease) was added to each 25 mg of tissue. The resulting solution wasincubated and mechanically disrupted at 6◦C. Depending on the sample, two different protocolswere used for mechanical disruption. The first protocol involved pipetting up and down for 15seconds every minute for a total of 15 minutes. The second protocol utilized the Miltenyi BiotecMACS Separator (programs h tumour 01, h tumour 02, h tumour 03) with the 6C proteasefor 30 minutes or 1 hour. Dissociation specifications for each sample are listed in Table 2.1.Following dissociation, cells were assessed for viability using the cell counter (5µl cells + 5µltrypan blue) under a microscope. Cells were then pelleted by centrifugation for 5 minutes at4◦C, resuspended in freezing media, placed in Mr. Frosty overnight, and frozen at -80◦C.Patient Sample Anatomic site Digestion 10x Method62 VOA11019SA-37RLQ sitemetastasis37C collagenase 4h 3’ GE62 VOA11019SA-CD3RLQ sitemetastasis37C collagenase 4h 3’ GE62 VOA11019SA-CD45RLQ sitemetastasis37C collagenase 4h 3’ GE3162 VOA11019SA-6RLQ sitemetastasiso/n 6C 3’ GE63 VOA11095SAPosterior Cul deSaco/n 6C63 VOA11095SB Splenic Stricture o/n 6C63 VOA11095SC Left Ovary o/n 6C63 VOA11095SD Epiploica Sigmoid o/n 6C63 VOA11095SE Right Ovary o/n 6C63 VOA11095SF Omentum o/n 6C64 VOA11213SA Ovary MACS 37C 1h 5’ GE64 VOA11213SB Omentum MACS 37C 1h 5’ GE64 VOA11213SC Bowel MACS 37C 1h 5’ GE64 VOA11213SC Bowel 37C collagenase 1h65 VOA11083A Pelvic o/n 6C65 VOA11083B Pelvic o/n 6C65 VOA11083C Right Ovary o/n 6C65 VOA11083D Omentum o/n 6C 3’ GE65 VOA11083E Cecum o/n 6C66 VOA11088ALeft FallopianTubeo/n 6C66 VOA11088B Omentum o/n 6C67 VOA11220SA Right Ovary MACS 37C 30min 5’ GE67 VOA11220SB Gastric Nodule MACS 37C 30min67 VOA11220SC Omental Nodule MACS 37C 30min 5’ GE67 VOA11220SD Rectal Sigmoid MACS 37C 30min 5’ GE68 VOA11243SA Uterus Surface 6C 1hr68 VOA11243SB Right Ovary 6C 1hr68 VOA11243SC Right Tube 6C 1hr68 VOA11243SD Left Ovary 6C 1hr68 VOA11243SE Omentum 6C 1hr68 VOA11243SF Pouch of Douglas 6C 1hr68 VOA11243SA Uterus Surface 6C O/N3268 VOA11243SB Right Ovary 6C O/N68 VOA11243SC Right Tube 6C O/N68 VOA11243SD Left Ovary 6C O/N68 VOA11243SE Omentum 6C O/N68 VOA11243SF Pouch of Douglas 6C O/N69 VOA11265SA Omentum MACS 6C 1hr69 VOA11265SBLeft fallopiantube noduleMACS 6C 1hr70 VOA11267-6 Left adnexal mass MACS 6C 1hr 5’ and 3’ GE70 VOA11267-37 Left adnexal mass MACS 37C 1hr 5’ and 3’ GE71 VOA11294SA Left Ovary MACS 6C 1hr71 VOA11294SA Left Ovary MACS 37C 1hr71 VOA11294SBSmall BowelTumourMACS 6C 1hr71 VOA11294SC Right Ovary MACS 6C 1hr71 VOA11294SC Right Ovary MACS 37C 1hr71 VOA11294SDLeft OvarianTumourMACS 6C 1hr71 VOA11294SDLeft OvarianTumourMACS 37C 1hr71 VOA11294SESurface UterineTumourMACS 6C 1hr71 VOA11294SF Omentum MACS 6C 1hr71 VOA11294SF Omentum MACS 37C 1hr72 VOA11558SA Omentum MACS 6C 1hr73 VOA11543SA Left Ovary MACS 6C 1hr 5’ GE73 VOA11543SB Right Ovary MACS 6C 1hr 5’ GEC2 VOA11229 Right Ovary MACS 6C 30minsC2 VOA11229 Right Ovary MACS 6C 1 hourE2 VOA11520SA Right Ovary MACS 6C 1hrE2 VOA11520SA Right Ovary MACS 37C 1hr33Table 2.1: Identifiers of samples used for 10x Chromium library preparation and/or sequencing.The tissue dissociation protocols and types of 10x Chromium library preparation are listed for eachsample.2.2.8.3 Post-dissociation washEnzymatically dissociated samples were thawed, spun down, and washed with 1mL PBS twiceto remove the dimethyl sulfoxide (DMSO) present in freezing media. Samples were then dilutedwith cold HFN and washed with trypsin, dispase, and DNAse while gently pipetting up anddown. Cold ammonium chloride was added to bloody samples. Cells were assessed for viabilityusing the cell counter (5µl cells + 5µl trypan blue) under a microscope, and kept on ice.2.2.9 Viability sorting and assessmentViability sorting was performed for samples with <75% viability after post-dissociation wash,with a target viability of ≥75% viability (Figure 2.2). Cells were spun down and the pelletresuspended in 100µl of Miltenyi Dead Cell Removal MicroBeads and incubated at roomtemperature for 15 minutes. Viable cell enrichment was performed using the positive selectioncolumn type MS with a MACS Separator. Cells were then placed on ice for 10X GenomicsscRNAseq library preparation.2.2.10 Single cell RNA-seq library preparation and quality controlSingle cell RNA-seq libraries were prepared following the 10x Genomics protocols for 3’ or 5’gene expression library construction [110]. The concentration and amount of cells and reagentsadded corresponded to the protocol requirements for obtaining 3000 cells with data [137].2.2.11 Sequencing of single cell RNA-seq librariesSequencing of 10x Genomics 3’ single cell RNA-seq or 5’ single cell RNA-seq libraries wasperformed on an Illumina NextSeq 500 at high throughput with 75bp paired-end reads at theUBC Biomedical Research Centre (sequencing the terminal 58bp of R2). The target sequencingdepth for each sample was 50,000 read pairs per cell, as recommended by 10x Genomics [137].However, as fewer cells (than targeted during library preparation) were recovered for severalsamples, the actual sequencing depth per cell varied from 50,000 to 1,000,000 read pairs/cell.342.2.12 Data curationEach patient enrolled in the study was assigned a unique anonymous research identifier. Samplesfrom the same patient were assigned unique identifiers associated with a common patientidentifier. To preserve patient privacy, the corresponding confidential patient identifiers, patientconsents, and clinical data (survival status, treatment, age, name, etc.) were stored in a clinicaldatabase inaccessible to research personnel.A structured query language (SQL) database was used to store sample information (collectiondate, anatomic site of collection, matched normal available) associated with each researchidentifier. Additionally, PDX model information associated with each specimen (transplantationdate, transplantation site, mouse model type, PDX identifier, mouse date-of-birth, passagenumber, euthanasia/termination date, necropsy findings, and tumour size) was stored in theSQL database. Physical sample locations were tracked in an OpenSpecimen database.2.3 Results2.3.1 Accrual of 41 HGSC casesForty-eight ovarian cancer cases – 8 single-site and 40 multi-site – were accrued from May 2015to September 2018 (Table 2.2). On confirmatory pathologic assessment, 41 were HGSC, 2 clearcell ovarian cancer, 2 endometroid, 1 serous borderline, 1 carcinoma with sarcoma elements, and1 Krukenberg (Table 2.2). In total, 148 HGSC samples have been collected from these cases(Figure 2.3). Whole-genome sequencing has been performed on 56 samples from 14 of theseHGSC cases (Table 2.3). Library construction for single cell RNA-seq has been performedon 43 samples from 17 HGSC cases; of these, 11 samples from 6 patients have been sequenced(Table 2.1).Patient Histotype # samples # transplanted24 HGSC 1 125 HGSC 4 426 HGSC 1 128 HGSC 3 329 HGSC 7 630 HGSC 7 731 HGSC 9 932 HGSC 6 63537 HGSC 4 438 HGSC 3 341 HGSC 4 442 HGSC 5 543 HGSC 5 044 HGSC 4 445 HGSC 3 346 HGSC 4 447 HGSC 5 548 HGSC 4 349 HGSC 4 450 HGSC 4 451 HGSC 3 352 HGSC 4 453 HGSC 3 356 HGSC 1 057 HGSC 4 359 HGSC 2 260 HGSC 2 261 HGSC 1 062 HGSC 1 063 HGSC 6 664 HGSC 3 365 HGSC 5 066 HGSC 2 267 HGSC 4 468 HGSC 6 669 HGSC 2 270 HGSC 1 071 HGSC 6 672 HGSC 1 073 HGSC 2 074 HGSC 2 236B1 Serous borderline 4 4C2 CCOC 1 0CCOC1 CCOC 4 4CS1 Carcinoma w/ sarcoma-like 4 4E1 Endometroid 5 5E2 Endometroid 2 2K1 Krukenberg 2 0Table 2.2: Patient identifiers, histotype as determined by final pathologic evaluation, and thenumber of tumour samples collected per case. The number of tumour samples transplanted (tocreate PDX models) is also shown.0501000 300 600 900 1200Days since start of accrualNumber of HGSC samplesFigure 2.3: HGSC sample accrual since the beginning of the study (first collected sample).37Patient Sample Anatomic site25 VOA6428AX omentum site 125 VOA6428BX right ovary site 125 VOA6428CX right ovary site 225 VOA6428DX right ovary site 326 VOA6491X left ovary site 128 VOA7640CX left fallopian tube distal site 128 VOA7640AX omentum site 128 VOA7640BX left fallopian tube proximal site 129 VOA7648DX omentum site 129 VOA7648CX ascites site 129 VOA7648BX cul-de-sac site 129 VOA7648EX sigmoid colon site 129 VOA7648FX round ligament site 129 VOA7648AX diaphragm site 130 VOA7652EX omentum site 130 VOA7652GX left pelvic wall site 130 VOA7652DX right fallopian tube site 130 VOA7652BX left ovary site 130 VOA7652FX sigmoid site 130 VOA7652AX right ovary site 130 VOA7652CX left fallopian tube site 131 VOA7668EX ileal tumour site 131 VOA7668JX left fallopian site 131 VOA7668AX pelvic sidewall site 131 VOA7668BX right ovary site 131 VOA7668DX small bowel serosa site 131 VOA7668HX uterine surface site 131 VOA7668CX right fallopian tube site 131 VOA7668GX left ovary site 131 VOA7668FX anterior cul-de-sac site 132 VOA7685FX omentum site 132 VOA7685AX left ovary site 132 VOA7685EX right ovary site 13832 VOA7685DX omental nodule site 132 VOA7685BX left ovary site 232 VOA7685CX left ovary site 337 VOA8841ax right ovary site 137 VOA8841bx retrosigmoid serosa site 137 VOA8841cx omentum site 138 VOA9127ax right ovary site 138 VOA9127bx left ovary site 141 VOA9465ax cul de sac site 141 VOA9465bx uterine serosa site 141 VOA9465cx omentum site 141 VOA9465dx left fallopian tube site 143 VOA7255ax right ovary site 143 VOA7255bx left ovary site 143 VOA7255cx right fallopian tube site 143 VOA7255ex peritoneal nodule site 144 VOA9655ax right ovary site 144 VOA9655bx left ovary site 144 VOA9655cx left ovarian cyst site 146 VOA9921ax left ovary site 146 VOA9921bx omentum site 146 VOA9921cx left fallopian tube site 147 VOA9955cx right ovary site 1Table 2.3: Sample and patient identifiers of HGSC samples used for whole genome sequencing.2.3.2 Construction of patient-derived xenograft modelsThus far, 128 samples from 38 HGSC cases have been engrafted in PDXs (total of 275 PDXs,see Table 2.4 for a full list and Figure 2.4 for established PDXs). Up to 4 models per passagewere created from each primary tumour. In total, 52 samples from 19 patients have grownmacroscopically visible tumours. Engraftment (growth/establishment) rates of HGSC tumoursat the model, sample, and patient level for NSG vs. NRG strains are shown in Table 2.5.Engraftment rates for NRG mice were lower than those for NSG mice (Table 2.5). Engraftmentrates as a function of time since transplant are shown in Figure 2.5. Approximately 70% of39models that eventually engrafted had already established a mass by 300 days (Figure 2.5A).Out of all models that were euthanized, approximately 50% had grown tumours after 100 dayssince surgery (Figure 2.5B). The rates shown in Figure 2.5B appear to decrease over timebecause the humane endpoint for most models that do not grow tends to occur later than forthose that do (Figure 2.6); models that do not grow are euthanized based on health/age, whilethose that do are euthanized based on health/tumour size.Patient Sample PDX ID Strain Grown24 VOA5576 Y55761 NSG 024 VOA5576 Y55762 NSG 124 VOA5576 Y55763 NSG 124 VOA5576 Y55764 NSG 124 VOA5576 Y557631 NSG 124 VOA5576 Y557632 NSG 124 VOA5576 Y557633 NSG 024 VOA5576 Y557634 NSG 125 VOA6428A Y6428A1 NSG 125 VOA6428A Y6428A2 NSG 025 VOA6428A Y6428A3 NSG 125 VOA6428B Y6428B1 NSG 025 VOA6428B Y6428B2 NSG 025 VOA6428B Y6428B3 NSG 125 VOA6428C Y6428C1 NSG 125 VOA6428C Y6428C2 NSG 025 VOA6428C Y6428C3 NSG 125 VOA6428D Y6428D1 NSG 125 VOA6428D Y6428D2 NSG 025 VOA6428B Y6428B31 NSG 125 VOA6428B Y6428B32 NSG 125 VOA6428B Y6428B33 NSG 125 VOA6428B Y6428B34 NSG 125 VOA6428B Y6428B341 NSG 125 VOA6428B Y6428B342 NSG 025 VOA6428B Y6428B343 NSG 125 VOA6428B Y6428B344 NSG 14026 VOA6491 Y64911 NSG 126 VOA6491 Y64912 NSG 126 VOA6491 Y64913 NSG 126 VOA6491 Y64914 NSG 128 VOA7640A Y7640A1 NSG 128 VOA7640A Y7640A2 NSG 128 VOA7640B Y7640B1 NSG 028 VOA7640B Y7640B2 NSG 028 VOA7640C Y7640C1 NSG 128 VOA7640C Y7640C2 NSG 129 VOA7648A Y7648A1 NSG 129 VOA7648A Y7648A2 NSG 129 VOA7648D Y7648D1 NSG 129 VOA7648D Y7648D2 NSG 129 VOA7648B Y7648B1 NSG 129 VOA7648B Y7648B2 NSG 129 VOA7648E Y7648E1 NSG 129 VOA7648E Y7648E2 NSG 129 VOA7648C Y7648C1 NSG 129 VOA7648C Y7648C2 NSG 129 VOA7648F Y7648F1 NSG 129 VOA7648F Y7648F2 NSG 130 VOA7652A Y7652A1 NSG 130 VOA7652A Y7652A2 NSG 130 VOA7652B Y7652B1 NSG 130 VOA7652B Y7652B2 NSG 130 VOA7652C Y7652C1 NSG 030 VOA7652C Y7652C2 NSG 130 VOA7652D Y7652D1 NSG 030 VOA7652D Y7652D2 NSG 130 VOA7652E Y7652E1 NSG 030 VOA7652E Y7652E2 NSG 030 VOA7652F Y7652F1 NSG 04130 VOA7652F Y7652F2 NSG 030 VOA7652G Y7652G1 NSG 130 VOA7652G Y7652G2 NSG 031 VOA7668A Y7668A1 NSG 131 VOA7668A Y7668A2 NSG 131 VOA7668B Y7668B1 NSG 131 VOA7668B Y7668B2 NSG 131 VOA7668C Y7668C1 NSG 131 VOA7668C Y7668C2 NSG 131 VOA7668D Y7668D1 NSG 131 VOA7668D Y7668D2 NSG 131 VOA7668E Y7668E1 NSG 131 VOA7668E Y7668E2 NSG 131 VOA7668F Y7668F1 NSG 131 VOA7668F Y7668F2 NSG 131 VOA7668G Y7668G1 NSG 031 VOA7668G Y7668G2 NSG 031 VOA7668H Y7668H1 NSG 031 VOA7668H Y7668H2 NSG 131 VOA7668J Y7668J1 NSG 131 VOA7668J Y7668J2 NSG 132 VOA7685A Y7685A1 NSG 032 VOA7685A Y7685A2 NSG 032 VOA7685B Y7685B1 NSG 132 VOA7685B Y7685B2 NSG 132 VOA7685C Y7685C1 NSG 032 VOA7685C Y7685C2 NSG 032 VOA7685D Y7685D1 NSG 132 VOA7685D Y7685D2 NSG 132 VOA7685E Y7685E1 NSG 132 VOA7685E Y7685E2 NSG 132 VOA7685F Y7685F1 NSG 132 VOA7685F Y7685F2 NSG 14237 VOA8841A Y8841A1 NSG 137 VOA8841A Y8841A2 NSG 037 VOA8841B Y8841B1 NSG 037 VOA8841B Y8841B2 NSG 037 VOA8841C Y8841C1 NSG 137 VOA8841C Y8841C2 NSG 037 VOA8841D Y8841D1 NSG 037 VOA8841D Y8841D2 NSG 038 VOA9127A Y9127A1 NSG 038 VOA9127A Y9127A2 NSG 038 VOA9127B Y9127B1 NSG 038 VOA9127B Y9127B2 NSG 038 VOA9127C Y9127C1 NSG 038 VOA9127C Y9127C2 NSG 041 VOA9465A Y9465A1 NRG 041 VOA9465A Y9465A2 NRG 141 VOA9465B Y9465B1 NRG 041 VOA9465B Y9465B2 NRG 041 VOA9465C Y9465C1 NRG 141 VOA9465C Y9465C2 NRG 141 VOA9465D Y9465D1 NRG 041 VOA9465D Y9465D2 NRG 042 VOA10243SA Y10243SA1 NSG 142 VOA10243SA Y10243SA2 NSG 142 VOA10243SB Y10243SB1 NSG 142 VOA10243SB Y10243SB2 NSG 142 VOA10243SD Y10243SD1 NSG 142 VOA10243SD Y10243SD2 NSG 142 VOA10243SC Y10243SC1 NSG 042 VOA10243SC Y10243SC2 NSG 142 VOA10243SE Y10243SE1 NRG 142 VOA10243SE Y10243SE2 NRG 144 VOA9655A Y9655A1 NRG 04344 VOA9655A Y9655A2 NRG 044 VOA9655B Y9655B1 NRG 044 VOA9655B Y9655B2 NRG 044 VOA9655C Y9655C1 NSG 044 VOA9655C Y9655C2 NSG 044 VOA9655D Y9655D1 NSG 044 VOA9655D Y9655D2 NSG 045 VOA9907A Y9907A1 NSG 145 VOA9907A Y9907A2 NSG 145 VOA9907B Y9907B1 NSG 045 VOA9907B Y9907B2 NSG 145 VOA9907C Y9907C1 NSG 045 VOA9907C Y9907C2 NSG 046 VOA9921A Y9921A1 NSG 146 VOA9921A Y9921A2 NSG 146 VOA9921B Y9921B1 NSG 046 VOA9921B Y9921B2 NSG 146 VOA9921C Y9921C1 NSG 146 VOA9921C Y9921C2 NSG 146 VOA9921D Y9921D1 NSG 046 VOA9921D Y9921D2 NSG 047 VOA9955A Y9955A1 NSG 047 VOA9955A Y9955A2 NSG 047 VOA9955B Y9955B1 NSG 047 VOA9955B Y9955B2 NSG 047 VOA9955C Y9955C1 NSG 047 VOA9955C Y9955C2 NSG 047 VOA9955D Y9955D1 NSG 047 VOA9955D Y9955D2 NSG 047 VOA9955E Y9955E1 NSG 147 VOA9955E Y9955E2 NSG 048 VOA7294A Y7294A1 NSG 048 VOA7294A Y7294A2 NSG 04448 VOA7294B Y7294B1 NSG 048 VOA7294B Y7294B2 NSG 048 VOA7294C Y7294C1 NSG 048 VOA7294C Y7294C2 NSG 049 VOA9186A Y9186A1 NSG 049 VOA9186A Y9186A2 NSG 049 VOA9186B Y9186B1 NSG 049 VOA9186B Y9186B2 NSG 049 VOA9186C Y9186C1 NSG 149 VOA9186C Y9186C2 NSG 049 VOA9186D Y9186D1 NSG 049 VOA9186D Y9186D2 NSG 050 VOA9453a Y9453a1 NRG 150 VOA9453a Y9453a2 NRG 150 VOA9453b Y9453b1 NRG 050 VOA9453b Y9453b2 NRG 050 VOA9453c Y9453c1 NRG 050 VOA9453c Y9453c2 NRG 150 VOA9453d Y9453d1 NRG 050 VOA9453d Y9453d2 NRG 151 VOA10288SA Y10288SA1 NRG 051 VOA10288SA Y10288SA2 NRG 051 VOA10288SB Y10288SB1 NRG 151 VOA10288SB Y10288SB2 NRG 151 VOA10288SC Y10288SC1 NRG 051 VOA10288SC Y10288SC2 NRG 052 VOA10429SA Y10429SA1 NRG 052 VOA10429SA Y10429SA2 NRG 052 VOA10429SB Y10429SB1 NRG 052 VOA10429SB Y10429SB2 NRG 052 VOA10429SC Y10429SC1 NRG 052 VOA10429SC Y10429SC2 NRG 052 VOA10429SD Y10429SD1 NRG 04552 VOA10429SD Y10429SD2 NRG 053 VOA10471SA Y10471SA1 NRG 053 VOA10471SA Y10471SA2 NRG 053 VOA10471SB Y10471SB1 NRG 053 VOA10471SB Y10471SB2 NRG 053 VOA10471SC Y10471SC1 NRG 153 VOA10471SC Y10471SC2 NRG 157 VOA10863SB Y10863SB1 NRG 057 VOA10863SB Y10863SB2 NRG 057 VOA10863SC1 Y10863SC11 NRG 057 VOA10863SC1 Y10863SC12 NRG 057 VOA10863SC2 Y10863SC21 NRG 057 VOA10863SC2 Y10863SC22 NRG 059 VOA10439SA Y10439SA1 NRG 059 VOA10439SA Y10439SA2 NRG 059 VOA10439SB Y10439SB1 NRG 059 VOA10439SB Y10439SB2 NRG 060 VOA10497SA Y10497SA1 NRG 060 VOA10497SA Y10497SA2 NRG 060 VOA10497SB Y10497SB1 NRG 060 VOA10497SB Y10497SB2 NRG 063 VOA11095SA Y11095SA1 NSG 063 VOA11095SA Y11095SA2 NSG 063 VOA11095SB Y11095SB1 NSG 063 VOA11095SB Y11095SB2 NSG 063 VOA11095SC Y11095SC1 NSG 063 VOA11095SC Y11095SC2 NSG 063 VOA11095SD Y11095SD1 NSG 063 VOA11095SD Y11095SD2 NSG 063 VOA11095SE Y11095SE1 NRG 063 VOA11095SE Y11095SE2 NRG 063 VOA11095SF Y11095SF1 NRG 063 VOA11095SF Y11095SF2 NRG 04664 VOA11213SA Y11213SA1 NRG 064 VOA11213SA Y11213SA2 NRG 064 VOA11213SB Y11213SB1 NRG 064 VOA11213SB Y11213SB2 NRG 064 VOA11213SC Y11213SC1 NRG 064 VOA11213SC Y11213SC2 NRG 066 VOA11088A Y11088A1 NRG 066 VOA11088A Y11088A2 NRG 066 VOA11088B Y11088B1 NRG 066 VOA11088B Y11088B2 NRG 067 VOA11220SA Y11220SA1 NRG 067 VOA11220SA Y11220SA2 NRG 067 VOA11220SB Y11220SB1 NRG 067 VOA11220SB Y11220SB2 NRG 067 VOA11220SC Y11220SC1 NSG 067 VOA11220SC Y11220SC2 NSG 067 VOA11220SD Y11220SD1 NSG 067 VOA11220SD Y11220SD2 NSG 068 VOA11243SA Y11243SA1 NRG 068 VOA11243SA Y11243SA2 NRG 068 VOA11243SB Y11243SB1 NRG 068 VOA11243SB Y11243SB2 NRG 068 VOA11243SC Y11243SC1 NSG 068 VOA11243SC Y11243SC2 NSG 068 VOA11243SD Y11243SD1 NSG 068 VOA11243SD Y11243SD2 NSG 068 VOA11243SE Y11243SE1 NRG 068 VOA11243SE Y11243SE2 NRG 068 VOA11243SF Y11243SF1 NRG 068 VOA11243SF Y11243SF2 NRG 069 VOA11265SA Y11265SA1 NRG 069 VOA11265SA Y11265SA2 NRG 069 VOA11265SB Y11265SB1 NRG 04769 VOA11265SB Y11265SB2 NRG 071 VOA11294A Y11294A1 NRG 071 VOA11294A Y11294A2 NRG 071 VOA11294B Y11294B1 NRG 071 VOA11294B Y11294B2 NRG 071 VOA11294C Y11294C1 NRG 071 VOA11294C Y11294C2 NRG 071 VOA11294D Y11294D1 NRG 071 VOA11294D Y11294D2 NRG 071 VOA11294E Y11294E1 NRG 071 VOA11294E Y11294E2 NRG 071 VOA11294F Y11294F1 NRG 071 VOA11294F Y11294F2 NRG 074 VOA11258SA Y11258SA1 NRG 074 VOA11258SA Y11258SA2 NRG 074 VOA11258SB Y11258SB1 NRG 074 VOA11258SB Y11258SB2 NRG 0B1 VOA7618A Y7618A1 NSG 0B1 VOA7618A Y7618A2 NSG 0B1 VOA7618B Y7618B1 NSG 0B1 VOA7618B Y7618B2 NSG 0B1 VOA7618C Y7618C1 NSG 0B1 VOA7618C Y7618C2 NSG 1B1 VOA7618D Y7618D1 NSG 0B1 VOA7618D Y7618D2 NSG 1CCOC1 VOA6851A Y6851A1 NSG 0CCOC1 VOA6851A Y6851A2 NSG 0CCOC1 VOA6851B Y6851B1 NSG 0CCOC1 VOA6851B Y6851B2 NSG 0CCOC1 VOA6851C Y6851C1 NSG 0CCOC1 VOA6851C Y6851C2 NSG 0CCOC1 VOA6851D Y6851D1 NSG 0CCOC1 VOA6851D Y6851D2 NSG 048CS1 VOA6873A Y6873A1 NSG 0CS1 VOA6873A Y6873A2 NSG 0CS1 VOA6873B Y6873B1 NSG 0CS1 VOA6873B Y6873B2 NSG 0CS1 VOA6873C Y6873C1 NSG 0CS1 VOA6873C Y6873C2 NSG 0CS1 VOA6873D Y6873D1 NSG 1CS1 VOA6873D Y6873D2 NSG 1E1 VOA7298A Y7298A1 NSG 0E1 VOA7298A Y7298A2 NSG 0E1 VOA7298B Y7298B1 NSG 0E1 VOA7298B Y7298B2 NSG 0E1 VOA7298C Y7298C1 NSG 0E1 VOA7298C Y7298C2 NSG 0E1 VOA7298D Y7298D1 NSG 0E1 VOA7298D Y7298D2 NSG 0E1 VOA7298E Y7298E1 NSG 0E1 VOA7298E Y7298E2 NSG 0E2 VOA11520SA Y11520SA1 NSG 0E2 VOA11520SA Y11520SA2 NSG 0E2 VOA11520SB Y11520SB1 NSG 0E2 VOA11520SB Y11520SB2 NSG 0Table 2.4: Inventory of ovarian PDXs created from patient primary tumours. The strain of mouseused (NSG or NRG) and whether or not a macroscopically visible tumour was grown and harvestedfrom each model (1 = grown, 0 = not grown) are indicated.4902550751000 250 500 750 1000Days since start of accrualNumber of established HGSC PDXsFigure 2.4: Accrual of established PDXs since the beginning of the study (first collected sample).NRG NSGModels 102 173Samples 51 77Patients 18 20Models grown 13 88Patients w/ grown models 5 14Samples w/ grown models 8 44Grown models (1 yr) 35% 58%Grown samples (1 yr) 50% 64%Grown patients (1 yr) 67% 77%50Table 2.5: Summary statistics for PDX collection by strain for HGSC tumours. The numberpatients and samples with at least 1 grown PDX, along with the engraftment rate (by model, sample,and patient) for models collected ≥ 1 year ago are shown.510%25%50%75%100%100 200 300 400Days from surgery to euthanasia% of established/all establishedType Models Samples Patientsa0%25%50%75%100%100 200 300 400Days from surgery to euthanasia% established/allType Models Samples PatientsbFigure 2.5: (a) Cumulative distribution function of engrafted and euthanized tumours for HGSCPDXs at the level of models, samples, and patients. (b) Percent of engrafted and euthanized tumours(out of all euthanized tumours) as a function of engraftment time. Statistics summarized at the levelof models, samples, and patients.520.0000.0050.0100.0150.020100 200 300 400Days from surgery to euthanasiaDensityEstablished 0 1Figure 2.6: Time from surgery to euthanasia for PDX models that do and do not grow tumours.2.4 DiscussionPrognosis for HGSC patients has remained poor (5-year survival approx. 35%) over the last fewdecades [138]. Despite the development of many new therapies including PARP inhibitors [16],platinum resistant HGSC remains difficult to manage [139]. Extensive clonal heterogeneity acrossspace is thought to be a key factor that engenders therapeutic resistance in this devastatingdisease. In some cases, tumour-intrinsic mechanisms of resistance such as BRCA mutationreversion [15, 140] and upregulation of drug efflux transporters have been observed. Initialin vitro work has revealed a microenvironment-mediated mechanism of resistance involvingfibroblasts and lymphocytes [59]. Yet, the mechanistic underpinnings of resistance in HGSCin the majority of cases – especially in platinum-refactory foldback inversion-subtype tumours53[3] – remain unknown. Patient-derived xenografts that faithfully recapitulate tumour histology,genomics, and expression patterns offer ideal model systems for studying drug response inHGSC. The combination of single cell-resolution assays and PDX models enables accuratecharacterization of rare clonal genotypes and phenotypes that underlie treatment resistance.Tracking clonal prevalence trajectories in PDXs will serve as the basis for understanding thecontribution of genomic subtypes to platinum resistance and for predictive modeling of clonaldynamics to inform therapeutic regimen choices for patients.The multi-site nature of HGSC necessitates collection of multiple tumour sites per patientto obtain a comprehensive view of the treatment-na¨ıve clonal repertoire. To this end, ourcohort constitutes the largest set of multi-site high-grade serous ovarian cancer patients we areaware of, with high-quality specimens for genomic, transcriptomic, and proteomic analysis andmatched patient-derived xenograft models. Our model-level engraftment rate of approximately50% after 100 days (Figure 2.5A) is within the range of previously described values (48% to90%) [141–143]. In addition, we have demonstrated that high-quality single cell copy numberprofiles can be derived from similarly constructed PDX models [97, 98]. Thus, our PDXs canbe leveraged to investigate drug response and clonal dynamics in HGSC over time and space.As the determinants of successful HGSC tumour engraftment are largely unknown, we notethat certain aspects of the engrafted cohort may not be completely representative of all HGSCtumours. Moreover, despite the similarity between model types, engraftment rates appeared tobe higher in NSG mice than NRG mice. Thus, analyses using these models will have to accountfor possible biases in cohort composition and differences between model types. Engraftmentrates may be improved by subcapsular transplantation in the kidney. Our future work willentail single cell whole-genome sequencing of carboplatin-treated PDX models to reveal anddevelop predictive models for clonal dynamics under treatment selection pressure. Additionally,we have established single cell dissociation protocols for single cell RNA-sequencing that willenable microenvironment decomposition.The cohort established in this chapter sets the groundwork for studying interactions betweenmalignant and immune cells in HGSC in Chapter 3 and studying single cell properties andmicroenvironment composition in Chapter 4.54Chapter 3The evolutionary interface betweentumour-infiltrating lymphocytes andcancer cells in multi-site HGSC3.1 IntroductionHigh-grade serous ovarian cancer (HGSC) exhibits the highest disease mortality among gyneco-logic cancers. Despite recent progress with poly ADP-ribose polymerase (PARP) inhibitor-basedsynthetic lethal approaches exploiting homologous recombination deficiency [16], HGSC remainsincurable in most cases. Characterized by profound genomic instability and clonal diversity,HGSC often presents with widespread peritoneal dissemination. Multi-site studies have revealedgenomic intratumoral heterogeneity (ITH) as a correlate to poor survival [28], as well as specificpatterns of malignant cell spread within the peritoneal cavity [27]. Importantly, the physicaldistribution of malignant clones across the peritoneal cavity is non-random, with the majorityof sites exhibiting clonal homogeneity and a minority of sites harboring diverse clones [29]. Thisraises the hypothesis that region-specific properties, including immunologic components of thetumor microenvironment, may modulate malignant cell invasion and expansion, thereby shapingevolutionary selection.HGSC patients with abundant CD8+, CD4+, CD20+, and plasma cell tumor-infiltratinglymphocytes (TILs) are associated with favorable clinical outcomes [36, 38, 41, 144]. TILscan respond to and temporally track neoantigens [39] and mitigate resistance to platinumchemotherapy [59]. However, much of our understanding of the immune response in HGSCderives from single biopsies; far less is known about spatial immunologic variation across distaltumor foci. Histologic imaging has revealed that lymphocyte abundance can vary betweentumor foci in HGSC [145]. Furthermore, lymphocyte expression signatures are linked topatterns of metastasis [146]. A single case report has described immunologic variation acrossrelapse specimens [147]; however, given the immunomodulatory effects of chemotherapy [148],55understanding of pre-treatment spatial variation is still lacking.Beyond immunologic features, prognostic mutational processes in HGSC through analysis ofpoint mutation, copy number, and rearrangement features has indicated a prominent associationbetween foldback inversions (FBIs) and poor response to platinum-based chemotherapy [3].FBI-dominated tumors, which comprise approximately 40% of HGSC, tend to be exclusiveto homologous-recombination-deficient (HRD) cases and bear a distinct pattern of high-levelamplifications colocalized with foldback rearrangements typical of breakage-fusion-bridge pro-cesses [3, 24]. How mutational processes co-vary with immune response characteristics in HGSCremains poorly understood. This will become of central importance as clinical trials assayingsynthetic lethal compounds targeting DNA repair processes combined with immune-modulationtherapies read out.We surmised that localized selective pressures imposed by immune microenvironments shapethe distribution of malignant clones during disease progression. Thus, we systematically profiledthe inter-relationship of clonal diversity, mutational processes, and immunologic response acrossa cohort of patients and multi-region samples. Genome-sequencing-based clonal decomposition,transcriptome-based T and B cell receptor sequencing, multicolor immunohistochemistry (IHC),and histologic image analyses were applied. Our results elucidate the landscape of cell-typeinteractions at the interface of malignant and immune cells across 212 samples from 38 patients.We show that samples robustly segregate into three distinct TIL subtypes, reflecting little or noimmune infiltration, stromal infiltration, and combined epithelial and stromal infiltration. Wereveal an association between these classes and malignant clone diversity properties. Regionswith highest levels of epithelial immune infiltration exhibit the lowest malignant clone diversity,neoantigen depletion, and subclonal loss of heterozygosity (LOH) at human leukocyte antigen(HLA) loci as evidence of purifying selection. Moreover, T cell clonotypes, but not B cellclonotypes, spatially track with tumor clones in patients with heavily infiltrated tumors. Finally,we show combinatorial prognostic effects between mutational processes and immune infiltrationwith foldback inversions exhibiting high risk even in the presence of high cytotoxicity. Inaggregate, our findings illuminate molecular and evolutionary properties at the immune-malignantinterface in HGSC with new insights on how tumor progression and clonal dissemination aredriven by immune-related selective pressures.563.2 Materials and Methods3.2.1 Experimental Model and Subject Details3.2.1.1 Sample acquisition, consent, & surgeryEthical approval for this study was obtained from the University of British Columbia (UBC)Research Ethics Board. Women (biological sex: XX) undergoing debulking surgery (primaryor recurrent) for carcinoma of ovarian/peritoneal/fallopian tube origin were approached forinformed consent to bank tumor tissue. Cases of high-grade serous carcinoma where more thanone sample was collected were chosen for this analysis. Clinicopathologic and outcome data werecollected by chart review. Consistent with the practice at UBC and BC Cancer, all patients withhigh-grade serous ovarian cancer (HGSC) are referred to the hereditary cancer clinic and offeredgenetic testing for BRCA1 and BRCA2 mutations (http://www.bccancer.bc.ca/screening/Documents/HCP_GuidelinesManuals-HBOCCriteria.pdf).For consented patients, when multiple tumor sites were encountered intraoperatively, effortwas made to bank as many sites as possible. Samples were flash frozen and stored according toconditions outlined below. For cases where multiple tumor sites were encountered but not allanatomic sites could be frozen (e.g., due to unavailability of trained staff), archival specimensstored within our pathology department were used. All samples were from removed structuresduring attempts at optimal debulking; hence the majority of samples were from omentum andovarian sites.Platinum sensitive is defined as no relapse within 6 months of the chemotherapy stop date.3.2.1.2 Sample preservation & histologic evaluationWhen adequate tumor volume was available, multiple tissue samplings were obtained from eachtissue specimen. Up to 5 samplings were taken from a given tumor, with effort made to equallyspace samples while staying within grossly apparent tumor tissue. Each sampling was cut intothree pieces, yielding two end-pieces for cryovials and a middle portion placed in 10% bufferedformalin. End pieces were homogenized manually and with a paddle blender (Stomacher). Allparaffin-embedded blocks, including formalin-fixed tumor samples and molecular-fixed fallopiantubes, were sectioned and stained with hematoxylin and eosin prior to expert histopathologicalreview to confirm the presence of high-grade serous carcinoma. Pieces from the same samplingwere given the same sample identifier for the analysis steps described below.573.2.2 Method Details3.2.2.1 WGSS library construction & sequencingFrozen tumor samples from 14 patients (patients 11-17, 25, 26, 28-32, total 71 samples) weresubmitted for library construction and sequencing. Sample size was determined by availabilityof resectable, cryopreserved tissue, and DNA quality. For all tumor and normal samples, DNAextraction was followed by library construction and sequencing using Illumina HiSeq2500 wholegenome shotgun v4 chemistry with paired-end 125bp reads. Samples were sequenced to anaverage of 96X coverage. Patients 1-4, 7, 9, and 10 were previously sequenced according tospecifications described in [29].3.2.2.2 Targeted bulk sequencing analysisTarget selection For each patient we performed targeted sequencing on (11-17), a total of 192positions were deeply sequenced, including 4 experimental controls, a TP53 variant, heterozygousgermline SNPs lost in dominant loss of heterozygosity (LOH) events, lost SNVs that could andcould not be explained by copy number events, and SNVs inferred to originate at each node ofthe sample phylogeny obtained by applying the stochastic Dollo approach (infinite sites withloss model) [29] (Supplemental Table A.1). SNVs were sampled as evenly as possible acrossnodes.Data for patients 1-4, 7, 9, and 10 was obtained from [29], and used as input for sectionClonal analysis onward.Primer design Primers targeting the positions described above were designed using primer3.The full list of primers is included in Supplemental Table A.1. Optimal primer lengthwas 27nt (18-30nt) and products were designed to be 150-250nt long with 53-61◦C meltingtemperature. Breslauer thermodynamic correction and Schildkraut and Lifson salt correctionsettings in primer3 were used. Additionally, primers targeting SNVs were required to pass thefollowing preliminary filters: minimum of 5 alignments to the genome as given by BLAT foreach primer, and each primer position at least 30nt away from the target SNV.Primers were additionally tested using a combination of UCSCs in silico PCR tool (http://genome.ucsc.edu/cgi-bin/hgPcr) aligned against the reference hg19 genome and customin-house code (Canadas Michael Smith Genome Sciences Centre) to verify a unique hit andcheck that the variant was located within 150bp of the nearest end of the amplicon to ensurecoverage in an Illumina NextSeq 150bp paired end read. The primers were tagged with Illuminaadapters to enable a direct sequencing approach that precludes the need for adaptor ligation58during sample preparation. The Illumina adaptor tags were: 5’-CGCTCTTCCGATCTCTG-3’ onthe forward amplicon primer and 5’-TGCTCTTCCGATCTGAC-3’ on the reverse amplicon primers.PCR and Illumina sequencing Genomic DNA templates were used as starting material togenerate PCR products. PCR was set up using Phusion DNA polymerase (Fisher Scientific,USA) according to the manufacturers specifications. The standard PCR conditions used werean initial denaturation at 98◦C for 30 s, followed by 35 cycles of 98◦C for 10 s, 60◦C for 15 sand 72◦C for 8 s, and a final extension at 72◦C for 10 minutes.Amplicons were pooled by template for sequencing sample preparation. Sample preparationinvolved a second round of amplification using Phusion DNA polymerase with 6 PCR cyclesusing PE primer 1.0-DS (5’-AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCTCTG-3’) and a custom PCR Primer (5’-CAAGCAGAAGACGGCATACGAGATNNNNNNGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTGAC-3’) that contains a unique six-nucleotide ’index’ shownas N’s. PCR products were cleaned up using PCRClean DX beads (Aline Biosciences, USA).DNA quality was assessed using the Caliper LabChip GX High Sensitivity Assay (Caliper LifeSciences, USA) and DNA quantity was measured using a Qubit dsDNA HS assay kit on a Qubitfluorometer (Life Technologies, USA).The indexed libraries were pooled together and sequenced on the Illumina NextSeq500platform with paired-end 150bp reads using v2 chemistry reagents.3.2.2.3 ImmunohistochemistryAll reagents were from Biocare Medical (Pacheco, CA) unless otherwise stated. Slides offormalin-fixed, paraffin embedded tissue were deparaffinized and rehydrated through xylene andgraded alcohols. Antigen retrieval was performed using Diva Decloaker in a Biocare decloakingchamber at 125◦C for 30 s. Slides were then rinsed with water, marked with PAP pen andloaded into the Biocare Intellipath FLX autostainer. Slides were blocked with peroxidazed-1and background sniper for 5 minutes and 10 minutes respectively then a cocktail of eitherCD8 (1/250, clone C8/144B, Cell Marque, Rocklin, CA) and CD3 (1/500, clone SP7, SpringBiosciences, Pleasanton, CA), or CD79a (1/400, clone SP18, Spring Biosciences, Pleasanton,CA) and CD138 (1/200, clone B-A38, Biocare Medical, Pacheco, CA) in Da Vinci Green diluentwas added for 30 minutes at room temperature. Following a wash step, Mach2 Doublestain#2 polymer was added for 30 minutes at room temperature and then antigens detected withIP Ferengi Blue chromogen for 7 minutes followed by IP DAB chromogen for 5 minutes. Todenature the first round of antibodies, slides were removed from the autostainer and placed inpre-warmed SDS-glycine pH 2.0 solution for 45 minutes at 50◦C with periodic agitation. Slides59were then washed with water and replaced in the autostainer for the 2nd round of staining.CD20 (1/300, clone L26, Biocare Medical, Pacheco, CA) diluted in Da Vinci Green diluentwas added to the slides and incubated for 30 minutes at room temperature. Mach2 Mouse-APpolymer or Mach2 Rabbit-AP polymer was added for 30 minutes at room temperature to detectCD20. Warp red chromogen was added to the slide for 7 minutes, hematoxylin at a 1/5 dilutionwas then added for 5 minutes. The slides were then washed, air-dried and coverslipped withEcomount coverslipping medium.3.2.2.4 Nanostring gene expressionFFPE samples were deparaffinised with xylene and washed with 100% ethanol. Tissue wasthen extracted using QIAGEN miRNeasy FFPE Kit, following the protocol for purificationof total RNA (including miRNA) from FFPE tissue sections. RNA quality was assessed withNanodrop. 500ng of high quality RNA (260/280 ratio of 1.7-2.3 and A260/230 ratio of 1.8-2.3)for each sample was used in the Nanostring assay (PanCancer Immune Profiling panel [149]additionally containing markers for high-grade serous ovarian cancer subtypes C1, C2, C4, andC5 [150]). Data was normalized with the voom function from the R package limma and TMMnormalization. Samples flagged by nSolver (Nanostring Technologies) were removed from furtheranalysis.3.2.2.5 TCR & BCR sequencingIn the text below, TRB and IGH refer to TCR-β chain and Ig-heavy chain, respectively.RNA was extracted from frozen tissue using the miRNeasy Mini kit. Quality (260/280)and quantity were determined using Nanodrop. Total RNA samples were also QC checkedusing the Caliper HT RNA HiSens assay (Caliper Life Sciences, USA). Samples ranging from60-255ng RNA were re-arrayed into a 96-well plate. First-strand cDNA was synthesized from thetotal RNA samples using the SMARTScribe Reverse Transcriptase from Clontech, BNA oligo,TRB and IGH gene specific primers at a concentration of 0.5uM. Reactions were incubatedon a tetrad using the following program: 90mins at 42◦C, 15mins at 70◦C and 2mins at 4◦C.Using cDNA as a template, first round PCR for TRB and IGH was set up using Phusion DNApolymerase (Fisher Scientific, USA) according to manufacturers specifications. The gene specificprimers used were TRB 5’-TCTCTGCTTCTGATGGCTCAAAC-3’ and IGH 5’-ACACCGTCACCGGTTCGG-3’. The PCR conditions used were an initial denaturation of 98◦C for 30 s, followed by 35cycles of 98◦C for 10 s, 55◦C for 10 s and 72◦C for 20 s, and a final extension at 72◦C for 5minutes. PCR products were size selected and cleaned up using PCRClean DX beads (Aline60Biosciences, USA). Using first round PCR product as a template, a nested round of PCR forTRB and IGH was set up using Phusion DNA polymerase (Fisher Scientific, USA) according tomanufacturers specifications. The gene specific primers used were TRB 5’-TGCTCTTCCGATCTGACAGCGACCTCGGGTGGGAACA-3’ and IGH 5’-TGCTCTTCCGATCTGACAAGACSGATGGGCCCTTGGT-3’.The PCR conditions used were an initial denaturation of 98◦C for 30 s, followed by 10 cyclesof 98◦C for 10 s, 65◦C for 10 s and 72◦C for 20 s, and a final extension at 72◦C for 5 minutes.PCR products were cleaned up using PCRClean DX beads (Aline Biosciences, USA).TRB and IGH amplicons were pooled by template for sequencing sample preparation. Samplepreparation involved a second round of amplification using Phusion DNA polymerase with 6PCR cycles using PE primer 1.0-DS (5’-AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCTCTG-3’) and a custom PCR Primer (5’-CAAGCAGAAGACGGCATACGAGATNNNNNNGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTGAC-3’) that contains a unique six-nucleotide ’index’shown here as N’s. Products were cleaned up using PCRClean DX beads (Aline Biosciences,USA). DNA quality was assessed using the Caliper LabChip GX High Sensitivity Assay (CaliperLife Sciences, USA) and DNA quantity was measured using a Qubit dsDNA HS assay kit on aQubit fluorometer (Life Technologies, USA).The indexed libraries were pooled together and sequenced on the Illumina HiSeq platformwith paired-end 250bp reads using v2 chemistry reagents.3.2.3 Quantification and Statistical Analysis3.2.3.1 WGSS analysisAlignment Reads were aligned to the hg19 reference genome downloaded from http://www.bcgsc.ca/downloads/genomes/9606/hg19/1000genomes/bwa_ind/genome/GRCh37-lite.fa.Alignments were performed using bwa [151] using the aln and sampe commands. Duplicateswere flagged with Picard http://broadinstitute.github.io/picard/.SNV and indel calling Somatic SNVs were called using both Strelka 1.0.14 [152] andMutationSeq 4.2.0 [153] with default parameters. Somatic indels were additionally called withStrelka. We considered a somatic SNV high quality if it was predicted by both MutationSeqand Strelka to be present in any sample from a patient, not necessarily the same sample foreach program. Germline SNVs and indels were called with samtools mpileup and bcftools call1.4.1, with default parameters.Gene name, predicted effect and impact of SNVs and indels were annotated using SnpEff4.0e. Mappability scores were annotated for each position using precomputed values down-61loaded from UCSC (http://hgdownload-test.cse.ucsc.edu/goldenPath/hg19/encodeDCC/wgEncodeMapability/release3/wgEncodeCrgMapabilityAlign50mer.bigWig). For downstreamanalysis we only considered variants with a mappability score > 0.99.Breakpoint calling We used deStruct [154] and lumpy [155] to call breakpoints from WGSdata. deStruct breakpoints were filtered for those with at least 2 discordant reads, and at least2 split reads. Additional filters removed breakpoints for which the reconstructed sequence wasless than 120nt, and removed breakpoints with read data likelihood less than 20. Followingthis, the intersection of deStruct and lumpy predictions was taken, and events lying withinpoor mappability regions, with break distance ≤ 30bp, and deletions with breakpoint size <1000bp were excluded [3]. Furthermore, breakpoints overlapping germline structural variationas determined from the database of genomic variants or identification of a similar event in thematched normal sample. Classification of breakpoint and rearrangement type was performedaccording to [3].Copy number calling We applied ReMixT [156] to predict allele and clone-specific copynumber from WGS samples. ReMixT jointly infers clone and allele specific copy number of bothsegments and breakpoints, allowing for increased statistical strength for detecting subclonalrearrangements associated with subclonal copy number changes. Additionally, ReMixT useshaplotype blocks obtained from phased SNPs to increase the power for detecting small allelicimbalances resulting from subclonal copy number changes. ReMixT was run on each patientsfull set of WGS samples with default parameters. Accurately inferred clone specific segmentcopy number was used to calculate the length-normalized proportion of segments predicted withdivergent clonal copy number.In order to call high-level amplification (HLAMP), we employed identical methods to [3].We ran TITAN [157] on WGS data to infer logR values; HLAMP was called for segments withmedian logR values > 1.Identifying BRCA variants Point mutations and indels in BRCA1 and BRCA2 were calledfrom germline and somatic WGS data, as described above. Variants with high SnpEff-annotatedimpact were used. Somatic BRCA status was determined from variant calls. Where available,clinical test results were used to determine germline BRCA status; germline variant calls wereused for patients that did not consent to clinical testing. Clinically-determined BRCA status isshown in Table 3.1.623.2.3.2 Clonal analysisMutation cluster inference We ran PyClone 0.13.0 [71] in multi-sample mode to performinitial clonal analysis. Parental copy number and tumor content estimates from ReMixT alongwith reference and alternative allele counts from deep sequencing data of SNVs (PCR andIllumina sequencing) were used as input for PyClone. The following SNVs were filtered out forclonal analysis: germline SNVs, SNVs absent (probability < 0.01) in all samples in a patient(probabilities computed from a binomial test, assuming a sequence error rate of 0.001), andSNVs on sex chromosomes. The MCMC chain was run for 100,000 iterations, with a burn-in of50,000. Posterior plots were visually inspected to confirm convergence. Flat cluster assignmentswere produced from posterior similarity matrices using the MPEAR method described in [71].SNVs with broad posterior cellular prevalence distributions (width of 95% credible interval ≥0.2) far from the corresponding cluster median (difference of ≤ 0.05) were excluded from furtheranalysis. Additionally, clusters absent or present at low prevalence in all samples (median clusterprevalence across SNVs ≤ 0.05 in all samples), with only one SNV, or with ≥ 50% SNVs lostwere filtered out.Archival samples without a corresponding flash frozen sample (i.e., no copy number pre-dictions) were excluded from this initial analysis. They are reintroduced in section Clonalphylogenies & postprocessing.Clonal phylogenies & postprocessing Filtered PyClone results were provided as input toLICHeE, a multi-sample cancer lineage inference method [75], to elucidate clonal phylogenies.LICHeE was run in cellular prevalence mode (-cp), with additional options -completeNetwork-sampleProfile. Other parameters were set to the defaults. The top ranking lineage tree fromLICHeE was kept.To remove artifacts (e.g., falsely called low prevalence clones) and obtain clonal prevalencesfor archival samples, clonal prevalences were refined by resampling alternative and reference allelecounts for deeply sequenced tumor samples and matched peripheral blood (normal) accordingto the following Bayesian generative model, adapted from [29]. We suppress indices for samplesas these can be treated independently.We assume that the alternative allele counts of SNV n in the matched normal and tumorsamples, bnnormal and bntumor, respectively, are distributed as:bnnormal|pnnormal ∼ Binomial(dnnormal, pnnormal) (3.1)63andbntumor|ψn, Zn = c, pnnormal ∼ BetaBinomial(dntumor, ξ(ψn, φc, t, pnnormal), σtumor) (3.2)where dnnormal and dntumor correspond to the total read depth of SNV n in the normal and tumorsample, respectively, pnnormal is the probability of observing the alternative allele of SNV n inthe normal sample, σtumor is the dispersion parameter, Zn is the cluster membership of SNV n,and ξ(ψn, φc, t, pnnormal), using similar notation to [71], is given by:ξ(ψn, φc, t, pnnormal) =(1− t)c(gN )Tnpnnormal +tφcψnTnwhere ψn is the copy number genotype of SNV n in the tumor variant population, t is tumorcontent, c(gN ) = 1 is the copy number genotype of the alternative allele in the normal population,total copy number Tn = 2(1 − t) + ψnt, and φc is the cellular prevalence of PyClone clusterc, which can be expressed as the summation of clonal prevalences fj over clones that containPyClone cluster c. That is:φc =∑j:Gcj=1fjwhere Gcj is a binary indicator of whether clone j contains PyClone cluster c. We then assumethe following distributions over the parameters in equations 3.1 and 3.2:f ∼ Dirichlet(κ)ψn ∼ Categorical(1)pnnormal ∼ Beta(ζ ∗ σnormal, (1− ζ) ∗ σnormal)with κ the Dirichlet parameter as defined in [29], and σnormal the dispersion parameter. Thevalue ζ corresponds to twice the mean allelic fraction of alternative alleles in the normal sample(twice because we model c(gN ) = 1). In essence, our model is analogous to that of [29], but wenow consider the probability of sampling a variant allele from non-tumor cells to be nonzero,equal to pnnormal, rather than 0.Informally, the model can be described as follows. For each tumor sample:1. Generate clonal prevalences642. Compute the cellular prevalence of a mutation n by summing the prevalences of all clonescontaining the PyClone cluster associated with n3. Generate the SNV-specific normal contamination fraction pnnormal and allelic count data forthe matched normal sample4. Based on the contamination fraction, apply a modified PyClone likelihood model to simulateallelic count data in the tumor sampleThe normal contamination fraction can be interpreted as the allelic fraction of SNV n inthe matched normal, likely due to sequence errors or contamination. Samples with low tumorpurity are particularly confounded by these issues; the addition of step 3 and modification ofstep 4 relative to [29] helps eliminate erroneously identified rare clones in these samples.We set the following hyperparameter values: σtumor = σnormal = 200 and κ as a repeatingvector of 0.01. The effect of our setting for κ is to assume clonal purity unless there is substantialevidence for the contrary.The Hamiltonian Monte Carlo chain was run for 10,000 iterations, with an additional burninof 5000. Posterior plots were visually inspected for convergence. Clones falling below a prevalencethreshold (< 90% of the posterior distribution of clonal prevalence > 0.01) were removed.Due to difficulties in lineage construction for patients with several samples composed ofdivergent clonal lineages [29], results for patients 3 and 9 were taken from previously analyzedsingle-cell sequencing data [29].Clonal architecture distance Pairwise similarity between clonal compositions (within agiven patient) was computed using a modified version of the weighted uniFrac measure, to simul-taneously incorporate clonal architecture and phylogeny information. First, clonal phylogeniesfrom Section 3.2.3.2 were taken as ground truth and used to recompute cellular prevalences forall SNVs (ψa and ψb) determined by WGS, where a and b denote the samples being compared.Clonal distance was computed as the summation of the differences in cellular prevalences acrossSNVs, or equivalently ||ψa −ψb||1.Measures of intratumoral heterogeneity Sample mixture entropy and clone divergencewere defined as in [29]. In order to compute divergence, SNVs from WGS data were assigned toPyClone clusters - and transitively, clones - by maximum likelihood according to the PyClonelikelihood model [71]. Proportion subclonality (copy number based measure) was computed asthe proportion of the genome with subclonal copy number according to results from ReMixT.Heterogeneity index, a combined measure of intratumoral heterogeneity incorporating both cloneprevalences and phylogenetic relationships, was computed as the sum of relative phylogenetic65divergence between all pairs of distinct clones, weighted by clonal prevalence. The heterogeneityindex is the mean phylogenetic divergence between a randomly selected pair of tumor cellsfrom a sample (based on inferred clonal composition). Formally, for a sample A with clone setC(A) = {ci} and corresponding prevalences pi (where 0 < pi < 1,∑i pi = 1):HI(A) =∑cj ,ck∈C(A)pjpkD(cj , ck)where D(cj , ck) is the relative phylogenetic divergence between clones cj and ck, defined as:D(cj , ck) =|Scj ∪ Sck | − |Scj ∩ Sck ||Scj ∪ Sck |where Sci is the set of WGS SNVs assigned to clone ci. By construction, the heterogeneity indexobtains values between 0 and 1. Intratumoral heterogeneity values for each sample are listed inSupplemental Table A.2.Samples were also assigned to clonal mixture classes (pure, chain, branched) based on thephylogenetic relationships between constituent clones. Pure samples contained a single clone;chain samples contained clones along a single lineage (in other words, the minimal spanning treeis a line); branched samples contained at least 2 clones that were not ancestors/descendants ofeach other (in other words, the minimal spanning tree contains a bifurcation).The significance of differences in the 3 clone-derived intratumoral heterogeneity measures(entropy, clone divergence, heterogeneity index) between the 3 TIL subtypes was assessed withthe Kruskal-Wallis test (Figure 3.7A). Post hoc comparisons were made with Dunns test(P-values were BH corrected).To assess the significance of differences in subclonal copy number proportion between the 3TIL subtypes, ANOVA was performed (aov function in R) with subclonal CN proportion as thedependent variable (logit-transformed, as subclonal CN proportion values lie between 0 and 1,exclusive), TIL subtype and cellularity as independent variables (to control for tumor cellularity).The residual plot did not indicate any substantial deviations from normality, with relativelyconstant variance across the fitted range. Post hoc comparisons were made with Tukeys rangetest (P-values were BH corrected).3.2.3.3 RNA-seq analysisRNA-seq raw counts for 54 primary HGSC tumors from the Australian Ovarian Cancer Study(OV-AU) [158] were downloaded from the International Cancer Genome Consortium (ICGC)Data Portal. Ensembl Gene IDs were mapped to gene symbols using biomaRt. Duplicate entries66were summarized by taking the mean of expression values. Raw counts were normalized usingvoom from the R package limma with quantile normalization.3.2.3.4 Mutation signature analysisData Mutation signatures were jointly inferred for 102 multi-site HGSC tumors (21 patients),62 primary HGSC tumors from the Australian Ovarian Cancer Study with BAM files [158], and133 additional ovarian tumors (59 HGSC, 35 clear cell, 10 germinal cell, and 29 endometrioid) [3](Supplemental Table A.5). Note that a POLE hypermutant (one of the endometrioid cases)was excluded from the original set of 133 cases described in [3], and while 93 cases were availablefrom the Australian Ovarian Cancer Study, only 62 had BAM files on the data portal. Similarlyprocessed variant calls to WGSS analysis were obtained from [3]. In order to avoid counting thesame variant more than once, the union of SNVs from all samples for each multi-site HGSCpatient was analyzed together as a ’meta-sample’.Signature inference & clustering Signatures and proportions were inferred from WGSSNV and rearrangement (structural variation, SV) calls (section WGSS analysis) by applyingthe multimodal correlated topic model method [26]. For SNVs, the pentanucleotide context ofeach variant is considered. Rearrangements (deletions, duplications, inversions, and foldbackinversions) were binned by breakpoint distance (<10kb, 10kb-100kb, 100kb-1Mb, 1Mb-10Mb,>10Mb) and microhomology length [26, 159]. The optimal number of SNV and SV clusters wasdetermined using the elbow method on model log-likelihoods [26]. The probable identity of eachpoint mutation signature is as follows: P-MMR-1 mismatch repair (MMR), P-HRD homologousrecombination deficiency (HRD), P-UM ultramutator-associated mutation signature (presentat very low levels in the HGSC samples; primarily observed because of an endometrial samplefrom [3]), P-APOBEC APOBEC, P-AGE age signature, and P-MMR-2 uncertain, but witha strikingly similar T→C substitution pattern to the MMR signature. Sample-specific andnon-ancestral mutation signatures were calculated by adding signature assignment weights forall constituent variants. For non-ancestral analysis (Figure 3.12D), non-ancestral SNVs weredefined as those not present (and not called as ancestral) in all samples from that patient, andsamples with fewer than 50 non-ancestral SNVs or SVs were excluded. Prior to clustering(Figure 3.12A, Figure 3.11C,D), signature proportions were scaled across the entire pooledcohort to a standard Gaussian distribution. Hierarchical clustering was performed with Wardsmethod and a Pearson correlation-based distance measure (d = (1− r)/2, where r is the Pearsoncorrelation coefficient). For patients in the discovery cohort with more than 2 samples, molecularsubtype annotations on the heatmap correspond to the mode of subtype assignments for each67patient. The 4 described subtypes (HRD-DEL, HRD-DUP, FBI, and TD) were recovered usingthe dynamicTreeCut R package (or equivalently, by cutting the dendrogram into 4 clusters).Association with immune markers RNA-seq expression data (see RNA-seq analysis, Nanos-tring analysis) from a set of 54 untreated primary OV-AU cases was used for the comparisondepicted in Figure 3.12C.Differential gene expression Differential gene expression analysis between mutation signa-ture clusters for ICGC OV-AU cases (see RNA-seq analysis) was carried out using the limmamethod (R package). limma results for HRD versus FBI, TD versus FBI, and HRD versus TDcontrast matrices were fed as input to the R package GAGE for gene set enrichment analysisusing KEGG pathways. Pathways significantly up- or downregulated with Q ¡ 0.01 were re-garded as significant. Results of differential expression analysis are shown in Figure 3.13 andSupplemental Table A.6.TCGA foldback inversions A set of n = 433 TCGA ovarian serous cystadenocarcinomacases with complete copy number, clinical, hg19 exome BAM files, and array-based geneexpression data was selected for analysis [14]. Selected TCGA cases are listed in SupplementalTable A.7. Expression data was downloaded from the TCGA data portal and clinical data wasdownloaded from the TCGA Pancancer project under Synapse (ID: syn1461171).Array gene expression data was preprocessed with the voom function from limma (R package),using quantile normalization. The median of normalized expression values for genes associatedwith cytotoxicity (derived from Nanostring PanCancer Immune Profiling Panel annotations[149]) was computed. Samples were stratified into immune-high and immune-low classificationsby thresholding on median cytotoxicity score across the cohort (Supplemental Table A.7). Tothreshold on FBI status, foldback-amplification colocalization status (FBI-AMP High, FBI-AMPLow, No AMP) for all cases was retrieved from [3]. We performed a survival analysis on FBIgroups after subsetting by immune cluster. The log-rank test was used to compare survivaloutcomes between subgroups.A Cox proportional hazards model was also fit to the overall survival data, using foldback-amplification colocalization status as a discrete explanatory variable, interaction terms betweencytotoxicity score and FBI-HLAMP status, along with control variables for age of pathologicdiagnosis and treatment regimen (columns immunotherapy, additional immunotherapy, additionaldrug therapy, and additional chemotherapy in the Synapse table). Age of diagnosis wasbinned into < 50, 50-70, and > 70 categories, and along with immunotherapy and additional68chemotherapy used as stratification variables (as these originally violated the proportionalityassumption). Patients without available data for age of diagnosis (5) were excluded. To assess thevalidity of the proportional hazards assumption, the cox.zph function the survival R packagewas used. None of the individual proportionality assumption tests or the global test wereviolated.The R formula for the model was:1 coxph ( s u r v i v a l ˜ mutation s i gna tu r e subgroup + cy t o t o x i c i t y : mutation s i gna tu r esubgroup + s t r a t a ( age binned ) + s t r a t a ( immuno therapy ) + s t r a t a ( add i t i ona lchemo therapy ) + add i t i o na l drug therapy + add i t i o na l immuno therapy , data )To evaluate the significance of the model including the cytotoxicity× FBI-HLAMP interactionterm, we constructed an identical model, but with a cytotoxicity score as an explanatory variablewithout the interaction terms with FBI-HLAMP. A likelihood ratio test was performed on theresulting fits of the 2 models.3.2.3.5 Immunohistochemistry analysisTissue segmentation & cell counting Slides were scanned using the Vectra MultispectralImaging System (Perkin Elmer) and 20 random 20× images (high-powered fields, HPFs) collectedfor each sample. The resulting multispectral images were then analyzed using Inform software(Perkin Elmer) with the resulting cell segregation data consolidated using Spotfire (Tibco).Phenotyping algorithms were created by 2 independent researchers (K.M., S.L.) and the resultsvalidated by a 3rd researcher (A.W.Z.). Briefly a training set of 10 images, selected to behistologically diverse on visual inspection, was used by each of the researchers to train Informto recognize the different phenotypes of interest in each image. Training was run until at least98% validation accuracy was achieved. The 2 algorithms were compared and visual inspectionused to confirm the cell counts. TIL densities for each image were calculated by normalizingvalidated TIL counts by total area covered by tissue in the image (in units of cells/HPF). OverallTIL densities for each slide were similarly calculated, but using the summation of TIL countsand area across all constituent images. Epithelial and stromal TIL densities employed similarcalculations, with counting and area restricted to epithelial/stromal regions identified by tissuesegmentation (Inform). Thus, a cell was called epithelial if it fell within epithelial regionsidentified by Inform, and stromal if it fell within identified stromal regions.69Correlations between TIL densities Correlations between TIL densities (epithelial andstromal CD8+, CD4+, CD20+, and plasma cell) were quantified with Spearmans correlationcoefficient (Figure 3.9A) and P values of their significance were adjusted for multiple testingwith the Benjamini-Hochberg method.Clustering Hierarchical clustering of TIL density profiles was performed using Wards methodwith Euclidean distance. Heatmap values were obtained by normalizing (to a standard Gaussiandistribution) across samples for each TIL type. For Figure 3.1B,C, only samples with validepithelial and stromal TIL densities (i.e., non-zero epithelial and stromal tissue area) are shown.Additionally, for Figure 3.1B, only samples with both TIL density and Nanostring expressiondata are shown. The optimal number of clusters (3) was determined with the Dunn index.Malignant clone similarity and TIL subtype To compare whether samples from thesame TIL subtype were more clonally similar (within patients), we used a nested ranks test(nestedRanksTest R package), treating patient as a random effect. Specifically, for each pairof samples within a patient, we (1) categorize them as belonging to the same, or different TILsubtypes (til cluster comparison); and (2) compute clonal composition similarity as perClonal architecture distance (clonal similarity). Then, we run:1 nestedRanksTest ( c l o n a l s im i l a r i t y ˜ t i l c l u s t e r comparison | pat i en t id , data )3.2.3.6 Nanostring analysisMolecular subtyping Ground truth molecular subtypes for a training set of 62 primaryHGSC tumors from [158] were obtained from the authors. Matched RNA-seq data for thesetumors was obtained from the International Cancer Genome Consortium (project OV-AU)and normalized according to section RNA-seq analysis. The resulting expression profiles werepooled with Nanostring-derived expression profiles, and subjected to batch effect correction withthe ComBAT R package. To confirm the effectiveness of batch correction, expression profilesfrom all samples were hierarchically clustered. Samples from different batches were not clearlysegregated.Following this, a k-nearest neighbors classifier (k = 5) was trained and applied to the datausing the [158] molecular subtypes as ground truth. Six-fold cross-validation accuracy of 85.8%on ground truth data was obtained, similar to that reported in [150]. As comparison, thediagonal LDA classifier attained an inferior 80.9% cross-validation accuracy and was thus not70used. To further test these molecular subtypes, a subset of 62 tumors was additionally profiledwith the Affymetrix U133A2 microarray platform. As described in [27], the expression data fromthese tumors was normalized with RMA and quantile normalization, corrected for batch effectswith ComBAT, pooled with TCGA array expression data (see TCGA foldback inversions), andsubjected to another level of batch effect correction with ComBAT. Following the methodsof TCGA [14], consensus non-negative matrix factorization (NMF) was applied to determinemolecular subtypes (k = 4). NMF-derived subtypes and k-nearest neighbor-derived subtypeswere largely concordant (mutual information: 0.74).Overrepresentation of each molecular subtype or set of molecular subtypes within eachIHC-based subgroup (N-TIL, S-TIL, ES-TIL) was computed relative to the other 2 subgroupsand other molecular subtypes with Fishers exact test.Pathway signature analysis Genes were grouped on the basis of pathway annotations fromthe Nanostring PanCancer Immune Profiling panel [149]. Metagene expression values wereconstructed by taking the median of expression values for constituent genes in each pathway.3.2.3.7 TCR/BCR-seq analysisAlignment and clonotype calling Alignment to germline TCR and BCR segments wasperformed with mixcr align from MiXCR 2.0 [86], using the human IMGT reference (https://github.com/repseqio/library-imgt/releases, commit d993d704553c0a1e905c702ab93c99c0001b30d9). Reads mapping to the same clonotype were clustered using mixcr assemble,and the resulting TRB and IGH clonotypes were exported with mixcr export. Clonotypeswere identified by V and J germline gene names and CDR3 nucleotide sequence. All other mixcrparameters were set to the defaults.Decontamination and quality control Clonotypes with fewer than 5 assigned reads wereimmediately removed. In order to filter out potential cross-sample contamination, clonotypesshared between samples from different patients were identified. Clonotypes present at an absoluteprevalence (read count) in one sample > 25 times lower than in another sample from a differentpatient were removed (from the former sample). Consistent with contamination, samples (fromdifferent patients) arranged close by on each 96-well PCR plate contained a larger number ofshared clonotypes. Finally, clonotypes that produced non-functional (frameshift or prematurestop) receptor sequences were removed.Prior to computing repertoire diversity or similarity, TCR/BCR reads were randomlydownsampled (using the minimal nonzero library size across the cohort, for TCR/BCR separately)71were randomly downsampled (10 times) with replacement from each sample to account fordifferences in library size. Mean clonotype abundances across these resamplings were used for thecomputations described below, and the corresponding statistics are reported in SupplementalTable A.2.Calculating repertoire diversity The following indices of diversity were calculated:• Number of unique clonotypes• Shannon’s entropy• Efron-Thisted index• D50 index (https://patents.google.com/patent/WO2012097374A1/en)The Efron-Thisted index estimates the total repertoire diversity (by estimating the numberof unseen clonotypes), and the D50 index quantifies the preponderance of rare clonotypes in arepertoire.Correlations between repertoire diversity and ITH were computed as Spearmans rankcorrelation, using the first 2 measures listed above.Repertoire similarity analysis Pairwise similarity between TCR/BCR repertoires A andB was calculated with the Morisita-Horn index (R package vegan):S(A,B) =2∑Ni=1AiBi|A||B|(∑Ni=1 A2i|A|2 +∑Ni=1B2i|B|2 )where Ai denotes the number of reads associated with clonotype i in repertoire A, |A| and |B|are the total number of clonotype reads in A and B, respectively, and N is the number of uniqueclonotypes in A ∪B.Correlation with clonal composition TCR repertoire and clonal dissimilarity matriceswere computed as described above. These dissimilarities were correlated with Mantels test.Uncorrected P -values are reported in Figure 3.10 and Figure 3.5.3.2.3.8 TCR clonotype classificationPrevious studies have revealed differences in the physicochemical properties of CDR3 sequences[160] and VJ (Vβ-Jβ) gene usage [161] between CD8+ and CD4+ T cells. We designed a binaryclassifier to predict the class (CD8+ or CD4+) of a T cell receptor based on both germline VJgenotype and physicochemical properties of the TCR CDR3 sequence.72Training data To train the classifier, unprocessed TCR sequence data from flow-sorted naiveCD8+ and CD4+ mononuclear cells derived from 18 unrelated healthy donors were obtainedfrom a previous study [162]. We made an effort to obtain TCR-sequence data of flow sortedCD8+ and CD4+ T cells from other sources as well [160, 161], but these data were short-reador had been preprocessed (with no raw sequence files available), and thus not amenable touniform downstream analysis. While these training data were derived from naive T cells, [161]have reported that there are no significant differences in Vβ and Jβ usage between naive andmemory T cells (for both CD4+s and CD8+s separately). For the analysis described below,we operated under the assumptions that differences in VJ gene usage patterns and CDR3physicochemical features between CD4+ and CD8+ T cells are similar in the training and multi-site HGSC datasets. We later assessed the validity of these assumptions by comparing predictedCD8/CD4 abundance with results from immunohistochemistry (see Classifier). Alignmentand clonotype calling were carried out according to the methods described in Alignment andclonotype calling. Twenty percent of the data, stratified by class, was randomly split off fortesting; 5-fold cross-validation was carried out on the remaining 80%.Features V and J genotypes were binarized (80 features). Additionally, Atchley factors (Rpackage HDMD) quantifying the physicochemical properties of amino acids at each position inthe CDR3 were used (5n features, where n is the CDR3 amino acid length). Separate classifierswere trained for each length category between 11 and 18 amino acids (0.70 of all clonotypes).The distribution of V and J gene usage was comparable between training and test data.Classifier A binary gradient-boosted tree classifier was trained on the data described in sectionAlignment and clonotype calling. Training with 5-fold cross-validation was allowed to proceeduntil 100 consecutive rounds of no improvement in validation accuracy. Based on area underthe receiver operating characteristic curve, the gradient-boosted tree classifier outperformedrandom forest, logistic regression, support vector machine (SVM), and extreme value regressionclassifiers. The classifier was then applied to clonotype calls from TCR-seq data of multisiteHGSC samples to predict whether each clonotype was CD8-type or CD4-type. Clonotypesassigned to either class with >80% probability were kept.Clonotype distribution broadness across tumor samples within each patient was computedwith Simpsons diversity index on the vector of per-sample relative clonotype prevalence values(R package vegan). The significance of differences in the distribution broadness between CD4+and CD8+ associated TCRs was evaluated by computing the average of CD4+ and CD8+ TCRdistribution broadness values within each patient, and applying the Wilcoxon signed-rank test73for paired data between the two groups.3.2.3.9 Neoantigen analysisHLA typing Four-digit HLA class I types were determined from WGS data for each multisiteand background patient (see Neoantigen depletion score) using OptiType [163]. OptiType wasrun on the WGS bam of the normal sample.Sample-level HLA LOH prediction For OV-AU and [3] patients, HLA class I loss-of-heterozygosity (LOH) was called from tumor and matched normal bams as well as OptiType4-digit HLA types using LOHHLA [52]. HLA LOH was called for an allele if the estimatedcopy number (with binning and B-allele frequency settings) was < 0.5 and the significance ofallelic imbalance p < 0.1 (paired t test, no duplicate counts). A less stringent P -value threshold(compared to [52]) was used due to the lower depth of the input bams.Clone-level HLA LOH prediction We devised a Bayesian statistical extension to callclone-level HLA LOH from multi-sample WGS data leveraging clonal phylogenies and clonalcompositions inferred from Clonal phylogenies & postprocessing as input. Inference is doneseparately for each heterozygous HLA locus and patient. We define:74T Tumor clone phylogenyc = {cj : j ∈ C} Set of HLA locus copy number genotypes, onefor each cloneθ ”Stay” rate between copy number statesfs,j Prevalence of clone j in tumor sample srs,i,1 ∈ N0 Read depth at polymorphic site i for allele 1 insample srs,i ∈ N0 Total read depth at polymorphic site i in samples (sum of allele 1 and 2)ρs Cellularity/tumor content of tumor sample sψs Ploidy of tumor sample sωs,i Allele 1 fraction at polymorphic site i in samplesνs Total copy number of HLA locus in sample sµs,i Mean parameter for total read depth at site i insample sMs Multiplicative factor between WGS library sizesof tumor sample s and the matched normal sam-pleNi ∈ N0 Observed read depth at site i in matched normalsampleL Set of all polymorphic sites between the 2 allelesat a given HLA locusS Set of all tumor samples for a given patientC Set of all clones in a given patientPloidy and cellularity estimates are assumed to be known and equal to the estimates fromReMixT [156]. We present our graphical model:75cTθωs,iρsfs,jrs,i,1µs,irs,iνsNiMsψss ∈ Si ∈ Lj ∈ CWe begin by defining the clone-specific copy number genotype at a given HLA locus cj as a(cj,1, cj,2) tuple (allele 1 copy number and allele 2 copy number, respectively), where allele 1can be arbitrarily assigned to either one of the 2 HLA alleles at a heterozygous locus withoutloss of generality. Given a clonal phylogeny T , we assume that the latent clone-specific copynumber genotype at a given HLA locus evolves according to a Markov chain with transition rate1− θ, ”stay rate” θ and the initial state distribution defined to be uniform across all possiblegenotypes. The transition and stay rates can be described by an n-by-n transition matrix P (nis the total number of genotype states) with diagonal entries Pii = θ and non-diagonal entriessatisfying∑j,j 6=i Pij = 1− θ. In addition, the total transition probability 1− θ is divided evenlyamongst all valid transitions (transitions from zero to non-zero allelic copy number are deemedinvalid, as an allele cannot be acquired from nothing).We use Markov chain Monte Carlo (MCMC) to sample from the posterior of c, the assignmentof genotypes to clones described above. In what follows we describe our proposed distributionsfor the observed data given c. We assume that, given c, the observed read depth of allele 1,rs,i,1, is distributed as:rs,i,1|c ∼ BetaBinomial(rs,i, ωs,i, σ),where σ is the dispersion parameter and ωs,i, the fraction of allele 1 in tumor sample s (accounting76for normal contamination), is given by:ωs,i =∑j∈Cρsfs,jcj,1 + (1− ρs).To then anchor the total copy number estimates, we use data from the matched normal bam.Given c, we assume that the total observed read depth at site i in sample s, rs,i, follows:rs,i|c ∼ NegBinomial(µs,i, α),where α is the hyperparameter of the Gamma-distributed rate parameter in the negative binomial,and µs,i, the expected read depth of polymorphic site i, can be computed as:µs,i =νsψsρs + 2(1− ρs) ×Ms ×Ni,with νs, the total copy number at the HLA locus under consideration for sample s, accountingfor normal contamination, given by:νs =∑j∈Cρsfs,j(cj,1 + cj,2) + 2(1− ρs).The space of possible clonal genotypes cj is restricted to those with total copy number≤ 6. The dispersion parameter σ for the beta binomial distribution is set to 200, and α for thenegative binomial distribution is set to 0.5.We consider the following prior distribution for the stay rate of the genotype Markov chain:θ ∼ TruncNormal(pi, δ, 0, 1),where 0 and 1 correspond to the lower and upper bounds of the truncated normal distribution,and the mean and standard deviation pi and δ were set to be relatively uninformative (0.75 and0.4, respectively).MCMC was run for 100,000 iterations, using 50,000 additional tuning iterations. HLA LOHfor a given clone j and allele a was called when ≥ 90% of the posterior trace supported cj,a = 0.Identification of putative neoepitopes All 8 to 11-mer peptides overlapping nonsynony-mous SNVs were considered candidate epitopes. MHC-I binding affinity was computed for everymutant and corresponding wild-type allele using netMHCpan-3.0 [164]. Percentile binding scoresof ≤ 2%, where the mutant epitope had equal or better affinity than the wild-type epitope, were77considered as putative neoepitopes. In cases of HLA LOH, predicted neoepitopes associatedwith the lost HLA allele were excluded (for subclonal HLA LOH, a neoepitope was only excludedif all clones containing the neoepitope also exhibited loss of the corresponding HLA allele).Neoantigen depletion score Neoepitopes were predicted from nonsynonymous SNVs in abackground set of ovarian tumors consisting of 62 primary HGSC tumors from the AustralianOvarian Cancer Study [158] and 59 additional HGSC tumors [3], following the methods describedabove. Following similar methods to [165], the probability of generating at least one overlappingneoepitope from each trinucleotide pattern was determined.For each considered tumor sample (from the multi-site HGSC cohort), the expected rate ofneoepitope-generating SNVs was calculated from the trinucleotide context of synonymous SNVsand the expected rate of nonsynonymous SNVs per synonymous SNV for each trinucleotidepattern. Mathematically, define N¯s to be the expected number of nonsynonymous SNVs persynonymous SNV with trinucleotide pattern s and B¯s to be the expected number of neoepitope-generating SNVs per nonsynonymous SNV with pattern s. Then, for a given sample i, define Yias the set of synonymous SNVs and Ni the set of nonsynonymous SNVs. We can write:Npred,i =Yi∑mN¯s(m)Bpred,i =Yi∑mN¯s(m)B¯s(m)where Npred,i and Bpred,i are the expected number of nonsynonymous SNVs and neoepitope-generating SNVs in sample i under the null model, respectively. s(m) is the trinucleotide patternfor synonymous SNV m. Denote Bobs,i to be the observed number of neoepitope-generatingSNVs in i, and Nobs,i = |Ni| the observed number of nonsynonymous SNVs in i. We then definethe neoantigen depletion score is:Ei =Bobs,iNobs,iBpred,iNpred,iLower values of this score were interpreted as evidence of higher neoantigen depletion.The within-patient relationship between the response, neoantigen depletion score and thecovariate, epithelial CD8+ TIL density was modeled with a Bayesian linear mixed model withpatient-specific random intercepts. Samples with fewer than 3 nonsynonymous mutations were78excluded. The corresponding R code (using the MCMCglmm R package) was:1 MCMCglmm( log ( observed neoant igen r a t i o / expected neoant igen r a t i o ) ˜ E CD8re s ca l ed , random=˜ pat i en t id , data=data , fami ly = ” gauss ian ” , n i t t = 500000 ,th in = 500 , burnin = 50000 , p r i o r = p r i o r )where observed neoantigen ratio/expected neoantigen ratio corresponds to Ei, epithelialCD8+ TIL density values were rescaled between 0 and 1, the residual covariance prior was setto be relatively uninformative (V = 1 and nu = 0.002 in R), and likewise for the random effectprior (V = 1, nu = 1, alpha.mu = 0, alpha.V = 1000 in R). For the fixed effect coefficient, anuninformative prior with mean 0 and variance 1010 was used. Lack of autocorrelation in theMCMC traces was confirmed with autocorr from the coda R package. Posterior densities ofparameter estimates were checked to ensure certain assumptions of the model (e.g. fixed effectbeing Gaussian-distributed) were met. Reported significance values correspond to area underthe (right) tail of the posterior distribution of the fixed effect coefficient.The across-patient relationship was computed similarly, but with no patient-specific interceptterm. To compute subclonal- or clonal-specific correlations, observed nonsynonymous mutations(and transitively, neoepitopes) were classified based on the clonal phylogenies inferred in Clonalphylogenies & postprocessing. Similar correlations between subclonal neoantigen depletion andepithelial CD8+ TIL densities were observed using multilevel analysis (intrapatient Spearman’scorrelation p = 0.034 across the cohort and p = 6.1105 in patients containing samples withhighest epithelial CD8+ TIL densities; all between-patient p > 0.2).Lymphocyte marker expression and HLA LOH CD3D, CD8A, and CD8B expressionvalues was extracted from Nanostring expression data for HGSC cases from [3] and RNA-seqexpression data from OV-AU cases (see RNA-seq analysis). As expression data from few geneswas available from the Nanostring data, expression values were modeled as a function of HLALOH using the nested ranks test (nestedRanksTest R package; gene expression as the dependentvariable, HLA LOH status as the explanatory variable, and cohort as a random effect). P -valuesrepresenting significance of the HLA LOH coefficient are shown in Figure 3.7H.The corresponding R code for the nested ranks test is:1 nestedRanksTest ( exp r e s s i on ˜ loh s t a tu s | cohort , data = data )793.2.3.10 Histologic image analysisCell classification and tissue segmentation QuPath v.0.1.2 (https://qupath.github.io/) was used to detect epithelial tissue and presumptive lymphocytes on hematoxylin & eosin(H&E) pathology slides. Briefly, slides were subjected to superpixel segmentation followingautomated tissue detection, and intensity features calculated for superpixels. A random treesclassifier was trained (by P.T.H.) to distinguish epithelial (tumor) and stromal regions fromwhitespace and other tissues using small sub-regions from 10 slides on the basis of 145 super-pixel features to produce tissue segmentation masks. QuPaths cell detection algorithm wassubsequently used to detect individual cells, and an additional random trees classifier trained todistinguish putative immune cells on the basis of 22 cellular features. Trained classifiers areavailable from the authors.Hotspot identification Cell location (coordinate) data for tumor epithelial regions from theclassifier were used as input for Getis-Ord Gi hotspot detection [166]. Getis-Ord Gi hotspotsdenote regions with statistically significant clustering of a variable of interest. Getis-Ord Gihotspots were identified for each cell type (cancer and lymphocyte).To identify hotspots, a grid composed of squares with side length s = 30 pixels was firstapplied to each tissue section image. Only epithelial regions of each image were considered.Grid squares devoid of cells were excluded from further analysis by applying a binary mask.Neighborhood weights were computed using a neighborhood size of 4s [167]. Getis-Ord GiG∗i values for each grid square i were computed using localG from spdep. For each image,permutation testing (400 random permutations of grid point counts) was applied to computeempirical P-values of G∗i . Regions with associated pi < 0.05 were called as hotspots.Samples with no identifiable epithelial regions from which to call hotspots were excluded.Cancer-immune hotspot colocalization Spatial colocalization between cancer and immunehotspots was computed with the following statistics [167]:• fC = proportion of cancer cell hotspots that are also lymphocyte hotspots• fI = proportion of lymphocyte hotspots that are also cancer cell hotspots• fCI = fractional area of tumor occupied by colocalized cancer-lymphocyte hotspots3.2.4 General statistical methodsUnless otherwise indicated, correlations between continuous data types were computed usingSpearman’s correlation coefficient and hierarchical clustering was performed with Ward’s method80on pairwise Euclidean distances. Sample sizes (n) for statistical comparisons are shown in therespective figures and supplemental figures. p < 0.05 was considered statistically significant (afteradjusting for multiple testing with the BH method). All Dunns test P -values were BH-adjusted.All boxplot whisker ends correspond to Q1 (first quartile) - 1.5IQR (interquartile range) and Q3+ 1.5IQR. Sample size estimation was not performed.3.3 Results3.3.1 High-Resolution Multi-site Profiling of Immune and Malignant Popu-lations in the HGSC Tumor MicroenvironmentWe assembled a cohort of 212 tumor samples from 38 HGSC patients (Figure 3.1). Multiplesamples per patient were collected via primary debulking surgery from ovary, omentum, and otherdistant metastatic sites (except some relapse samples from patients 7, 11, and 23; Table 3.1).TIL densities were measured by multicolor IHC, cell-type colocalization with 20× histologicimages, clonotype diversity in T and B cell populations with T and B cell receptor sequencing(TCR-/BCR-seq), total mRNA gene expression from the 770-gene Nanostring PanCancer ImmuneProfiling Panel (Cesano, 2015) augmented with 39 molecular subtyping probes [150], mutationalsignatures and clonal diversity of malignant cells from whole-genome sequencing (WGS; meandepth: 86×), and deep amplicon sequencing (mean depth: 16 278×, median number of loci: 188,Supplemental Table A.1) (Figure 3.2). Both WGS and immune data (IHC, TCR/BCR-seq,or Nanostring) were obtained for 101 samples from 21 of 38 patients.81Figure 3.1: (A and B) (A) Experiments conducted on each tumor sample. Hierarchical clustering(Wards method on L2-distances) of TIL densities from (B) discovery cohort of 119 samples from 20patients. (C) Additional cohort of 69 samples from 17 patients. Median expression of select immunepathways also shown in (B). Heatmap values standardized and clipped between 2 and 2. Sampleswith zero epithelial/stromal areas were removed. (D) Distribution of TIL subtypes by patient.82Patient Age Stage Recurrence RFS Status OFS BRCA status1 72 IIIC no N/A NED 71 screen negative2 76 IIIC yes 12 DOD 45 screen negative3 69 IIIC yes 25 AWD 73 screen negative4 53 IIIA yes 50 AWD 71 screen negative7(a) 47 IIIC yes 8 DOD 52 screen negative8 62 IIIC no N/A NED 65BRCA1 mut andunclassifiedBRCA2 variant9 53 IIIB yes 5 DOD 32 unknown10 74 IIIC no N/A NED 59 unknown11(b) 53 IIIB yes 32 AWD 174 BRCA2 mut12 62 IIIC yes 15 DOD 44 screen negative13 80 IV no N/A NED 40 screen negative14 58 IIIC yes 7 DOD 36 screen negative15 61 IIIC no N/A NED 38 BRCA1 VUS16 72 IIIC yes 23 AWD 35 screen negative17 56 IIIC yes 19 AWD 32BRCA2 andMUTYH variant18 56 IIIC yes 19 DOD 34 unknown19 59 IIIA no N/A NED 32 screen negative20 64 IIIA no N/A NED 10 unknown21 79 IIIC yes 4 DOD 45 screen negative22 73 IIIC yes 22 AWD 75rare BRCA2variant (2680GA)likely benign23(c) 65 IIIC yes 9 DOD 75 screen negative24 40 IIIB yes 22 DOD 66 screen negative25 46 IIIC yes 6 AWD 23 screen negative26 55 IB no N/A NED 14 unknown28 83 IIIC yes 4 AWD 16 unknown29 19 IV yes 5 AWD 16 BRCA1 mut30 38 IIIC no N/A NED 13 screen negative8331 38 IIIC yes 10 AWD 16 BRCA1 mut32 46 IIIC yes 1 AWD 14 screen negative37 81 IIIB no N/A NED 12 unknown38 80 IIC no N/A NED 6 unknown41 68 IIIC no N/A NED 8 screen negative42 54 IIIC no N/A NED 4 unknown43 70 IIIC yes 5 AWD 20 screen negative44 35 IIIC no N/A NED 6 unknown45 79 IIIC no N/A NED 5 unknown46 77 IIIC no N/A NED 6 unknown47 45 IIIC no N/A NED 4 unknownTable 3.1: Studied patients and samples. Age refers to age (in years) at diagnosis. Recurrence-freesurvival (RFS) and overall survival (OFS) are indicated in months. BRCA status was determinedthrough clinical testing. Current disease status: NED, no evidence of disease; AWD, alive withdisease; DOD, dead of disease. (a): BrnM and BrnMA1 14 months, RPvM and BwlImA6 33 monthspost-diagnosis; (b): Pv1, Rct1, Rct2 139 months post-diagnosis; (c): LOv1 14 months post-diagnosis84Figure 3.2: Schematic Diagram Depicting Sample Collection, Experimental Modalities, andAnalysis Workflows Applied to the Data.3.3.2 Tumor-Infiltrating Lymphocyte Subtypes Reveal Extensive Intrapa-tient Variation in Immune Responses across Peritoneal SitesWe began by profiling 188 tumor samples from 37 patients with multicolor IHC for CD8+T cells (CD3+CD8+), CD4+ T cells (CD3+CD8-), CD20+ B cells (CD20+), and plasmacells (CD79a+CD138+). All but three patients were surveyed at multiple sites, providing anunprecedented view of intrapatient spatial variation. CD8+ T cells were the most abundant TILtype (0-1125.65 cells per high-powered field [HPF], median: 53.08), while CD20+ B cells werethe rarest (0-136.77 cells per HPF, median: 2.74). Densities of all TIL types were correlated(Figure 3.3), with extensive variation across the cohort (Figure 3.3).85Figure 3.3: (A) Correlations between overall TIL densities. Color indicates Spearmans , P-valuesshown inside each cell. (B) Overall CD8+, CD4+, CD20+, and plasma cell densities across thecohort. Bars colored by patient.Using TIL densities as input features, we first analyzed a discovery cohort of 119 samplesfrom 20 patients. Hierarchical clustering revealed three major TIL subtypes: N-TIL (tumorssparsely infiltrated by TILs), S-TIL (tumors dominated by stromal TILs), and ES-TIL (tumorswith substantial levels of both epithelial and stromal TILs) (Figure 3.1 and SupplementalTable A.2). Based on orthogonal Nanostring probe counts, gene expression values for immune-associated pathways, including cytotoxicity, cytokines, and T cell- and B cell-associated genes,86were comparable between S-TIL and ES-TIL but lower in N-TIL (Figure 3.1). The threeTIL subtypes mapped to previously described gene expression subtypes (C1, C2, C4, and C5)of HGSC [150]. N-TIL was enriched for C4 and C5 tumors (p < 10−5, Fishers exact test),while S-TIL was overrepresented for C1 tumors (p < 0.01, Fishers exact test) and ES-TILfor C2 tumors, respectively (p < 10−5, Fishers exact test; Figure 3.1 and SupplementalTable A.2), suggesting previously reported HGSC gene expression subtypes [14, 101] largelyreflect immune cell content. We analyzed IHC data from an additional cohort of 69 samplesfrom 17 patients and observed a similar N-TIL, S-TIL, and ES-TIL distribution (Figure 3.1),indicating reproducibility of the TIL subtypes. Among patients with ≥2 treatment-naive samples,14 of 31 patients harbored only one TIL subtype: seven were N-TIL only, six were ES-TIL only,and one was S-TIL only. The remaining 17 of 31 patients harbored tumors from more thanone TIL subtype (Figure 3.1), and five patients harbored samples from all three subtypes,indicating extensive variation in immune response within patients.While the ES-TIL pattern suggests active cytolytic TIL response against tumor cells, thepresence of TILs in an epithelial region does not necessarily indicate active engagement withmalignant cells. We therefore used histologic image analysis to profile microscopic spatialrelationships between cancer cells and TILs. For each sample, we leveraged hematoxylinand eosin (H&E) images to identify cancer cell and lymphocyte “hotspots” within the tumorepithelium–i.e., regions of local aggregation relative to epithelial cellular density (Figure 3.4).We computed three measures of cancer-lymphocyte hotspot colocalization [167]: fC (the fractionof cancer cell hotspots that are lymphocyte hotspots); fI (the fraction of lymphocyte hotspotsthat are cancer cell hotspots), and fCI (fractional tissue area occupied by colocalized cancer-lymphocyte hotspots) (Supplemental Table A.2). ES-TIL tumors exhibited high levels ofoverlap between cancer and lymphocyte hotspots, while S-TIL samples contained relativelylow overlap (all p < 0.05, Kruskal-Wallis test, Figure 3.4). Thus, in S-TIL tumors, the rareimmune cells that enter epithelial compartments appear to fail to engage with tumor cells,possibly due to lack of recognition. Although N-TIL tumors have negligible levels of TIL, theynonetheless showed occasional immune cells that could be evaluated by hotspot analysis. Wheremeasureable, N-TIL tumors showed similar levels of colocalization as ES-TIL (Figure 3.4).87Figure 3.4: (A) Epithelial cancer cell and lymphocyte hotspots for representative N-, S-, andES-TIL examples. (B) Histology of a cancer cell hotspot. Yellow arrow: cancer cell. (C) Histologyof a colocalized cancer-lymphocyte hotspot. Blue arrow: lymphocyte. (D) Comparison of epithelialcancer-lymphocyte hotspot colocalization between TIL subtypes. P -values from Kruskal-Wallis test;post hoc comparisons (Benjamini-Hochberg adjusted) from Dunn’s test. Whisker ends correspond toQ1 -1.5*IQR and Q3 +1.5*IQR.883.3.3 Evidence for Purifying Malignant Clonal Selection at Tumor Sites withHigh Epithelial Lymphocyte InfiltrationWe next evaluated whether regional variation in TIL subtypes provided insight into the evolution-ary trajectories and dissemination patterns of malignant clones. Using WGS on cryopreservedtissues (102 samples from 21 patients, of which 31 from 7 patients were previously described in[29]), we profiled somatic single-nucleotide variants (SNVs), allele-specific copy number, andrearrangements (Supplemental Table A.2) as markers of malignant clones. In addition, weperformed deep amplicon sequencing on 97 samples from 14 of these patients (66 frozen and31 formalin-fixed samples) to calculate clonal phylogenies and the clonal composition of eachsample (Figure 3.5). We then related quantitative attributes of malignant clone compositionto the N-TIL, S-TIL, and ES-TIL subtypes.89Figure 3.5: Patients are ordered by significance of the association between BCR repertoireand clonal composition dissimilarities. Chords denote shared clonotypes, width proportional toclonotype count, colored by publicity (number of samples containing a clonotype). Shared clonotypes:publicity≥2, private clonotypes: publicity = 1. Arc length (along circumference) is proportionalto total clonotype count. Tumor clone composition and phylogenies shown external to each circle.Samples without BCR-seq data shown separately below each circle. TIL subtypes indicated by N(N-TIL), S (S-TIL), and ES (ES-TIL) labels. Uncorrected Mantel’s test P -values between BCRrepertoire dissimilarity and clonal dissimilarity shown below patient labels.90For each sample, we computed three continuous measures of malignant clone complexity:mixture entropy (the mixture distribution of clones present within a sample), clone divergence(the maximum phylogenetic distance between clones present within a sample; see [29]), andheterogeneity index (the mean phylogenetic distance between a randomly selected pair ofclones within a sample, weighted by abundance). We also computed an orthogonal measurefrom WGS directly with copy-number analysis ([156]; Supplemental Table A.2). All fourmeasures of ITH were correlated (all p ¡ 0.1, significance of Spearman ρ; Figure 3.6). Forquality control, we confirmed entropy, clone divergence, and heterogeneity index were notcorrelated with tumor purity (all p > 0.2; Figure 3.6). We evaluated the associations betweenmeasures of malignant clone complexity and the three TIL subtypes over all treatment-na¨ıvesamples. ES-TIL samples were lower for all four ITH measures relative to S-TIL and N-TILsamples (Figure 3.7; accounting for tumor purity in the subclonal copy-number comparison)with mixture entropy, heterogeneity index, and subclonal copy number statistically significant.Accordingly, clonally pure tumors had the highest epithelial CD8+ TIL densities (Figure 3.6).Despite the association between TIL and ITH, clonal similarity between intrapatient siteswas not associated with TIL subtype (p > 0.3, nested ranks test; Figure 3.6). For example,omentum sites 1 and 2 from patient 17 had comparable clonal composition, while ovary site 1contained different clones (Figure 3.5); however, omentum site 1 was ES-TIL subtype, whereasomentum site 2 and ovary site 1 were N-TIL subtype (Supplemental Table A.2). Together,these data are consistent with epithelial TIL abundance as a negative determinant of regionalmalignant clonal complexity.91Figure 3.6: (A) Correlations between ITH measures. Asterisks indicate significance of Spearmanscorrelation (legend shown in D). (B) Correlations between tumor cellularity and ITH. P -values ofSpearmans correlation are shown. (C) Epithelial CD8+ TIL densities for pre-treatment samples,stratified by clonal mixture type. P-value from the Kruskal-Wallis test shown. Whisker endscorrespond to Q1 - 1.5*IQR and Q3 + 1.5*IQR. Significance of post hoc Dunn’s test shown (legendin D). (D) Degree of similarity in tumor clone composition for pre-treatment samples with differentor identical TIL subtypes. Subtype comparisons were made within patients; mean similarity acrossall comparisons was used. Lines connect comparisons made within the same patient. Nested rankstest P -value is shown. Whisker ends correspond to Q1 -1.5*IQR and Q3 +1.5*IQR. (E) Correlationmatrix between TIL densities and Nanostring-derived expression of inhibitory immune checkpointgenes. Asterisks indicate significance of Spearmans correlation.9293Figure 3.7: (A) Clonal measures of ITH by TIL subtype. p values from Kruskal-Wallis tests;asterisks indicate post hoc significance (Benjamini-Hochberg adjusted) from Dunns test. Whiskerends correspond to Q1 -1.5*IQR and Q3 +1.5*IQR. (B) Subclonal copy-number proportion by TILsubtype. P value from ANOVA, controlling for cellularity. Asterisks indicate post hoc significancefrom Tukey’s range test. Whisker ends correspond to Q1 -1.5*IQR and Q3 +1.5*IQR. (C) Ratiobetween observed and expected neoantigen rates for pre-treatment samples in patients with highestsample-level epithelial CD8+ densities (indicated by bar color). (D-G) For patients with subclonalHLA class I LOH, (left) clonal phylogeny showing HLA LOH events and (right) logR values forsamples with and without HLA LOH based on clonal composition. a:RPvM and LOv1 did not haveIHC data. (Bottom) Clonal composition and epithelial CD8+ density of each sample. (D) Patient1. (E) Patient 15. (F) Patient 7. HLA-C*07:01:01:01 was not as visually depleted in RPvM dueto low cellularity (38%). (G) Patient 13. Sample labels defined in Supplemental Table A.2. (H)Expression of lymphocyte markers in cases with none or any HLA LOH for [3] (Nanostring) andOV-AU (RNA sequencing) cohorts. P values from nested ranks test. Whisker ends correspond toQ1 -1.5*IQR and Q3 +1.5*IQR.The negative association between epithelial TIL densities and malignant clone diversity could beexplained by clonally complex tumors suppressing development of ES-TIL microenvironmentsand/or tumor clones undergoing immune-mediated purifying selection in the presence of highepithelial TIL density. In the latter scenario, subclonal (non-ancestral) neoepitopes might serveas targets of T cell recognition and hence show evidence of depletion at ES-TIL sites. To test this,we used NetMHCpan [164] to computationally predict neoepitopes from nonsynonymous somaticSNVs (Supplemental Table A.3), categorizing each neoepitope as clonal or subclonal throughphylogenetic analysis. For each sample, we then quantified neoantigen depletion by comparingobserved to expected (computed on an independent cohort of 121 primary HGSC samples)neoantigen rates. Within patients, samples with higher epithelial CD8+ density exhibited higherlevels of subclonal neoantigen depletion (lower observed/expected subclonal neoantigen rate,p = 0.09, linear mixed model; Supplemental Table A.3), but not clonal neoantigen depletion(p > 0.3), compared to other samples from the same patient. This association was pronouncedin patients containing samples with the highest epithelial CD8+ TIL densities (p = 0.001, linearmixed model; Figure 3.7). In contrast, no significant association was observed between stromalCD8+ TIL density and clonal or subclonal neoantigen depletion (all p > 0.2, linear mixedmodel). Thus, samples with high epithelial CD8+ TILs show evidence of immune editing ofsubclonal neoantigens, raising the possibility that immune-driven purifying selection underliesthe observed reduction in malignant cell diversity at TIL-rich sites.In tumors with high epithelial CD8+ TIL densities, we postulated that the few remainingtumor clones might have avoided immune-related negative selection through clonal expansion ofcells lacking neoantigen- or other tumor antigen-presenting HLA alleles. We used a Bayesianstatistical extension of the LOHHLA algorithm [52] to analyze WGS data for clone-specific HLA94class I allele loss. Of 14 patients evaluated, we identified four patients harboring clonal HLALOH and four with subclonal HLA LOH (one patient had both; Supplemental Table A.4).In three out of four patients with subclonal HLA LOH, the samples with the highest epithelialCD8+ TIL densities contained tumor clones with subclonal HLA LOH (Figure 3.7), includingtwo of the patients (1 and 15) that demonstrated subclonal neoantigen depletion. An exceptionwas patient 13, where subclonal HLA LOH was observed despite all samples having low epithelialCD8+ TIL density (Figure 3.7; no samples were ES-TIL). Nevertheless, these findings suggestthat tumor clones at ES-TIL sites have, in some cases, escaped immune clearance by somaticgenomic loss of HLA haplotypes. We next examined the prevalence of HLA LOH in orthogonalWGS external cohorts [3, 158]. HLA LOH was found in 33.3% of samples (OV-AU: 34.7%,Wang: 32.1%) and was associated with significantly higher expression of lymphocyte markers(Figure 3.7), establishing a link between HLA LOH and higher TIL levels.To provide context, we also considered other known mechanisms of immune escape, includinganatomic site, disruption of antigen presentation machinery [168], and expression of immuno-suppressive factors [165, 169]. TIL subtype was not significantly associated with any specificanatomic location (Fishers exact test, p > 0.05, Supplemental Table A.2), and no pointmutations, indels, or copy-number losses in antigen presentation machinery molecules wereobserved in ES-TIL samples. However, consistent with expectation from previous reports [165],we found that inhibitory immune checkpoint molecules were generally upregulated in tumorswith high epithelial CD8+ TIL density (Figure 3.6).3.3.4 T Cell, but Not B Cell, Clonotypes Show Evidence of Tumor CloneTrackingWe next investigated whether T and B cell clonotypes associate with tumor clones. We appliedTCR β chain and BCR heavy-chain sequencing to total RNA from 116 samples (27 patients) anddefined the clonotype-level composition of T and B cell populations in each sample (Figure 3.8and Figure 3.9, Supplemental Table A.2). TCR diversity was positively correlated withIHC-based CD8+ and CD4+ TIL densities (all Spearman p < 10−5; Figure 3.9). Similarly,BCR diversity was positively correlated with CD20+ and plasma cell densities (all Spearmanp < 0.01; Figure 3.9). S-TIL and ES-TIL tumors had significantly more diverse TCR and BCRrepertoires than N-TIL tumors (Figure 3.8 and Figure 3.9) and a higher proportion of rareclonotypes (Figure 3.9). None of the four ITH measures were significantly associated withTCR or BCR diversity across treatment-naive samples (all Spearman p > 0.3), indicating thatdiverse malignant populations do not recruit similarly diverse TIL repertoires.95Figure 3.8: (A) Number of unique TCR clonotypes, prevalences of top 10 clonotypes (gray = allothers), and CD8+ and CD4+ TIL density for each sample. (B) Comparison of unique TCR andBCR clonotype counts between TIL subtypes. p values from Kruskal-Wallis tests; asterisks indicatepost hoc significance (Benjamini-Hochberg adjusted) from Dunn’s test. (C) Distribution of pairwiseTCR similarity for each patient. Whisker ends correspond to Q1 -1.5*IQR and Q3 +1.5*IQR. (Dand E) Scatterplot of mean intrapatient TCR similarity and (D) CD8+ TIL density and (E) CD4+TIL density. P value of Spearman ρ shown. (F) Mean repertoire broadness for CD8+ and CD4+type clonotypes in each patient. P value from Wilcoxon signed-rank test. Post-treatment tumorsexcluded in (B), (C), (D), (E), and (F).9697Figure 3.9: (A) Unique BCR clonotype count, relative frequencies of the top 10 BCR clonotypes(gray = all other clonotypes), CD20+ and plasma TIL density for each sample. (B and C) Correlationsbetween (B) overall CD8+ and CD4+ densities and TCR diversity; (C) overall CD20+ and plasmadensities and BCR diversity. Diversity was quantified as the (1) number of unique clonotypes and(2) the entropy of the clonotype abundance distribution. BH-adjusted P -values of Spearmans areshown. (DF) TCR and BCR repertoire diversity across TIL subtypes. Diversity was measured by(D) Shannon entropy of clonotype prevalences, (E) Efron-Thisted index, and (F) D50 index. Whiskerends correspond to Q1 - 1.5*IQR and Q3 + 1.5*IQR. P -value from Kruskal-Wallis tests shown;asterisks indicate post hoc Dunn’s test significance. (G) Correlation between mean intrapatient TCRand BCR repertoire similarity. Spearman correlation P-value is shown. (H) Correlation betweenmean intrapatient BCR repertoire similarity and CD20+ TIL density. P -value of Spearman ρis shown. (I) Correlation between mean intrapatient BCR repertoire similarity and plasma celldensity. P -value of Spearman ρ is shown. (J) Consistency between CD8+/CD4+ ratios fromimmunohistochemistry and from TCR-based prediction. P -value of Spearman ρ is shown.We next ascertained the degree of homogeneity (similarity) between TCR and BCR repertoiresacross spatial samples within patients. This revealed marked variation in both intrapatient TCRand BCR similarity across the cohort (Figure 3.8 and Figure 3.9). Considering patients withat least three samples, the extent of intrapatient TCR and BCR repertoire similarities werecorrelated (Spearman p < 0.1), but with notable exceptions (Figure 3.9). Patient 15 had highTCR similarity (ranked 2nd out of 20 patients), but not BCR similarity (14th), while patients 10and 21 had high BCR similarity (3rd and 5th), but not TCR similarity (15th and 20th). Meanintrapatient BCR similarity was not significantly correlated with IHC-based CD20+ or plasmacell density (all Spearman p > 0.2, Figure 3.9). However, mean intrapatient TCR similarity wasstrongly associated with CD8+ (Spearman p < 0.01), but not CD4+, TIL density (Figure 3.8),suggesting that CD8+ TILs were more broadly distributed (shared) across tumor sites comparedto CD4+ TILs. To test this, we trained a classifier to separate TCRs as CD8+ type or CD4+type on the basis of V/J genes and physicochemical properties of the hypervariable domain.The ratio of CD8+-/CD4+-type TCRs was correlated with the ratio of CD8+/CD4+ densitiesby IHC (Spearman p < 0.01; Figure 3.9). Corroborating our predictions, CD8+-type TCRswere significantly more broadly distributed than CD4+-type TCRs (p < 0.001; Figure 3.8).Having established that TCR-/BCR-based immune profiles vary across space, we asked how thisvariation is related to the spatial distribution of tumor clones. Pairwise T cell repertoire similaritywas significantly correlated with malignant clone composition similarity in 7 out of 13 patients(Figure 3.10). Importantly, this relationship was significant in 5 of 6 patients with the highestepithelial CD8+ TIL densities (patients 1, 2, 9, 15, and 17), consistent with T cell clonotypesspatially tracking tumor clones in patients with high epithelial CD8+ TILs. This associationheld in the same six patients when considering only major TCR clonotypes (most abundantclonotypes constituting the top 50% of reads within each patient), but was only significant in98patients 2, 9, and 12 when considering minor clonotypes (all other clonotypes), indicating thatthe most abundant clonotypes drove this effect. In contrast, pairwise BCR similarity was notsignificantly correlated with tumor clone similarity in any patient (Figure 3.5), suggesting anabsence of spatial tracking between B cells and tumor clones.99Figure 3.10: Patient 2 aside, cases ordered by significance of association between TCR repertoireand clonal composition dissimilarities (uncorrected Mantel’s test p values). Chords denote sharedclonotypes, width proportional to clonotype count, colored by publicity (number of samples containinga clonotype). Shared: publicity ≥ 2; private: publicity = 1. Purple arrow: chord denoting clonotypesshared only between right ovary sites 1 and 2. Green arrow: clonotypes shared only between omentumsites 1 and 2. Tumor clone composition and phylogenies next to each circle. TIL subtypes indicatedas N (N-TIL), S (S-TIL), and ES (ES-TIL). Patient 7 excluded, as only two samples had TCR andtumor clone data.1003.3.5 Mutation Signatures Prognostically Associate with Patient-Level Im-munologic FeaturesWe next investigated interaction of malignant and immune infiltration from the perspective ofmutational processes operating in HGSC. We previously identified two prognostically relevantmutation signature-associated subtypes: H-HRD and H-FBI [3]. Here, we explored whetherthose subtypes could explain the observed variation in immune infiltration within and betweenpatients. We pooled WGS data from our 21 cases with 195 additional single-site ovarian cancercases (133 from [3] and 62 from OV-AU in the International Cancer Genome Consortium [ICGC])and applied a novel multimodal correlated topic model (MMCTM) [26], identifying six SNV andseven rearrangement signatures (Figure 3.11 and Supplemental Table A.5). Hierarchicalclustering by signature proportions identified four major clusters (Figure 3.12, SupplementalTable A.5): one subtype (HRD-DEL) dominated by the point mutation signature associatedwith homologous recombination deficiency (P-HRD) along with a short deletion signature(R-SDEL) associated with BRCA2 mutations [159], a second subtype (HRD-DUP) with P-HRDand a short tandem duplication signature (R-SDUP) associated with BRCA1 mutations [159], athird subtype (FBI) characterized by an FBI rearrangement signature (R-FB) associated withbreakage-fusion-bridge [3], and a fourth, minor subtype distinguished by medium and largetandem duplications (TDs) (R-MDUP and R-LDUP, respectively) associated with CDK12 pointmutations [25, 26].101102Figure 3.11: (A and B) Jointly inferred SNV and rearrangement signature profiles from MMCTM.Point mutation signatures: P-AGE, age associated; P-HRD, homologous recombination deficiency;P-APOBEC, APOBEC associated; P-MMR-1 + P-MMR-2, mismatch-repair associated. P-UM:ultramutator-associated (virtually absent in HGSC). Rearrangement signatures: R-TDUP, tandemduplications; R-SDUP, short duplications; R-MDUP, medium length duplications; R-LDUP, longduplications; R-SDEL, short deletions; R-MDEL, medium length deletions; R-FB, foldback inversions;R-TR, translocations. Pentanucleotide contexts are shown for each SNV signature and relativeprevalences of deletions, duplications, inversions, foldback inversions, and translocations are shownfor each rearrangement signature. For rearrangements, microhomology length is labeled. (C)Standardized proportions of each mutation signature for multisite HGSC, OV-AU, and HGSC [3]samples, showing clustering of samples from the same patient. (D) Standardized proportions of eachmutation signature for multisite HGSC, OV-AU, and HGSC [3] samples, where only non-ancestralmutations were considered for the multisite HGSC cohort. Heatmap values were clipped between 4and 4.103104Figure 3.12: (A) Signature proportions in HGSC cases standardized and clipped from 4 to 4.Dendrogram computed with Ward’s method on Pearson correlation dissimilarities. ITH: multisitecohort from this study. (B) Fractions of ES-none, ES-mixed, and ES-pure patients across mutationalsubtypes. (C) Expression of select immune-associated pathways across mutational subtypes inOV-AU. P values (Benjamini-Hochberg adjusted) from Kruskal-Wallis test; asterisks indicate posthoc significance (Benjamini-Hochberg adjusted) from Dunn’s test. Survival analysis of 433 TCGApatients. Whisker ends correspond to Q1 -1.5*IQR and Q3 +1.5*IQR. (D) Hazard ratios, 95%confidence intervals, and P values from Cox regression of overall survival. Interaction terms indicatedby colons; e.g., CTX:No AMP: effect of cytotoxicity in No AMP subtype. (E) Differences in overallsurvival between FBI-HLAMP subgroups for tumors with low/high cytotoxicity. P values fromlog-rank test.Using this grouping of samples, we asked how immune response characteristics co-segregatedwith mutational signatures. Unlike TIL subtypes, mutational subtypes were largely invariantwithin patients (Figure 3.11), indicating that mutational processes cannot explain intrapatientheterogeneity in TIL subtypes. We next asked whether mutational subtypes related to themixture of TIL subtypes within each patient. Focusing on the ES-TIL subtype, we categorizedpatients with multi-sample IHC data as ES-none (no ES-TIL samples), ES-mixed (both ES-TILand N-TIL/S-TIL samples), or ES-pure (all samples ES-TIL). The HRD subtypes contained theonly three ES-pure patients (out of 12 HRD patients), although this did not reach significancewith respect to the other mutational subtypes (Fishers exact test, p = 0.23; Figure 3.12).Expression values of immune-associated pathways [149] for 54 OV-AU cases revealed thatcytotoxicity, antigen processing, cytokine, and T cell markers were highest among HRD tumors(Figure 3.12), concordant with similar findings in ER+ breast cancer [170] and among BRCA1-mutated tumors in HGSC [34]. Relative to HRD tumors, TD tumors had similar expressionof immune markers, whereas FBI tumors were significantly depleted of these (Figure 3.12).Corroborating these findings, differential expression analysis of OV-AU cases revealed thatantigen processing, TCR/BCR signaling, cytotoxicity, and cytokine pathways were upregulatedin HRD and TD relative to FBI (Q < 0.01), while none of these were differentially expressedbetween HRD and TD (Figure 3.13 and Supplemental Table A.6).105Figure 3.13: Pathway annotations derived from KEGG. Fold change indicated as, e.g., HRDversus TD: mean log2 fold change in expression from HRD to TD (> 0 = higher in HRD). Q-valuescomputed with the BH procedure; Q < 0.01 was considered significant. Selected immunologicpathways are highlighted; significant hits are additionally labeled.Colocalized foldback inversions and focal high-level amplifications (HLAMPs), thought to bereflective of breakage-fusion-bridge, have been associated with poor outcomes in HGSC [3]. Weasked whether immune activity could be used to further stratify foldback-enriched tumors into106subgroups with distinct survival outcomes. Using gene expression data for 433 ovarian cystade-nocarcinoma cases from the Cancer Genome Atlas (TCGA [14]; Supplemental Table A.7),we jointly modeled the effects of colocalized foldback-HLAMP events and cytotoxicity expressionwith a Cox proportional hazards model, controlling for age of diagnosis and therapeutic regimen.In agreement with [3], high levels of colocalized foldback-HLAMP events were associated withsignificantly shorter overall survival (hazard ratio: 1.64, 95% CI: 1.06–2.52, p < 0.05; Fig-ure 3.12). The association between cytotoxicity and survival differed between FBI-HLAMPgroups (p < 0.05, likelihood ratio test between Cox models with and without cytotoxicity ×FBI-HLAMP interaction). In cases with no HLAMP events, cytotoxicity was significantly associ-ated with a decreased hazard ratio (0.52, 95% CI: 0.29–0.92, p < 0.05; Figure 3.12). However,among cases with colocalized foldback-HLAMP events, the hazard ratio for cytotoxicity wasnot significant (FBI-AMP low: 0.97, 95% CI: 0.71–1.34, p > 0.3; FBI-AMP high: 1.24, 95% CI:0.91–1.69, p > 0.1; Figure 3.12), suggesting that HLAMP-positive foldback-containing tumorsharbor prognostic effects that are independent of immune response. We then median-stratifiedcases into low- and high-cytotoxicity groups. Low FBI was associated with significantly longeroverall survival among tumors with high cytotoxicity (log-rank p < 0.05; Figure 3.12), but notlow cytotoxicity (log-rank p > 0.2; Figure 3.12). Together, the covarying effects of immuneactivity and mutational processes suggest a combinatorial prognostic effect with high immuneactivity and low prevalence of FBIs leading to the best outcomes, while FBI-bearing patientshave poor outcomes even in the presence of high immune activity.3.4 DiscussionOur results illuminate evolutionary properties at the malignant-immune interface of HGSC. Inpatients with the highest epithelial TIL densities, our data are consistent with active pruning ofmalignant cell diversity by TIL through subclonal neoepitope recognition, resulting in expansionof clones harboring neoantigen loss and/or HLA LOH. The underlying mechanism likely involvestracking of tumor clones across peritoneal space by T cell clones, but not B cell clones. As such,immune infiltrates impose selective constraints, shaping patterns of malignant spread and clonaldiversity in HGSC. Our findings do not exclude the possibility that T cells can also recognizeclonal neoepitopes [171]; however, subclonal neoepitopes, which have been reported to havehigher predicted immunogenicity than clonal neoepitopes [147], may be under stronger negativeselection. Moreover, depletion of clonal neoantigens could result in complete tumor eliminationand therefore go clinically undetected.The presence of extensive intrapatient immune variation prior to treatment highlights potential107shortcomings of prognostic stratification and study of the immune microenvironment fromsingle biopsies. The widespread multi-site variation we observed suggests that even a single siteharboring relative immune privilege may be sufficient to engender resistant disease, regardless ofactive immune responses in distal intraperitoneal regions. We suggest immunologically shelteredhavens may plausibly act as reservoirs of clonal diversity from which malignant clones impactingdisease relapse might emerge. As a preliminary illustrative example, ES-pure patients had betteroutcomes (5 of 6 no evidence of disease [NED] or alive with disease [AWD], 5 of 6 platinumsensitive, median progression-free survival [PFS] for relapsed patients was 19 months) thanES-mixed and ES-none patients (8 of 11 and 11 of 14 NED or AWD, 7/9 and 7/11 platinumsensitive, median PFS for relapsed patients was 9.3 and 7.1 months, respectively).Our data show for the first time a prognostically relevant interaction between mutationalprocesses and immune response in HGSC. Notably, foldback inversions associate with pooroutcomes, even in highly cytotoxic tumor microenvironments. Thus, in contrast to pointmutations resulting from mismatch repair deficiency [48], FBIs likely represent a class of non-immunogenic genomic aberrations. Conversely, our findings also provide context for explainingsuperior outcomes observed in BRCA1- and BRCA2-mutated HGSC [34]. In contrast to previousreports that BRCA1 disruption, but not BRCA2 disruption, is associated with elevated TILs[34, 172], we observe comparably high immune activity between BRCA1-associated (HRD-DUP), BRCA2-associated (HRD-DEL), and TD subtypes. Shared deficiencies in homologousrecombination between HRD and TD subtypes [173] may result in patterns of rearrangementsor point mutations [3] responsible for eliciting these immune responses [170].Our study provides context for clinical trials investigating various classes of immunotherapyin ovarian cancer (e.g., immune checkpoint blockade, adoptive T cell transfer, neoepitopevaccination, combination immunotherapy with PARP inhibition). A recent case study trackingimmune response over time in a HGSC patient with a remarkable clinical trajectory [147]demonstrated that spatiotemporal variation of the immune microenvironment relates specificallyto treatment sensitivity of malignant clones. We reveal that immune-microenvironment spatialvariation exists prior to treatment and is prevalent in the HGSC patient population. Given thatefficacy of PD-1 axis blockade hinges on pre-existing adaptive immunity [174], immunologicallyprivileged sites on an otherwise highly infiltrated background may explain the limited successof immunotherapy in HGSC to date [49, 50]. While some tumors contain abundant TILs,lack of cancer cell-lymphocyte colocalization and reduced tumor-immune engagement in S-TILsites may result from a failure of immune recognition or region-specific barriers to infiltration.Consequently, TIL abundance alone is an insufficient predictor of active immune response. Evenat sites patterned by extensive epithelial TILs, neoantigen depletion and apparent positive108selection of clones harboring HLA LOH may render checkpoint blockade ineffective.Despite these challenges, our findings inform on several potential therapeutic solutions. WhileFBI cases exhibit poor prognostic profiles independently of immune properties, HRD cases,typically associated with fewer foldback inversions, likely represent optimal candidates forimmunotherapy approaches. Thus, mutational processes considered in conjunction with immuneproperties will aid in interpretation of newly initiated clinical trials examining combinationPARP inhibition with checkpoint-blockade approaches. Furthermore, if obstacles to infiltrationat immunologically privileged sites can be surmounted, our findings hint at the tantalizingpotential that such tumor sites may represent targetable cancer cell populations, owing to theirlimited neoantigen and HLA depletion at baseline.As the cancer evolution field progresses toward a more rigorous understanding of the fitness ofheterogeneous clones within disease spectra and over temporal dimensions [175], external selectivepressures imposed by the immune system must be considered as highly relevant factors. Here weshow that high-resolution measurement of the immune microenvironment together with clonaldecomposition analysis is tractable and yields novel insight into forces shaping malignant celldiversity and intraperitoneal spread. Broadly disseminated intraperitoneal disease at diagnosisin HGSC remains a formidable clinical problem. Our study informs on how regional variationat the interface of immunological and cancer cells controls dissemination and diversification ofclones and simultaneously identifies microenvironmental and malignant cell properties to exploitin future immuno-oncologic therapeutic strategies for HGSC.109Chapter 4Probabilistic cell type assignment ofsingle-cell transcriptomic datareveals spatiotemporalmicroenvironmental dynamics inhuman cancers4.1 IntroductionGene expression observed at the single-cell resolution in human tissues enables studying the celltype composition and dynamics of mixed cell populations in a variety of biological contexts,including cancer progression. Cell types inferred from single-cell RNA-seq (scRNA-seq) data aretypically annotated in a two-step process, whereby cells are first clustered using unsupervisedalgorithms and then clusters are labeled with cell types according to aggregated cluster-levelexpression profiles [176]. Myriad methods for unsupervised clustering of scRNA-seq have beenproposed, such as SC3 [177], Seurat [178], PCAReduce [179], and PhenoGraph [180], along withstudies evaluating their performance across a range of settings [181, 182]. However, clusteringof low-dimensional projections may limit biological interpretability due to i) low-dimensionalprojections not encoding variation present in high-dimensional inputs [183] and ii) overclusteringof populations that are not sufficiently variable.Furthermore, even in the context of robust clustering which recapitulates biological cell statesor classes, few principled methods for annotating clusters of cells into known cell types exist. Incontrast to unsupervised statistical frameworks, this latter step is a supervised, or classificationproblem. Typical workflows employ differential expression analysis between clusters to manuallyclassify cells according to highly differentially expressed markers, aided by recent databases110linking cell types to canonical gene-based markers [184]. In situations where investigators wishto identify and quantify specific cell types of interest with known marker genes across multiplesamples or replicates, such workflows can be cumbersome, and differences in clustering strategiescan affect downstream interpretation [181]. Alternatively, cell types may be assigned by gatingon marker gene expression, but this strategy is difficult to implement in practice as (i) gatingis difficult for more than a few genes and relies on knowledge of marker gene expression levelsand (ii) cells that fall outside these gates will not be assigned to any type, rather than beingprobabilistically assigned to the most likely cell type.Another approach to cell type annotation is to leverage ground-truth single-cell transcriptomicdata from labeled and purified cell types to establish robust profiles against which new data canbe compared and classified. For example, scmap-cluster [185] calculates the medioid expressionprofile for each cell type in the known transcriptomic data, and then assigns input cells based onmaximal correlation to those profiles. However, this approach requires existing scRNA-seq datafor purified cell populations of interest. Given the technical effects associated with differences inexperimental design and processing, expression profiles for reference populations may not bedirectly comparable to those for other single-cell RNA-seq experiments [186].We assert that statistical cell type classification approaches leveraging prior knowledge in theliterature (or from experiments) will be an effective complement to unsupervised approaches forquantitative decomposition of heterogeneous tissues from scRNA-seq data. Therefore, to addressthe analytical challenges inherent in both clustering and mapping approaches, we developedCellAssign, a scalable statistical framework that annotates and quantifies both known and denovo cell types in scRNA-seq data. CellAssign automates the process of annotation by encodinga set of a priori marker genes for each cell type. The statistical model then classifies the mostlikely cell type for each cell in the input data, using a marker gene matrix (cell type-by-gene).The model allows for flexible expression of marker genes, assuming that marker genes are morehighly expressed in the cell types they define relative to others. Implemented in Google’sTensorflow framework, CellAssign is highly scalable, capable of annotating thousands of cells inseconds while controlling for inter-batch, patient and site variability. We evaluated CellAssignacross a range of simulation contexts and on ground truth data for FACS-purified H7 humanembryonic stem cells (HSCs) at various differentiation stages [187], showing that CellAssignoutperforms both clustering and correlation based methods—more readily discriminating closelyrelated cell types—and is robust to errors in marker gene specification. In addition, we appliedCellAssign to two novel datasets generated to profile spatiotemporal tumor microenvironment(TME) dynamics in human cancers. Using the CellAssign approach, we demonstrated tumor‘ecosystem’ spatial diversity in untreated high-grade serous ovarian cancer through variable111composition in stromal and immunologic cell types comprising the TME and variation in keypathways across malignant cell populations including immune evasion, epithelial-mesenchymaltransition and hypoxia. Temporal dynamics were also exemplified using the CellAssign approach.We generated scRNA-seq libraries from matched diagnostic and relapsed pairs of follicularlymphoma samples, with one case having undergone histologic transformation to an aggressivelymphoma. We show compositional and phenotypic changes, including T-cell activation andHLA downregulation in cancer cells upon transformation, pointing towards an evolutionaryinterplay with cancer cells escaping immune recognition following transformation. In aggregatewe conclude the CellAssign approach provides a robust new statistical framework through whichdisease dynamics in tissues comprised of mixed cell populations can be quantified and interpretedto ultimately uncover new properties and understanding of disease progression.4.2 Methods4.2.1 The CellAssign model4.2.1.1 Model descriptionLet Y be a cell-by-gene expression matrix of raw counts for N cells and G genes. Supposeamong those cells we have C total cell types, each of which is defined by high expression ofseveral “marker” genes. We encode the relationship between cells and marker genes through abinary matrix ρ, where ρgc = 1 if gene g is a marker for cell type c and 0 otherwise. To relatecells to cell types, we introduce an indicator vector z = {zn} that encodes to which of the C celltypes each cell belongs -zn = c if cell n of type c.In order to assign cells to cell types we perform statistical inference of the probability that eachcell is of a given cell type for which we must compute the quantity p(zn = c|Y, Θˆ), where Θˆ arethe MAP estimates of the model parameters.Let sn be the size factor for cell n and X be a P ×N matrix of P covariates (such as patient oforigin). Then our model isE[yng|zn = c] = µngc112whereMean log expression︷ ︸︸ ︷logµngc = log sn︸ ︷︷ ︸Cell size factor+Cell type︷ ︸︸ ︷δgcρgc + βg0︸︷︷︸Base expression+Other covariates (incl. batch)︷ ︸︸ ︷P∑p=1βgpxpnwith the constraint that δgc > 0.The intuition here is that we expect the expression of gene g in cell type c is multifactorial,influenced by a cell type specific factor δgc if gene g is a marker for cell type c, combined withcovariates expected from batch effects and other arbitrary sources. In this way we put norestriction that marker genes can’t be expressed in other cell types and that they must be highlyexpressed in their cell type, only that they exhibit higher expression in the cells of type forwhich they are a marker. The quantity δgc corresponds to the average log fold change that geneg is over-expressed in cell c, which only occurs for marker genes for cell types since ρgc mustequal 1 for this to contribute to the likelihood. By default we impose a lower bound such thatδ > log 2, making the interpretation that a marker gene must be over-expressed by a factor of 2relative to cells for which it is not a marker, but this is left as an option for the user. We alsocontrol for technical or sample effects through the matrix X.The user can specify whether or not to put a lognormal shrinkage prior over δgc values, wherethe mean and variance parameters of the lognormal are initialized to 0 and 1, respectively. Inplot labels, cellassign shrinkage refers to the version of CellAssign with this option turnedon.4.2.1.2 InferenceThe likelihood is given byyng|zn = c ∼ NB(µngc, φ˜ngc)where NB is the negative binomial distribution parametrized by a mean µ and a µ-specificdispersion φ˜ngc. We define φ˜ngc as a sum of radial basis functions dependent on the modelledmean µngc as proposed by a recent publication [188]:φ˜ngc =B∑i=1ai × exp(−bi × (µngc − xi)2)where ai and bi represent RBF parameters to be fitted, B is the total number of centers in theRBF, and xi is center i. The centers are set to be equally spaced apart from 0 to the maximum113number of counts max yng.Using EM for inference, the latent variables are z ≡ {zn} while the model parameters to bemaximized are δ = {δgc}, β = {βg0, βgp}, a = {ai}, and b = {bi}.E-step Computeγnc := p(zn = c|yn, δ(t−1),β(t−1),a(t−1),b(t−1)) =∏gNB(µngc, φ˜ngc)∑c′∏g′ NB(µng′c′ , φ˜ng′c′),where θ(t) is the value of some parameter θ at iteration t. We then form the Q functionQ(δ(t),β(t),a(t),b(t)|δ(t−1),β(t−1),a(t−1),b(t−1))= Ez|Y,δ(t−1),β(t−1),φ(t−1)[log p(Y|pi, δ(t),β(t),φ(t))]=N∑n=1C∑c=1γncG∑g=1logNB(yng|µngc, φ˜ngc)M-step During the M-step we optimize the above Q-function using the ADAM optimizer[189] as implemented in Google’s Tensorflow [190]. By default we use a learning rate of 0.1,allow a maximum of 105 ADAM iterations per M-step, and consider the M-step converged whenthe change in the Q function value falls below 10−4%. By default we consider the EM algorithmconverged when the change in the marginal log likelihood falls below 10−4%.Initialization The following initializations are used for model parameters:• βgp is drawn from a N (0, 1) distribution• log δgc is drawn from a N (0, 1) distribution truncated at [log(δmin), 2]• a is initialized to 0• b is initialized to twice the square difference between successive spline basesTo deal with convergence to local optima, multiple random initializations of log δgc and βgp canbe used for each run (5 by default). The number of spline bases is set to 20 by default, but themodel appears to be fairly insensitive to this setting in the tested range of 5 to 40.1144.2.2 Simulation4.2.2.1 Model description and rationaleInitially, we attempted to simulate multi-group data from the splatter model. We employed10x Chromium data for peripheral blood mononuclear cells (PBMC) [110] with cell type labelsderived from [191] to determine realistic parameter estimates for the differential expressioncomponent of the model (see below). In order to do so, group-specific log fold-change (logFC)values were drawn from a mixture distribution of a central, narrow Gaussian-Laplace mixture(representing non-differentially expressed genes) and two flanking, absolute value-transformedGaussians (representing downregulated/upregulated genes). This mixture distribution was fittedto logFC values derived from differential expression analysis (see below).However, inspection of posterior predictive samples for multiple fits, using labeled single cellRNA-seq data from [110] and FACS-purified data from Koh et al. [187] (Figure 4.1A,B,Figure 4.2A,B), revealed that this model systematically underestimates extreme logFC values(Figure 4.1C, Figure 4.2C). Thus, to accommodate the heavier tails present in observed data,we augmented the splatter model by replacing the flanking absolute value-transformed Gaussiancomponents with bounded Student’s t distributions. Posterior predictive logFC distributions fromthis modified model better fit the observed data (Figure 4.1D, Figure 4.2D). Consequently,we used this model to perform simulation analysis.1150.020 0.015 0.010 0.005 0.000 0.005 0.010 0.015 0.020logFC0200400600800Frequencya0.020 0.015 0.010 0.005 0.000 0.005 0.010 0.015 0.020logFC0500100015002000250030003500Frequencyb0.4 0.2 0.0 0.2 0.4 0.6 0.8 1.0Quantiles of observed logFC0.40.20.00.20.40.60.81.0Quantiles of posterior predictive samplesc0.4 0.2 0.0 0.2 0.4 0.6 0.8 1.0Quantiles of observed logFC0.40.20.00.20.40.60.81.0Quantiles of posterior predictive samplesdFigure 4.1: Fitting single cell RNA-seq simulation models to the Zheng PBMC 68k dataset, usingcell type annotations provided in [191]. (a) Log fold change values computed from differentialexpression analysis between naive CD8+ and naive CD4+ T cells. (b) ‘Null’ log fold change valuescomputed by randomly splitting naive CD8+ T cells into equally sized halves 10 times. (c) Quantile-quantile (QQ) plot comparing observed log fold change values between naive CD8+ and naive CD4+T cells and posterior predictive samples from the splatter model (Methods). (d) Quantile-quantile(QQ) plot comparing observed log fold change values between naive CD8+ and naive CD4+ T cellsand posterior predictive samples from the modified model (Methods).1164 3 2 1 0 1 2 3 4logFC050100150200250300Frequencya4 3 2 1 0 1 2 3 4logFC02004006008001000120014001600Frequencyb6 4 2 0 2 4 6Quantiles of observed logFC6420246Quantiles of posterior predictive samplesc6 4 2 0 2 4 6Quantiles of observed logFC6420246Quantiles of posterior predictive samplesdFigure 4.2: Fitting single cell RNA-seq simulation models to the Koh et al. [187] dataset ofFACS-purified cell types. (a) Log fold change values computed from differential expression analysisbetween human embryonic stem cells (hESCs) and day 3 somite cells (ESMT). (b) ‘Null’ log foldchange values computed by randomly splitting naive anterior primitive streak cells into equally sizedhalves 10 times. (c) Quantile-quantile (QQ) plot comparing observed log fold change values betweenhESC and ESMT cells and posterior predictive samples from the splatter model (Methods). (d)Quantile-quantile (QQ) plot comparing observed log fold change values between hESC and ESMTcells and posterior predictive samples from the modified model (Methods).4.2.2.2 Model fittingThe models described above were fit to logFC values derived from real data. Using thelabeled 10x Chromium data for 68k PBMCs [110], differential expression was performed withthe findMarkers function from the R package scran [192]. To generate corresponding null117distributions of logFC values for non-differentially expressed genes, we split data for each celltype into equally sized halves 10 times, running findMarkers to compare the resulting halves. Acentral Gaussian-Laplace mixture (µ = 0) was first fit to the null logFC values. The distributionof posterior predictive logFC values appeared to be consistent with observed logFC values forthis null component (Figure 4.1D). Following this, the entire mixture distribution was fittedto logFC values for pairs of distinct cell types, using maximum a posteriori (MAP) estimatesof parameters for the central Gaussian-Laplace component. Posterior distributions of modelparameters were inferred using the no U-turn sampler (NUTS) in pymc3, using 4 independentchains, 1000 tuning iterations, and 2500 additional iterations per chain. Trace plots and theGelman-Rubin diagnostic were used to assess convergence.4.2.2.3 Simulating multi-group dataExpression count matrices were simulated using a modified version of the splatter package.Log fold change values were simulated according to our model instead of the splatter model.Other settings were kept identical. We used MAP estimates of µ+, µ−, σ+, σ−, ν+, and ν−,determined by fitting our simulation model to (1) logFC values between naive CD4+ andnaive CD8+ T cells (Figure 4.1A); and (2) logFC values between B cells and CD8+ T cells(Section 4.2.2.1) for the differential expression component. The proportion of downregulatedgenes out of differentially expressed genes was set to 0.5 (i.e. equally probable for a differentiallyexpressed gene to be downregulated vs. upregulated). Three “groups” (cell types) were simulatedat equal proportions. Other parameters for splatter were fitted from 10x Chromium data for4,000 T cells available from 10x Genomics.To assess the performance of CellAssign relative to other clustering methods across a rangeof pd values (proportion of genes differentially expressed between each pair of cell types), pdwas chosen from {0.05, 0.15, 0.25, 0.35, 0.45, 0.55}. (The true MAP estimate of pd was 0.0746for naive CD4+ vs. naive CD8+ T cells, and 0.153 for B vs. CD8+ T cells.) The number ofsimulated cells, n, was set to 2000, and 1000 were randomly set aside for training (for scmapand correlation-based supervised clustering).To assess the robustness of CellAssign to misspecification of the marker gene matrix ρ, pd wasset to 0.25 and the number of simulated cells n to 1500.Simulations were run 9 times with unique random seeds for each combination of parametersettings.1184.2.2.4 Clustering multi-group dataCount matrices were normalized with scater normalize and the top 50 principal componentswere computed from the top 1000 most variable genes. For phenograph, Seurat (resolu-tion ∈ {0.4, 0.8, 1.2}), k-means, densitycut, and dynamicTreeCut, unsupervised clustering wasperformed on the values of these top 50 PCs. For SC3, the entire normalized SingleCellExperi-ment object was passed as input instead. For supervised methods (scmap-cluster [185] andcorrelation-based [110]), expression data for both training and evaluation sets was provided.For CellAssign, the raw count matrix was provided as input, along with a set of marker genesselected based on simulated log fold change and mean expression values. Specifically, a genewas defined as a marker gene if it was in the top 5th percentile of differentially expressed genesaccording to logFC and the top 10th percentile of differentially expressed genes according tomean expression. In simulations of robustness to marker gene misspecification, a proportion ofrandomly selected entries in the marker gene matrix ρ were flipped from 0 to 1 (or vice versa).4.2.2.5 Mapping clusters to true groupsFor assignments derived from unsupervised clustering, clusters were mapped to simulated groupsby first performing differential expression between each cluster and the remaining cells. Followingthis, we computed the Spearman correlation between these logFC values and the simulated(true) logFC values for each simulated group. Each inferred cluster was mapped to most highlycorrelated simulated group based on Spearman’s ρ where ρ > 0 and P ≤ 0.05. Clusters thatcould not be mapped based on these criteria were marked as ‘unassigned’.4.2.2.6 BenchmarkingWe generated synthetic datasets for benchmarking from the modified splatter model (Sec-tion 4.2.2.1) with Student’s t parameters µ = 0.1, σ = 0.1, ν = 1 and the proportion ofdifferentially expressed genes per cell type set to 20%. Synthetic datasets of various sizes(number of cells N ∈ {1000, 2000, 4000, 8000, 10000, 20000, 40000, 80000} and number of celltypes C ∈ {2, 4, 6, 8} with a balanced number of cells per type were generated. Markers forCellAssign were selected from genes in the top 20th percentile in terms of log fold change amongdifferentially upregulated genes and the top 10th percentile in terms of expression. CellAssignwas run with 2, 4, 6, and 8 markers per cell type, with a maximum minibatch size of 5000 cells.Five separate CellAssign runs were timed for each combination of parameters.1194.2.3 Koh et al. datasetThis section refers to the scRNA-seq dataset from [187].4.2.3.1 Preprocessing and normalization of single cell RNA-seq dataPreprocessed data was obtained from the R package DuoClustering2018 [181, 187]. Celltypeswith both single cell RNA-seq data and bulk RNA-seq data were used: hESC (day 0 humanembryonic stem cell), APS (day 1 anterior primitive streak), MPS (day 1 mid primitive streak),DLL1pPXM (day 2 DLL1+ paraxial mesoderm), ESMT (day 3 somite), Sclrtm (day 6 sclerotome),D5CntrlDrmmtm (day 5 dermomyotome), D2LtM (day 2 lateral mesoderm). Normalizationand dimensionality reduction was performed with scater normalize, runPCA, runTSNE, andrunUMAP. The top 500 most variable genes were used to compute the top 50 principal components,and the top 50 PCs were used as input for t-SNE and UMAP.4.2.3.2 Identification of marker genes from bulk RNA-seq dataDifferential expression analysis results for bulk RNA-seq data for the same cell types was usedto compute the relative expression of each gene in each cell type. Briefly, bulk RNA-seq logfold change values obtained from [187] were used to compute log-scale relative gene expressionlevels. Next, we identified gene-specific thresholds for defining the cell types in which each geneis a marker. For each gene, relative expression levels across cell types were sorted in ascendingorder, denoted as E1, ..., EC , where C is the total number of cell types. The maximum differencebetween sorted expression levels, max1≤i<C(Ei+1 − Ei), was then computed. Denote the indexi for gene g in which this difference is maximal ig. For gene g, cell types in which relativeexpression values were equal to or greater than Eig+1 were considered cell types with gene g asa marker. Genes with a maximum difference value in the the top 20th percentile were used asmarker genes.4.2.3.3 CellAssignCellAssign was run on count data using the marker gene matrix defined from bulk RNA-seqdata described above. Three random initializations of expectation-maximization were used withshrinkage priors on δgc turned on (Section 4.2.1.1). Results from the run that reached thehighest marginal log-likelihood at convergence were kept.1204.2.3.4 Unsupervised clusteringUnsupervised clustering was performed on the top 50 PCs with phenograph [193] and Seurat [178](resolution ∈ {0.4, 0.8, 1.2} and on the SingleCellExperiment object of raw and normalizedcounts with SC3 [177]. Inferred clusters were mapped to true (FACS-purified) cell types bycomputing the pairwise Spearman correlation between mean expression vectors for each clusterand each true cell type. Each cluster was treated as the cell type it was most strongly positivelyassociated with by Spearman’s ρ.4.2.4 High-grade serous ovarian cancer4.2.4.1 Sample preparation, library preparation, and sequencingSample preparation, library preparation, and sequencing steps are described in Chapter 2(see Section 2.2.8). Cell dissociation was carried out at 6◦C to maximize lymphocyte yield(O’Flanagan et al., unpublished). The 10x Chromium 5’ gene expression kit was used for singlecell RNA-seq library preparation.4.2.4.2 Preprocessing and normalization of single cell RNA-seq dataRaw sequence files were processed with CellRanger v2.1.0. The resulting filtered count matriceswere read into SingleCellExperiment objects. Outlier cells according to quality controlparameters (≥ 3 median absolute deviations from the median) were filtered out using the scaterR package. Additionally, cells with ≥ 20% mitochondrial UMIs or ≥ 50% ribosomal UMIs wereremoved. Size factors were computed using quickCluster and computeSumFactors from thescran R package. Following this, data normalization was performed using scater normalize.Principal components analysis was performed on the resultant normalized logcounts for the top1000 most variable genes. The first 50 PCs were used as input for t-SNE and UMAP.4.2.4.3 CellAssignThe following marker gene list was used for CellAssign [194–197]:• B cells: VIM c, MS4A1 c, CD79Ac, PTPRC c, CD19 c, BANK1 [194]• CD4 T cells: VIM c, CD2 c, CD3Dc, CD3E c, CD3Gc, CD28 c, PTPRC c, CD4 c• Cytotoxic T cells: VIM c, CD2 c, CD3Dc, CD3E c, CD3Gc, CD28 c, PTPRC, CD8Ac, CD8Bc,PRF1 c, GNLY c, NKG7 c, KLRC1 c• Monocyte/Macrophage: VIM c, CD14 c, FCGR3Ac, CD33 c, ITGAX c, ITGAM c, CD4 c,PTPRC c, LYZ c121• Epithelial/cancer cell: EPCAM c, MUC1 c, CDH1 c, MYC c• Stromal cell: VIM c, ECEL1 [198], KLHDC8A[198], MUM1L1 [198], ARX [198], ACTA2 c• Endothelial cells: VIM c, EMCN c, CLEC14A [199], CDH5 c, PECAM1 cc: canonical markerCellAssign was run with default parameters, the shrinkage prior on δgc values turned on, and 5random initializations. Patient was added as an additional covariate into the design matrix X(Section 4.2.1.1). The best result according to marginal log-likelihood at convergence was kept.Optimization was considered converged after 3 consecutive rounds of no improvement (relativechange in log-likelihood < 10−5). MAP assignments from CellAssign were used for downstreamanalysis.4.2.4.4 Unsupervised clusteringCells with a total probability of at least 0.99 for the stromal cell type were subsetted. The top50 PCs from preprocessing were provided as input to densitycut, which was run with defaultparameters.4.2.5 Follicular lymphoma4.2.5.1 Sample preparationLeftovers from clinical flowed samples were collected and frozen in fetal calf serum containing10% DMSO. Cells were thawed and washed according to the steps outlined in the 10X GenomicsSample Preparation Protocol. Cells were stained with PI for viability and sorted in the BDFACSAria Fusion using a 85um nozzle. Sorted cells were collected in 0.5 ml of medium,centrifuged and diluted in 1X PBS with 0.04% bovine serum albumin.4.2.5.2 Library preparation and sequencingCell concentration was determined by using a Countess II Automated Cell Counter and approx-imately 3,500 cells were loaded per well in the Single Cell 3’ Chip. Single cell libraries wereprepared according to the Chromium Single Cell 3’Reagent Kits V2 User Guide. Single celllibraries from two samples were pooled and sequenced on one HiSeq 2500 125 base PET lane.4.2.5.3 Preprocessing and normalization of single cell RNA-seq dataRaw sequence files were processed with CellRanger v2.1.0. The resulting filtered count matriceswere read into SingleCellExperiment objects. Outlier cells according to quality control122parameters (≥ 3 median absolute deviations from the median) were filtered out using the scaterR package. Additionally, cells with ≥ 10% mitochondrial UMIs or ≥ 60% ribosomal UMIs wereremoved. Size factors were computed using quickCluster and computeSumFactors from thescran R package. Following this, data normalization was performed using scater normalize.Principal components analysis was performed on the resultant normalized logcounts for the top1000 most variable genes. The first 50 PCs were used as input for t-SNE and UMAP.Cell cycle scores were computed with cyclone from the scran package [192, 200].4.2.5.4 scvis analysisscvis train (v0.1.0) [201] was run with default settings on the top 50 PCs to produce a 2-dimensional embedding of the follicular lymphoma data. Early stopping was added to scvis,so that the model would terminate after 3 successive iterations of no improvement (relativeimprovement in ELBO < 10−5). The resultant model was saved and used for mapping inSection 4.2.6.4.4.2.5.5 CellAssignThe following marker gene list was used for CellAssign:• B cells: CD19 c, MS4A1 c, CD79Ac, CD79Bc, CD74 c, CXCR5 [202]• Cytotoxic T cells: CD2 c, CD3Dc, CD3E c, CD3Gc, TRAC c, CD8Ac, CD8Bc, GZMAc,NKG7 c, CCL5 c, EOMES c• Follicular helper T cells: CD2 c, CD3Dc, CD3E c, CD3Gc, TRAC c, CD4 c, CXCR5 c,PDCD1 c, TNFRSF4 [194], ST8SIA1 [194], ICA1 [194], ICOS [194]• Other CD4+ T cells: CD2 c, CD3Dc, CD3E c, CD3Gc, TRAC c, CD4 c, IL7R [194]c: canonical markerCellAssign was run with default parameters, the shrinkage prior on δgc values turned on, and 5random initializations. Patient was added as an additional covariate into the design matrix X(Section 4.2.1.1). The best result according to marginal log-likelihood at convergence was kept.Optimization was considered converged after 3 consecutive rounds of no improvement (relativechange in log-likelihood < 10−5). MAP assignments from CellAssign were used for downstreamanalysis.4.2.5.6 Classifying B cellsB cells from CellAssign were further subclassified into ‘malignant’ or ‘nonmalignant’ groupsaccording to expression of the constant region of the immunoglobulin light chain (kappa or123lambda type) and the results of PCA. Seurat [178] (resolution = 0.8) was used to separate Bcells into clusters, based on the top 50 PCs. Following this, the sole cluster associated withIGKC (immunoglobulin light chain kappa-type constant region) expression was designated asnonmalignant. We further reasoned this was the case based on the cluster containing a mixtureof T1 and T2 cells and constituting only a minor subset of the B cells.4.2.5.7 Differential expression between timepointsDifferential expression analysis between timepoints for a given celltype and patient was performedusing voom from the limma package for each patient and cell type separately, with timepointas the independent variable. Genes with low expression (< 500 UMIs in total across all cells)were removed. P -values were adjusted with the Benjamini-Hochberg method, and genes withQ ≤ 0.05 were considered differentially expressed. Differential expression between malignantand nonmalignant B cells was performed similarly, but using the formula ~malignant status +timepoint + malignant status:timepoint to control for timepoint and any interactions.4.2.5.8 Reactome pathway enrichment analysisPathway analysis was performed for the top 50 most upregulated and top 50 most downregulatedgenes (separately) by log fold change from limma (where Q ≤ 0.05, filtering out ribosomal andmitochondrial genes). Overrepresentation of Reactome [203] pathways was assessed using the Rpackage ReactomePA. Pathways were considered significantly overrepresented if the adjustedP -value ≤ 0.05 and at least 2 genes from the pathway were present.4.2.5.9 Comparing malignant cells between timepointsLog fold change values from the findMarkers function (filtering out ribosomal and mitochondrialgenes) from scran were used as input for gene set enrichment analysis with the fGSEA R package,using default parameters with 10,000 permutations, and the hallmark pathway gene set [204].Annotations for cell cycle-associated pathways (E2F targets, G2M checkpoint, and mitoticspindle) were taken from [204]. BH-adjusted P -values for differences in proliferation markerexpression (MKI67 and TOP2A) were also computed with the findMarkers function fromscran, using default parameters.4.2.5.10 Somatic variant callingSomatic single-nucleotide variants (SNVs), indels, and breakpoints for both cases were obtainedfrom [205]. Annotations from the Nanostring PanCancer Immune Profiling panel were used to124identify antigen processing and presentation genes [149].4.2.5.11 HLA loss-of-heterozygosity analysisHLA class I typing was performed using matched normal bams [205] with OptiType [163].Following this, HLA class I loss-of-heterozygosity (LOH) was called from tumor and matchednormal bams as well as OptiType 4-digit HLA types using LOHHLA [52]. HLA LOH was calledfor an allele if the estimated copy number (with binning and B-allele frequency settings) was< 0.5 and the significance of allelic imbalance p < 0.05 (paired t test, no duplicate counts).4.2.6 Reactive lymph node data4.2.6.1 Sample preparationCell suspensions from patients with reactive lymphoid hyperplasia but no evidence of malignantdisease and collagen disease were used. Leftovers from clinical flowed samples were collected andfrozen in FCS containing 10%DMSO. The day of the experiment cell suspensions were rapidlythawed at 37◦C, and washed according to the steps outlined in the 10X Genomics SamplePreparation Protocol. Cells were stained with DAPI and viable cells (DAPI negative) weresorted on a FACS ARIAIII or FACS Fusion (BD Biosciences) instrument.4.2.6.2 Library preparation and sequencingApproximately 8,700 cells per sample were loaded into a Chromium Single Cell 3’ Chip kit v2(PN-120236) and processed according to the Chromium Single Cell 3’Reagent kit v2 User Guide.Libraries were constructed using the Single 3’ Library and Gel Bead Kit v2 (PN-120237) andChromium i7 Mulitiplex Kit v2 (PN-120236). Single cell libraries from two samples were pooledand sequenced on one HiSeq 2500 125 base PET lane.4.2.6.3 Preprocessing and normalization of single cell RNA-seq dataPreprocessing steps for the reactive lymph node data were identical to those for single cellRNA-seq data, described in Section 4.2.5.3.4.2.6.4 scvis analysisThe identities of the top 1000 most variable genes and PCA loadings from follicular lymphomadata analysis were used to compute a 50-dimensional embedding for the reactive lymph node125data. Following this, the resultant 50 PCs were provided as input to scvis map [201], using themodel trained in Section 4.2.5.4 and default settings.4.3 Results4.3.1 Automated assignment of cell types with CellAssignThe CellAssign statistical framework (Figure 4.3) models observed gene expression as acomposite of cell type-specific, library size, and batch effects, using raw single cell RNA-seqcounts from a heterogeneous cellular population as input. To enable automated cell typeclassification, marker gene information is provided a priori to CellAssign in the form of a set ofmarker genes for each modeled cell type. The sole assumption for a marker gene to be indicativeof a cell type is that it is should be over-expressed in that cell type relative to all others - it maystill be expressed in all cells and variable between others. Information on other experimental andbiological covariates - such as batch and patient-of-origin - can also be encoded as a standarddesign matrix. Using this information, CellAssign employs a hierarchical Bayesian statisticalframework to determine the probability that each cell belongs to each of the modeled cell types,along with estimates of the model parameters including the relative expression of marker genes ineach cell type and the systematic effects of other covariates on marker gene expression patterns.To prevent misclassification when unknown cell types are present, CellAssign can assign cellsthat do not belong to any of the provided cell types to an ‘unassigned’ group.126CellAssign: automated probabilistic assignment of cells to known cell typesPrior knowledge of marker genesT cell: CD2, CD3, CD45B cell: CD19, CD20, CD79A, CD45Tumor cell: EPCAMFibroblast: COL1A1, ACTA2, THY90Dim-1Dim-2Cell type assignmentsCellsProbabilitiesExpression count matrix forheterogeneous populationa bcGenesFigure 4.3: (a) Overview of CellAssign. CellAssign takes raw count data from a heterogeneous single-cell RNA-seq population,along with a set of known markers for various cell types understudy. Using CellAssign for inference, each cell is probabilisticallyassigned to a given cell type without any need for manual annotation or intervention, accounting for any batch or sample-specificeffects. (b) An overview of the CellAssign probabilistic graphical model. (c) The random variables and data that form the model,along with the distributional assumptions and description.1274.3.2 Performance of CellAssign relative to state-of-the-art unsupervisedand supervised classification methodsWhile CellAssign is the only method to-date that can automatically assign cells to cell types basedon prior marker gene associations, we sought to demonstrate its performance was competitivecompared to standard workflows including unsupervised clustering followed by manual curationand methods that map cells to existing data from purified populations. We employed an adaptedversion of the splatter model to simulate single cell RNA-seq data for multiple cell populations(Methods). On simulated data for 80000 cells from 2 cell types, CellAssign completed in under2 minutes, and appeared to scale at worst linearly in the number of cell types and marker genesused per cell type (Figure 4.4). In order to select realistic parameter settings for simulation,we fitted the splatter model to data for na¨ıve CD8+ and CD4+ T cells from peripheral bloodmononuclear cell (PBMC) data. Simulations were conducted across a wide range of values for thefraction of differentially expressed genes (0.05 to 0.55), to represent cellular mixtures of similarand distinct cell types. Following this, we evaluated the performance of each unsupervised(Seurat [178], SC3 [177], phenograph [193], densitycut [206], dynamicTreeCut [207], k-means)and supervised (scmap-cluster [185], correlation-based [110]) clustering methods for single cellRNA-seq data (Methods). Half of the simulated cells (n=2000 total, n=1000 training, n=1000evaluation) were set aside for training the supervised methods. Marker genes for CellAssignwere selected based on simulated log-fold change values and mean expression (Methods),and maximum a posteriori (MAP) cell type probability estimates were treated as cell typeassignments. For all values of the fraction of differentially expressed genes, CellAssign performedcomparably or superior to alternative workflows in terms of accuracy and F1 score (Figure 4.5A,Supplemental Table B.1). As expected, supervised methods generally performed better thanunsupervised methods (Figure 4.5A).128501001000 2000 4000 8000 10000 20000 40000 80000Number of cellsTime to convergence (seconds)aMarkers per celltype2 4 6 82550751002 4 6 8Number of cell typesTime to convergence (seconds)bFigure 4.4: Benchmarking results for CellAssign across a range of simulated data set sizes (numberof cells), number of cell types being inferred, and number of marker genes per cell type. (a) Runtime(to convergence, defined as a relative change in log-likelihood < 10−3 between successive iterations,as a function of data set size and the number of marker genes used per cell type, on simulated data(Methods). Two cell types were used. (b) Runtime (to convergence, defined as a relative change inlog-likelihood < 10−3 between successive iterations, as a function of the number of cell types andthe number of marker genes used per cell type, on simulated data. One thousand cells were used.1290.05 0.15 0.25 0.35 0.45 0.55F1Accuracycellassign_shrinkagecellassignZheng_corscmapSC3densitycutphenographseurat_0.8seurat_1.2seurat_0.4dynamicTreeCutkmeanscellassign_shrinkagecellassignZheng_corscmapSC3densitycutphenographseurat_0.8seurat_1.2seurat_0.4dynamicTreeCutkmeanscellassign_shrinkagecellassignZheng_corscmapSC3densitycutphenographseurat_0.8seurat_1.2seurat_0.4dynamicTreeCutkmeanscellassign_shrinkagecellassignZheng_corscmapSC3densitycutphenographseurat_0.8seurat_1.2seurat_0.4dynamicTreeCutkmeanscellassign_shrinkagecellassignZheng_corscmapSC3densitycutphenographseurat_0.8seurat_1.2seurat_0.4dynamicTreeCutkmeanscellassign_shrinkagecellassignZheng_corscmapSC3densitycutphenographseurat_0.8seurat_1.2seurat_0.4dynamicTreeCutkmeans0.000.250.500.751.000.000.250.500.751.00MethodScore% of genes differentially expressed per cell typea0.05 0.15 0.25 0.35 0.45 0.55F1Accuracycellassign_shrinkagecellassignZheng_corscmapseurat_1.2seurat_0.8seurat_0.4SC3kmeansphenographdensitycutdynamicTreeCutcellassign_shrinkagecellassignZheng_corscmapseurat_1.2seurat_0.8seurat_0.4SC3kmeansphenographdensitycutdynamicTreeCutcellassign_shrinkagecellassignZheng_corscmapseurat_1.2seurat_0.8seurat_0.4SC3kmeansphenographdensitycutdynamicTreeCutcellassign_shrinkagecellassignZheng_corscmapseurat_1.2seurat_0.8seurat_0.4SC3kmeansphenographdensitycutdynamicTreeCutcellassign_shrinkagecellassignZheng_corscmapseurat_1.2seurat_0.8seurat_0.4SC3kmeansphenographdensitycutdynamicTreeCutcellassign_shrinkagecellassignZheng_corscmapseurat_1.2seurat_0.8seurat_0.4SC3kmeansphenographdensitycutdynamicTreeCut0.000.250.500.751.000.000.250.500.751.00MethodScore% of genes differentially expressed per cell typebR = 0.958, Rs = 0.962 R = 0.976, Rs = 0.982DE prob = 0.15 DE prob = 0.450 2 4 6 0 2 4 602468True logFCInferred logFCcF1Accuracy0 0.1 0.2 0.3 0.4 0.5 0.60 0.1 0.2 0.3 0.4 0.5 0.60.000.250.500.751.000.000.250.500.751.00Proportion of incorrect entries in rhoScoredMethodcellassign cellassign_shrinkageFigure 4.5: Performance of CellAssign on simulated data. (a) Accuracy and cell-level F1 score(Methods) for varying proportions of differentially expressed genes per cell type, with otherdifferential expression parameters set to MAP estimates determined from comparing naive CD8+and naive CD4+ T cells (Methods). CellAssign was provided with a set of marker genes (Methods);all other methods were provided with all genes. cellassign shrinkage refers to a version of CellAssignwith a shrinkage prior on δ (Methods). (b) Accuracy and cell-level F1 score for varying proportionsof differentially expressed genes per cell type, with other differential expression parameters set toMAP estimates determined from comparing naive CD8+ and naive CD4+ T cells. All methodswere provided with the same set of marker genes. (c) Correspondence between true simulated logfold change values and log fold change (δ) values inferred by CellAssign. R and Rs refer to thePearson correlation between true and inferred logFC values for cellassign and cellassign shrinkage,respectively. (d) Performance of CellAssign where a certain proportion of entries in the marker genematrix are flipped at random. Differential expression parameters used for these simulations werebased on those determined from comparing B and CD8+ T cells.130In case CellAssign’s superior performance was due to being provided solely with informativemarker genes compared to transcriptome-wide data provided to other methods, we repeated oursimulations providing other methods with exactly the same data as CellAssign. Nonetheless,CellAssign performed superiorly to the other tested methods (Figure 4.5B). Similar resultswere obtained on data simulated from parameter estimates fitted to B cells and CD8+ Tcells (Figure 4.6A,B, Supplemental Table B.1). Moreover, CellAssign accurately infers therelative expression of marker genes in each cell type (all R >0.958; Figure 4.5C, Figure 4.6C).1310.05 0.15 0.25 0.35 0.45 0.55F1Accuracycellassigncellassign_shrinkageSC3Zheng_corseurat_0.4seurat_0.8seurat_1.2densitycutphenographdynamicTreeCutscmapkmeanscellassigncellassign_shrinkageSC3Zheng_corseurat_0.4seurat_0.8seurat_1.2densitycutphenographdynamicTreeCutscmapkmeanscellassigncellassign_shrinkageSC3Zheng_corseurat_0.4seurat_0.8seurat_1.2densitycutphenographdynamicTreeCutscmapkmeanscellassigncellassign_shrinkageSC3Zheng_corseurat_0.4seurat_0.8seurat_1.2densitycutphenographdynamicTreeCutscmapkmeanscellassigncellassign_shrinkageSC3Zheng_corseurat_0.4seurat_0.8seurat_1.2densitycutphenographdynamicTreeCutscmapkmeanscellassigncellassign_shrinkageSC3Zheng_corseurat_0.4seurat_0.8seurat_1.2densitycutphenographdynamicTreeCutscmapkmeans0.000.250.500.751.000.000.250.500.751.00MethodScore% of genes differentially expressed per cell typea0.05 0.15 0.25 0.35 0.45 0.55F1Accuracycellassigncellassign_shrinkagescmapZheng_corseurat_0.4seurat_0.8dynamicTreeCutkmeansseurat_1.2densitycutphenographSC3cellassigncellassign_shrinkagescmapZheng_corseurat_0.4seurat_0.8dynamicTreeCutkmeansseurat_1.2densitycutphenographSC3cellassigncellassign_shrinkagescmapZheng_corseurat_0.4seurat_0.8dynamicTreeCutkmeansseurat_1.2densitycutphenographSC3cellassigncellassign_shrinkagescmapZheng_corseurat_0.4seurat_0.8dynamicTreeCutkmeansseurat_1.2densitycutphenographSC3cellassigncellassign_shrinkagescmapZheng_corseurat_0.4seurat_0.8dynamicTreeCutkmeansseurat_1.2densitycutphenographSC3cellassigncellassign_shrinkagescmapZheng_corseurat_0.4seurat_0.8dynamicTreeCutkmeansseurat_1.2densitycutphenographSC30.000.250.500.751.000.000.250.500.751.00MethodScore% of genes differentially expressed per cell typebMethodcellassigncellassign_shrinkagedensitycutdynamicTreeCutkmeansphenographSC3scmapseurat_0.4seurat_0.8seurat_1.2Zheng_corR = 0.988, Rs = 0.988 R = 0.982, Rs = 0.981 R = 0.984, Rs = 0.979 R = 0.977, Rs = 0.97DE prob = 0.05 DE prob = 0.15 DE prob = 0.25 DE prob = 0.350 2 4 6 0 2 4 6 0 2 4 6 0 2 4 602468True logFCInferred logFCcMethodcellassign cellassign_shrinkageFigure 4.6: Simulation performance across a range of proportions of differentially expressed genes,using differential expression parameters derived from comparing B and CD8+ T cells. (a) Accuracyand cell-level F1 score (Methods) for varying proportions of differentially expressed genes per celltype. CellAssign was provided with a set of marker genes (Methods); all other methods wereprovided with all genes. (b) Accuracy and cell-level F1 score for varying proportions of differentiallyexpressed genes per cell type. All methods were provided with the same set of marker genes. (c)Correspondence between true simulated log fold change values and log fold change (δ) values inferredby CellAssign. R and Rs refer to the Pearson correlation between true and inferred logFC values forcellassign and cellassign shrinkage, respectively.132We next assessed the robustness of CellAssign to mis-specification of the association of markergenes to cell types. While CellAssign assumes information for user-provided marker genesis complete and correct, in reality this may not always be the case. For example, a sharedmarker gene may be incorrectly specified as a cell type-specific marker gene due to incompleteprior information or human error. Thus, we tested the robustness of CellAssign to markergene mis-specification by changing a proportion of entries in the binary marker gene matrixfrom 0 to 1 or vice-versa at random. Supplied with data for 5 marker genes per cell type,CellAssign maintained comparable performance in scenarios where up to 30% of matrix entrieswere mis-specified (Figure 4.5D, Supplemental Table B.1). This robustness to markermis-specification is maintained even when the simulated cells belong to transcriptionally similarcell types containing few highly differentially expressed genes. For example, when cells weresimulated based on the degree of dissimilarity between na¨ıve CD4+ and na¨ıve CD8+ T cells,the accuracy of CellAssign predictions was comparably high in scenarios where 20% of markergene matrix entries were mis-specified (Figure 4.7, Supplemental Table B.1).F1Accuracy0 0.1 0.2 0.3 0.4 0.5 0.60 0.1 0.2 0.3 0.4 0.5 0.60.000.250.500.751.000.000.250.500.751.00Proportion of incorrect entries in rhoScoreMethod cellassign cellassign_shrinkageFigure 4.7: Simulation performance across a range of proportions of randomly flipped entries inthe binary marker gene matrix, using differential expression parameters derived from comparingnaive CD8+ and naive CD4+ T cells.133We also assessed the performance of CellAssign on real single cell RNA-seq data. We firstapplied CellAssign to data for FACS-purified H7 human embryonic stem cells in various stagesof differentiation [187]. Using bulk RNA-seq data from the same cell types, we defined aset of 84 marker genes for CellAssign based on differential expression results (SupplementalTable B.2; Methods). CellAssign performed superiorly to the most competitive unsupervisedmethods from systematic analysis (SC3, Seurat) [181] in terms of accuracy and cell type-levelF1 score (Figure 4.8A-D,F; Methods). Similar results were obtained for comparisons usingonly expression data for the marker genes (Figure 4.8E,G). Crucially, CellAssign was able todistinguish anterior primitive streak (APS) and mid primitive streak (MPS) cells, while no othermethod could reliably do so (Figure 4.8).134−50510−10 0 10TSNE−1TSNE−2Celltype APSD2LtMD5CntrlDrmmtmDLL1pPXMESMThESCMPSSclrtmCelltypeaAccuracy = 0.948F1 = 0.947−50510−10 0 10TSNE−1TSNE−2CellAssign APSD2LtMD5CntrlDrmmtmDLL1pPXMESMThESCMPSSclrtmCellAssignbAccuracy = 0.677F1 = 0.538−50510−10 0 10TSNE−1TSNE−2cluster 123456SC3 (full)cAccuracy = 0.778F1 = 0.674−50510−10 0 10TSNE−1TSNE−2cluster 012345Seurat (res = 0.8, full)dAccuracy = 0.93F1 = 0.841−50510−10 0 10TSNE−1TSNE−2cluster 0123456Seurat (res = 0.8, markers)eAccuracy = 0.778F1 = 0.674−50510−10 0 10TSNE−1TSNE−2cluster 012345Seurat (res = 1.2, full)fAccuracy = 0.93F1 = 0.841−50510−10 0 10TSNE−1TSNE−2cluster 0123456Seurat (res = 1.2, markers)gFigure 4.8: Performance (accuracy and cell type-level F1 score, Methods) of CellAssign and thebest-performing clustering methods evaluated by [181] on FACS-purified H7 human embryonic stemcells in various stages of differentiation. t-SNE plots of (a) ground-truth FACS annotations; (b)CellAssign-derived annotations; (c) SC3 clusters (using all genes); (d) Seurat clusters (resolution =0.8, using all genes); (e) Seurat clusters (resolution = 0.8, using the same marker gene set used byCellAssign); (f) Seurat clusters (resolution = 1.2, using all genes); (g) Seurat clusters (resolution =1.2, using the same marker gene set used by CellAssign).1354.3.3 Profiling the malignant and nonmalignant composition of high-gradeserous ovarian cancerTo profile the microenvironment of HGSC, we sequenced the transcriptomes of 6298 cells from 4spatially collected pre-treatment biopsies of 3 patients with HGSC (Table 4.1; sample identifierscorrespond to those from Chapter 2). Following data preprocessing, we used CellAssign toidentify 7 known cell types including epithelial cells, endothelial cells, other stromal cells, CD4T cells, cytotoxic T cells, B cells, and monocytes, and visualized the normalized data usingprincipal components analysis (PCA) and t-distributed stochastic neighbour embedding (t-SNE)(Figure 4.9A,B, Figure 4.10A,B, Methods). These cell type assignments appeared to be con-sistent with the expression patterns of well-known marker genes [104, 195–197] (Figure 4.9C).Overall, epithelial cells were the dominant cell type (52.8% of assigned cells), composing 20% to86% of assigned cells across samples (Figure 4.10A). Multiple clusters of epithelial cells wereobserved in both samples from patient 73, suggestive of clonally distinct malignant cell subpop-ulations or a mixture of malignant and nonmalignant epithelial populations (Figure 4.9A,B).While most epithelial cell clusters in t-SNE space were largely patient-specific, most nonmalignantclusters, such as endothelial cells, stromal cells, and monocytes, contained cells from multiplepatients (Figure 4.9A,B). Lymphocytes were rare, composing only 4.3% of all cells (B cells:1.4%, CD4 T cells: 2.2%, cytotoxic T cells: 0.7%). Most B cells appeared to express CD79A andSDC1 (CD138 ) but not MS4A1 (CD20 ), consistent with plasma cells [208] (Figure 4.9C).Patient Sample ID Site Raw Filtered Temperature70 70LAdnx6 VOA11267Leftadnexalmass506 280 672 72Om VOA11558SA Omentum 559 282 673 73LOv VOA11543SA Left ovary 2818 2707 673 73ROv VOA11543SB Right ovary 2415 2125 6Table 4.1: HGSC samples profiled by single cell RNA-seq. Raw and filtered correspond to raw andpreprocessed cell counts, respectively.136−2002040−20 0 20 40TSNE−1TSNE−2Sample 70_LAdnx−6 72_Om 73_LOv 73_ROva−2002040−20 0 20 40TSNE−1TSNE−2Celltype B cellsCD4 T cellsCytotoxic T cellsEndothelial cellsEpithelial cellMonocyte/MacrophageStromal cellUnassignedb−2002040−20 0 20 40TSNE−1TSNE−2CD3E−2002040−20 0 20 40TSNE−1TSNE−2GZMA−2002040−20 0 20 40TSNE−1TSNE−2EPCAM−2002040−20 0 20 40TSNE−1TSNE−2VIM−2002040−20 0 20 40TSNE−1TSNE−2PECAM1−2002040−20 0 20 40TSNE−1TSNE−2LYZ−2002040−20 0 20 40TSNE−1TSNE−2CD79A−2002040−20 0 20 40TSNE−1TSNE−2SDC1−2002040−20 0 20 40TSNE−1TSNE−2MS4A1c0 1 2 3 4ExpressionFigure 4.9: CellAssign infers the composition of the HGSC microenvironment. (a) t-SNE plot ofHGSC single cell expression data, labeled by sample. (b) t-SNE plot of HGSC single cell expressiondata, labeled by maximum probability assignments from CellAssign. (c) Expression of select markergenes CD3E (for T cells [104]), GZMA (for CD8 T cells [104]), EPCAM (for epithelial cells [195]),VIM (for mesenchymal cells), PECAM1 (for endothelial cells [197]), CD79A (for B cells [104]), LYZ(for monocytes [104]), SDC1 (for plasma cells [208]), and MS4A1 (for non-plasma B cells).1370%25%50%75%100%70_LAdnx−672_Om73_LOv73_ROvSampleProportionaB cellsCD4 T cellsCytotoxic T cellsMonocyte/MacrophageEpithelial cellStromal cellEndothelial cellsotherSampleCelltype00.20.40.60.81bCelltypeB cellsCD4 T cellsCytotoxic T cellsEndothelial cellsEpithelial cellMonocyte/MacrophageStromal cellUnassignedSample64_Ov64_Om67_ROv70_LAdnx−670_LAdnx−3772_Om73_LOv73_ROvFigure 4.10: Proportions and probabilities of cell type assignments. (a) Proportions of eachCellAssign-assigned celltype in each sample. (b) Cell-level probabilities from the CellAssign model,labeled by maximum probability celltypes and sample.4.3.4 Stromal subpopulations in the ovarian cancer microenvironmentHaving profiled the immune microenvironment in Chapter 3, we next surveyed stromal cellpopulations in the HGSC microenvironment. Considering cells assigned to the stromal cell type(i.e. stromal cells with the exception of endothelial cells) with a probability of at least 0.99 byCellAssign, we performed unsupervised clustering with densitycut [206] (Figure 4.11A,B),revealing 6 clusters. Reasoning that stromal cells from different anatomic sites may expressdifferent markers, we interrogated the expression profiles of ovarian stroma-specific markers.Based on the expression of MUM1L1 and ARX [198], we annotated clusters 1 and 2 asovarian stromal cells, consistent with the ovarian or adnexal origin of cells from these clusters(Figure 4.11A-C). In contrast, cluster 5, which corresponds to cells from an omental sample(72 Om) did not express these markers (Figure 4.11A-C). Following this, we explored genes138associated with other microenvironmental populations, including pericytes, myofibroblasts, andsmooth muscle cells. Based on the expression of PLN, MYH11, and ACTA2 [209–211], weputatively labeled cluster 4 as vascular smooth muscle (Figure 4.11D). Most smooth musclecells are VIM (vimentin) negative and desmin positive, but vascular smooth muscle contains apredominance of vimentin [212] (Figure 4.9C). While cluster 3 also contained MYH11 -expressingcells, the pattern of expression in t-SNE space was not homogeneous. MYH11 expression incluster 3 negatively correlated with the expression of pericyte markers THY1, CD248, andPDGFRB [213] (Figure 4.11D). As such, cluster 3 likely contains a mixture of pericytesand myofibroblasts. Thus, the HGSC microenvironment contains multiple phenotypically-distinct stromal subpopulations that resemble ovarian stromal cells, extraovarian stromal cells,myofibroblasts, pericytes, and vascular smooth muscle cells.139−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2Sample 70_LAdnx−6 72_Om 73_LOv 73_ROva−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2Cluster 123456b−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2MUM1L1c−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2ACTA2d−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2MYH11−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2PLN−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2ARX−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2CD248−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2PDGFRB−20020−30 −20 −10 0 10 20 30TSNE−1TSNE−2THY10 1 2 3 4ExpressionFigure 4.11: Stromal subpopulations in the HGSC microenvironment. (a) t-SNE projection ofstromal (and non-endothelial) populations in the HGSC microenvironment, labeled by sample. (b)t-SNE projection of stromal populations in the HGSC microenvironment, labeled by CellAssign-assigned cell type. (c) Expression of ovarian stromal marker genes MUM1L1 and ARX in the HGSCmicroenvironment [198]. (d) Expression of various stromal, pericyte, and muscle-associated genes inthe HGSC microenvironment.1404.3.5 Dissecting the lymphocyte composition of follicular lymphomaTo demonstrate the utility of CellAssign for microenvironment analysis in another cancer type,we sequenced the transcriptomes of 9754 cells from serially collected lymph node biopsies of 2patients with follicular lymphoma (FL1018: 4321 cells; FL2001: 5433 cells). Histopathologicaltransformation to an aggressive subtype of B cell lymphoma, diffuse large B cell lymphoma(DLBCL), occurred in one patient (FL1018), while early progression (4 years after initialtreatment with rituximab) occurred in the other (FL2001) (Figure 4.12A). Following datapreprocessing and normalization, we applied principal components analysis (PCA) and uniformmanifold approximation and projection (UMAP, [214]), revealing 5 major clusters in the reduced-dimension projections (Figure 4.12B). Three clusters appeared to be relatively pure for cellsfrom a single patient, while the other 2 clusters comprised a mixture of cells from both patients.Leveraging marker gene information derived from the literature (Supplemental Table B.2;Methods), we applied CellAssign to identify 4 major B and T cell populations across theseclusters (Figure 4.12C). One of the mixed clusters exclusively contained T cells, while the other4 clusters were largely B cell-specific (Figure 4.12C). These cell type assignments appearedto be consistent with the expression patterns of well-known marker genes, such as CD3D andCD2 for T cells, CD79A and MS4A1 (CD20 ) for B cells, CCL5, CD8A, and GZMA for CD8+T cells, CD4, CXCR5, and ICOS for T follicular helper cells, and CD4 for other CD4+ Tcells (Figure 4.12D, Figure 4.13, Methods). No evidence of regulatory T cells (FOXP3 andIL2RA expression), NK cells (NCAM1 expression), and myeloid cells (CD14 /CD16 and LYZexpression) was detected.141T1 (FL) T2 (DLBCL)T1 (FL) T2 (FL)1452 28692736 2697FL2001FL10180 3 6 9Time since first biopsy (years)a−10−50510−10 −5 0 5 10UMAP−1UMAP−2SampleFL1018T1 FL1018T2 FL2001T1 FL2001T2b−10−50510−10 −5 0 5 10UMAP−1UMAP−2Predicted celltypeB cells CD4 T cells Cytotoxic T cells Tfh Unassignedc−10−50510−10 −5 0 5 10UMAP−1UMAP−2CD79Ad−10−50510−10 −5 0 5 10UMAP−1UMAP−2CD3D−10−50510−10 −5 0 5 10UMAP−1UMAP−2CCL5−10−50510−10 −5 0 5 10UMAP−1UMAP−2ICOS0 1 2 3ExpressionFigure 4.12: CellAssign infers the composition of the follicular lymphoma microenvironment. (a)Sample collection times for FL1018 (transformed FL) and FL2001 (progressed FL). FL1018 is alivewhile FL2001 was lost to followup (indicated by the red rectangle). The number of cells collected foreach sample is indicated. (b) UMAP plot of follicular lymphoma single cell expression data, labeledby sample. (c) UMAP plot of follicular lymphoma single cell expression data, labeled by maximumprobability assignments from CellAssign. (d) Expression of select marker genes CD79A (for B cells),CD3D (for T cells), CCL5 (for CD8+ T cells), and ICOS (for T follicular helper cells).142−10−50510−10 −5 0 5 10UMAP−1UMAP−2CD2−10−50510−10 −5 0 5 10UMAP−1UMAP−2MS4A1−10−50510−10 −5 0 5 10UMAP−1UMAP−2CD8A−10−50510−10 −5 0 5 10UMAP−1UMAP−2GZMA−10−50510−10 −5 0 5 10UMAP−1UMAP−2CD4−10−50510−10 −5 0 5 10UMAP−1UMAP−2CXCR5−10−50510−10 −5 0 5 10UMAP−1UMAP−2IL7R−10−50510−10 −5 0 5 10UMAP−1UMAP−2ICA1a0 1 2 3ExpressionCD19TRACCD2CD3DCCL5GZMANKG7CXCR5PDCD1EOMESCD8ACD8BTNFRSF4ICA1CD4ICOSST8SIA1IL7RCD3ECD3GCD74CD79BCD79AMS4A1CelltypeCelltypeB cells (malignant)CD4 T cellsCytotoxic T cellsTfhB cellsUnassigned02468bFigure 4.13: (a) Expression of select marker genes CD2 (for T cells), MS4A1 (for B cells), CD8Aand GZMA (for CD8+ T cells), CD4 (for CD4+ T cells and T follicular helper cells) and CXCR5 andICOS (for T follicular helper cells). (b) Heatmap of marker gene expression, labeled by maximumprobability CellAssign-inferred cell types.We next interrogated the identity of each B cell cluster. Of the B cell clusters, three were almostexclusively comprised of cells from a single patient, while one was a mixture of cells from bothpatients (Figure 4.14A). Reasoning that nonmalignant B cells were likely more phenotypicallysimilar across timepoints than cancer cells, we hypothesized that the mixed cluster containednonmalignant B cells. To explore this further, we examined immunoglobulin light chain constantdomain expression across these clusters (Figure 4.14B). Each clonally identical population of B143cells produces immunoglobulins containing a single class of immunoglobulin light chain (κ/IGKCor λ/IGLC ), created through V(D)J recombination. Whereas normal lymphoid organs typicallycontain a 60:40 ratio of κ- to λ-expressing B cells [215], we observed a substantial departurefrom this ratio among all B cells in both patients (Figure 4.14B). Notably, the majority of thethree patient-specific clusters were exclusively IGLC+, consistent with expansion of a malignantIGLC+ cell, while the mixed cluster contained both IGKC+ and IGLC+ cells (Figure 4.14B).Hypothesizing that the IGKC :IGLC ratio in nonmalignant B cells should be similar to thatfor normal lymphoid organs, we applied CellAssign to the mixed cluster, using IGKC as amarker of IGKC+ cells and IGLC2 and IGLC3 as markers of IGLC+ cells (IGLC1 and IGLC7were not expressed; Supplemental Table B.2). Out of the 774 cells that were assigned toeither group, 456 cells (58.9%) were classified as IGKC+ (FL1018: 67/106 (63.2%), FL2001:389/668 (58.2%)), consistent with results for normal lymphoid organs (Figure 4.15). Finally,we attempted to delineate nonmalignant B cells by comparing B cell expression patterns tothose derived from lymph node B cells from healthy donors. Using scvis [201], we first traineda variational autoencoder to produce a 2-dimensional embedding of the follicular lymphomasingle cell RNA-seq data. Following this, we applied scvis to map similarly processed single cellRNA-seq data for reactive lymph node (RLN) B and T cells from four healthy donors onto thisembedding. Concordant with our other predictions, RLN-derived T cells mapped to follicularlymphoma-derived T cells and RLN-derived B cells mapped to the mixed cluster of follicularlymphoma-derived B cells (Figure 4.14C, Figure 4.16). Thus, the mixed cluster is comprisedof nonmalignant B cells, while the other 2 clusters represent malignant B cells. Corroboratingthis, differential expression analysis revealed that these malignant B cells express significantlyhigher levels of follicular lymphoma-associated markers, such as BCL2 and BCL6 [216–218],than nonmalignant B cells (all Q < 1.8e-07; Figure 4.17, Supplemental Table B.3).144−50510−5 0 5UMAP−1UMAP−2Sample FL1018T1FL1018T2FL2001T1FL2001T2a−50510−5 0 5UMAP−1UMAP−2Celltype B cellsB cells (malignant)Unassigned−50510−5 0 5UMAP−1UMAP−2IGKC−50510−5 0 5UMAP−1UMAP−2IGLC2−50510−5 0 5UMAP−1UMAP−2IGLC30 1 2 3 4 5 6Expressionb−1001020−15 −10 −5 0 5 10 15scvis−1scvis−2Celltype B cell B cell (malignant) RLN T cell UnassignedcFL1018 FL2001T1 T2 T1 T20.000.250.500.751.00TimepointProportionCelltype nonmalignant malignantd−4−2024−4 −2 0 2 4UMAP−1UMAP−2e−4−2024−4 −2 0 2 4UMAP−1UMAP−2FL1018 FL2001T1 T2 T1 T20.20.40.6TimepointProportionfSampleFL1018T1FL1018T2FL2001T1FL2001T2CelltypeTfh CD4 T cells Cytotoxic T cellsP = 4.45e−08P = 0.847P = 0.000803P = 0.00549P = 1.41e−09P = 0.017P = 5.61e−15P = 0.506GZMA PRF1 CD69 IFNGFL1018FL2001T1 T2 T1 T2 T1 T2 T1 T20123401234TimepointExpressiongFigure 4.14: Temporal changes in nonmalignant cells in the follicular lymphoma microenvironment.(a) Left: UMAP plot of CellAssign-inferred B cells, labeled by sample. Right: UMAP pot ofCellAssign-inferred B cells, labeled by putative malignant/nonmalignant status. (b) Expression of κ(IGKC ) and λ (IGLC2 and IGLC3 ) light chain constant region genes. (c) Scvis plot of follicularlymphoma data and single cell RNA-seq data of lymphocytes from reactive lymph nodes from healthypatients. The follicular lymphoma data was used to train the variational encoder and produce thetwo-dimensional embedding. Indicated cell types are B cell (nonmalignant B cell from FL), B cell(malignant) (malignant B cell from FL), T cell (T cell from FL), RLN (reactive lymph node cell).(d) Relative proportion of B cell subpopulations over time. (e) UMAP plots of FL T cells, labeledby sample and CellAssign-inferred celltype. (f) Relative proportion of T cell subpopulations overtime. (g) Normalized expression of CD8+ T cell activation markers over time. P -values computedwith the Wilcoxon rank-sum test and adjusted with the Benjamini-Hochberg method.145IGKCIGLC2IGLC3IGLC1IGLC7ClassClassIGKCIGLCUnassigned01234567Figure 4.15: Expression of κ and λ light chain constant region genes in nonmalignant B cells.Class assignments were determined by CellAssign (Methods).146−10010−10 −5 0 5 10scvis−1scvis−2CD2−10010−10 −5 0 5 10scvis−1scvis−2CD3D−10010−10 −5 0 5 10scvis−1scvis−2CD3E−10010−10 −5 0 5 10scvis−1scvis−2CD79A−10010−10 −5 0 5 10scvis−1scvis−2MS4A1−10010−10 −5 0 5 10scvis−1scvis−2CD190 1 2 3ExpressionFigure 4.16: Expression of selected marker genes (CD2, CD3D, and CD3E for T cells; CD79A,MS4A1, and CD19 for B cells) in scvis embedding of reactive lymph node data.147IGHE STAG3IGLC3IGHG4SYNE2IGKCB2MIGHMPLAC8IGHDBCL2BCL6050100150200−2 0 2log2 (fold change)−log 10 (q−value)Significance P <= 0.05 P > 0.05FL1018aIGLC2FCRL2RGS13PCDH9TCL1AIGKC FCER2HLA−DRB1LTBHLA−DPA1BCL2BCL60100200300−4 −2 0 2 4log2 (fold change)−log 10 (q−value)Significance P <= 0.05 P > 0.05FL2001bFigure 4.17: Differential expression results for malignant vs. nonmalignant B cells in (a) FL1018and (b) FL2001. Comparisons was performed accounting for timepoint and potential interactionsbetween malignant status and timepoint using a multivariate linear model described in Methods.Genes upregulated among malignant cells have logFC values > 0. P -values were adjusted with theBenjamini-Hochberg method.1484.3.6 CellAssign uncovers compositional and phenotypic changes in the fol-licular lymphoma microenvironmentWe next asked how the relative abundance of each cell type differed after transformation orearly progression. While the overall proportion of B cells in both cases was higher in the secondtimepoint, the relative proportion of nonmalignant B cells decreased dramatically (FL1018:12.2% to 1.4%; FL2001: 44.4% to 1.4%) (Figure 4.14D). Thus, malignant B cells appear todominate the B cell population upon transformation or early progression. Among T cells, therelative proportions of each cell subtype were comparable in FL1018 and FL2001 at the firsttimepoint, with T follicular helper cells and CD4+ T cells composing the majority of T cells andcytotoxic T cells the minority (Figure 4.14E,F). However, compared to the consistent patternof B cell dynamics seen across both patients, T cell compositional dynamics upon transformationor early progression appeared to be divergent (Figure 4.14F). Cytotoxic T cells dominatedthe recurrence sample in FL1018, whereas T follicular helper cells became the major T cellpopulation following progression.Additionally, we looked at whether these compositional changes in the follicular lymphomamicroenvironment were accompanied by phenotypic changes in nonmalignant cell populationsbetween timepoints. To address this, we performed differential expression analysis acrosstimepoints for each patient and cell type separately (Supplemental Table B.3). In FL1018,differential expression analysis revealed upregulation of immune-associated pathways [203],such as cytokine signalling and toll-like receptor pathways, among cytotoxic T cells in thetransformed sample (Figure 4.18A). Similar results were observed for T follicular helperand CD4+ T cells (Supplemental Table B.3). T cell activation and effector moleculesincluding CD69, IFNG, GZMA, and PRF1 were also significantly upregulated in the secondtimepoint among cytotoxic T cells in FL1018 [219] (Figure 4.14G). Likewise, CD69 wassignificantly upregulated among T follicular helper and CD4+ T cells in the recurrence sample(all Q < 9.1e-07; Figure 4.19). Among cytotoxic T cells in FL2001, other immune relatedpathways such as antigen presentation and TCR/BCR signalling were upregulated in the earlyprogressed sample, but GZMA and IFNG were not significantly differentially expressed betweentimepoints (Figure 4.18A, Figure 4.14G). Ubiquitin-associated genes and pathways weresignificantly upregulated in T follicular helper cells and CD4+ T helper cells in FL2001 as well(Supplemental Table B.3). In nonmalignant B cells, no significantly differentially expressedpathways were observed in either patient apart from upregulation of general adaptive immunesystem genes encompassing canonical B cell markers such as CD79A, CD79B, and HLA moleculesin FL2001 (Supplemental Table B.3). Thus, transformation, and to a lesser extent early149progression, appears to be accompanied by T cell activation.Signaling by InterleukinsInterleukin−10 signalingG alpha (i) signalling eventsCytokine Signaling in Immune systemSenescence−Associated Secretory Phenotype (SASP)Platelet degranulation Response to elevated platelet cytosolic Ca2+Stimuli−sensing channelsOxidative Stress Induced Senesc nceMAPK family signaling cascadesTLR10 CascadeTLR5 CascadeMyD88 cascade initiated on plasma membraneTRAF6 mediated induction of NFkB and MAP kinases upon TLR7/8 or 9 activationTLR7/8 CascadeMyD88 dependent cascade initiated on endosomeMyD88:MAL(TIRAP) cascade initiated on plasma membraneTLR9 CascadeToll Like Receptor TLR1:TLR2 CascadeToll Like Receptor TLR6:TLR2 CascadeTLR2 CascadeNegative regulation of MAPK pathwayFc epsilon receptor (FCERI) signalingTLR3 CascadeMyD88−independent TLR4 cascade TRIF(TICAM1)−mediated TLR4 signaling RAF/MAP kinase cascadeMAPK1/MAPK3 signalingNF−kB is activated and signals survivalPeptide ligand−binding receptorsFL1018, CD8 T cellsa3e−04 1e−03 3e−03 1e−02 3e−02Adjusted p value Genes2 5 10Immunoregulatory interactions between a Lymphoid and a non−Lymphoid cellER−Phagosome pathwayAntigen processing−Cross presentationEPH−Ephrin signalingAxon guidanceBeta−catenin independent WNT signalingAdaptive Immune SystemPCP/CE pathwaySignaling by the B Cell Receptor (BCR)NOTCH3 Activation and Transmission of Signal to the NucleusRHO GTPases Activate ROCKsHost Interactions of HIV factorsSmooth Muscle ContractionRHO GTPases activate PAKsGlycogen metabolismRHO GTPase EffectorsDownstream signaling events of B Cell Receptor (BCR)Downstream TCR signalingRegulation of activated PAK−2p34 by proteasome mediated degradationVpu mediated degradation of CD4Autodegradation of the E3 ubiquitin ligase COP1Ubiquitin−dependent degradation of Cyclin D1Ubiquitin Mediated Degradation of Phosphorylated Cdc25Ap53−Independent DNA Damage Responsep53−Independent G1/S DNA damage checkpointUbiquitin−dependent degradation of Cyclin DRegulation of ApoptosisVif−mediated degradation of APOBEC3GAsymmetric localization of PCP proteinsFBXL7 down−regulates AURKA during mitotic entry and in early mitosisFL2001, CD8 T cellsbFigure 4.18: Significantly enriched Reactome pathways (BH-adjusted P -value ≤ 0.05) among thetop 50 most highly upregulated genes (ranked by log fold change) in (a) FL1018 and (b) FL2001.Up to 30 pathways are shown in either plot (Methods).150RGS1TSC22D3KLF6DUSP1EIF1HSPA1AHSPA1BEEF1GNBEAL1TOX2CD69051015−2 −1 0 1 2log2 (fold change)−log 10 (q−value)Significance P <= 0.05 P > 0.05T follicular helperaTSC22D3CXCR4EIF1BTG1 ZFP36L2HSPA1BHSPA1A NBEAL1EEF1GCCR7CD690102030−2 −1 0 1 2log2 (fold change)−log 10 (q−value)Significance P <= 0.05 P > 0.05CD4 T cellsbFigure 4.19: Differentially expressed genes for (a) T follicular helper and (b) other CD4 T cellsbetween T2 vs. T1. Genes upregulated in T2 have log fold change values > 0. The activation markerCD69 is highlighted. P -values were adjusted with the Benjamini-Hochberg method.1514.3.7 Malignant cell dynamics associated with early progression and trans-formationHaving explored the temporal dynamics of nonmalignant cells, we next investigated tran-scriptomic changes in malignant B cells. At a high level, malignant B cells from FL1018T1and FL1018T2 were more distinct in UMAP space and were less concordant than those fromFL2001T1 and FL2001T2 (Pearson’s correlation coefficient between mean sample expressionprofiles = 0.97 and 0.982 in FL1018 and FL2001, respectively), suggesting higher levels oftranscriptomic divergence upon transformation compared to early progression (Figure 4.14A).To analyze the nature of these differences in malignant cell transcriptomes, we performed differ-ential expression and gene set enrichment analysis of malignant B cells across timepoints usingcancer hallmark pathways (Supplemental Table B.3). Proliferation and cell cycle-associatedpathways, including MYC targets, E2F targets, and G2M checkpoint-associated genes, were sig-nificantly upregulated in the recurrence sample of FL1018 (all Q <0.0016), suggesting an increasein proliferative potential following transformation [204] (Figure 4.20A). While MYC targetswere also upregulated following recurrence in FL2001 (Q =0.0043), the cell cycle-associatedE2F and G2M pathways were not (all Q >0.34), and the cell cycle-associated mitotic spindlepathway was significantly downregulated (Q = 0.0043; Figure 4.20B). Based on these findings,we explored the distribution of cell cycle-associated genes in malignant cells. Overlaying theexpression of MKI67 and TOP2A onto the UMAP embedding for malignant B cells revealedputative replicative clusters in both patients (Figure 4.20C,E). In the transformed but notearly progressed case, a larger proportion of malignant cells from the recurrence sample appearedto be associated with these replicative clusters (Figure 4.20C,E), and differential expressionanalysis showed significant upregulation of MKI67 and TOP2A in the recurrence sample (all Q <7e-07. Correspondingly, cell cycle analysis with cyclone [200] revealed that a higher proportionof cycling (S or G2/M phase) malignant cells was present following transformation in FL1018(Figure 4.20D). In contrast, we observed a reduced proportion of cycling T follicular helper,cytotoxic, and CD4+ T cells in FL1018 (Figure 4.20D). Consistent with the findings frompathway analysis, there were fewer cycling malignant cells in the recurrence sample of FL2001(Figure 4.20F). Therefore, transformation, but not progression, appears to be associated withincreased replication among malignant cells.152INTERFERON_GAMMA_RESPONSEALLOGRAFT_REJECTIONAPICAL_JUNCTIONANDROGEN_RESPONSEKRAS_SIGNALING_UPCOMPLEMENTIL6_JAK_STAT3_SIGNALINGIL2_STAT5_SIGNALINGTNFA_SIGNALING_VIA_NFKBMYOGENESISMITOTIC_SPINDLEMTORC1_SIGNALINGADIPOGENESISHYPOXIAUNFOLDED_PROTEIN_RESPONSEGLYCOLYSISOXIDATIVE_PHOSPHORYLATIONG2M_CHECKPOINTMYC_TARGETS_V2E2F_TARGETSMYC_TARGETS_V1−2 −1 0 1 2Normalized Enrichment ScorePathwayaMITOTIC_SPINDLEUV_RESPONSE_DNG2M_CHECKPOINTE2F_TARGETSMTORC1_SIGNALINGALLOGRAFT_REJECTIONADIPOGENESISGLYCOLYSISFATTY_ACID_METABOLISMMYC_TARGETS_V1OXIDATIVE_PHOSPHORYLATION−2 −1 0 1 2Normalized Enrichment ScorePathwaybGene set size 50 70 100 SignificanceP <= 0.05 P > 0.05048−5 0 5UMAP−1UMAP−2c048−5 0 5UMAP−1UMAP−2MKI67048−5 0 5UMAP−1UMAP−2TOP2A0.250.300.350.400.45T1 T2TimepointProportion of cells in S/G2/MdTimepointT1 T2 0 1 2 3Expression CelltypeB cells (malignant)CD4 T cellsCytotoxic T cellsTfhB cellsTotal proportion0.01 0.10 0.30−6−303−5 0 5UMAP−1UMAP−2e−6−303−5 0 5UMAP−1UMAP−2MKI67−6−303−5 0 5UMAP−1UMAP−2TOP2A0.10.20.30.40.50.6T1 T2TimepointProportion of cells in S/G2/MfB2M HLA−A HLA−B HLA−C HLA−DRA HLA−DRB1FL1018FL2001malignant, T1malignant, T2nonmalignant, T1nonmalignant, T2malignant, T1malignant, T2nonmalignant, T1nonmalignant, T2malignant, T1malignant, T2nonmalignant, T1nonmalignant, T2malignant, T1malignant, T2nonmalignant, T1nonmalignant, T2malignant, T1malignant, T2nonmalignant, T1nonmalignant, T2malignant, T1malignant, T2nonmalignant, T1nonmalignant, T202460246B cell populationLog normalized countsgFigure 4.20: Temporal changes in malignant cells in the follicular lymphoma microenvironment.(a,b) Pathway enrichment scores computed by fGSEA for differentially enriched (adjusted P ≤ 0.05)and cell cycle-associated pathways among malignant cells between timepoints for (a) FL1018 and (b)FL2001 (Methods). Pathways with a positive enrichment score are upregulated in T2 comparedto T1 samples. P -values were adjusted with the Benjamini-Hochberg method. (c,e) UMAP plots,labeled by sample and proliferation marker expression (MKI67 and TOP2A), for (c) FL1018 and (e)FL2001. (d,f) Proportion of cells assigned to be in non-G1 cell cycle phases (S/G2/M) by cycloneacross timepoints in (d) FL1018 and (f) FL2001. (g) Normalized expression of HLA class I genesand select HLA class II genes across timepoints in FL1018 and FL2001.153Several immune-associated pathways, including complement and interferon gamma response, weredownregulated in the recurrence sample of FL1018 (Figure 4.20A, Supplemental Table B.3).To interpret these findings, we looked at genes that were most downregulated in the recurrencesample based on effect size and significance (Figure 4.21). Among these were HLA class I andII genes, including HLA-A, HLA-B, HLA-C, B2M, HLA-DRA, and HLA-DRB1. In order tofurther investigate the temporal dynamics of HLA expression, we analyzed HLA expression levelsin both nonmalignant and malignant B cells across timepoints. In both patients, HLA class I andII genes were expressed at lower levels in malignant B cells compared to nonmalignant B cells(all Q < 0.037; Figure 4.20G,H; Methods). Moreover, HLA expression levels in nonmalignantB cells were similar between timepoints (Figure 4.20G,H). However, while the expression ofHLA genes in malignant cells from FL2001 was comparable across timepoints, malignant cellsin the transformed case expressed significantly lower levels of HLA genes at recurrence (all Q <9.6e-24; Figure 4.20G,H). Corroborating these findings, differential expression and pathwayanalysis revealed that the HLA class I antigen presentation pathway was downregulated inmalignant cells from FL1018 upon transformation (Q = 0.019; Figure 4.22, SupplementalTable B.3). Coupled with the increase in cytotoxic proportion among T cells and upregulationof T cell activation markers in FL1018T2 compared to FL1018T1, these results are consistentwith immune escape in response to T cell activation following histologic transformation. Askingwhether these results could be explained by genomic changes in antigen processing or presentationgenes, we analyzed whole-genome sequencing data to profile somatic single nucleotide variants(SNVs), indels, and copy number alterations. No variants in these genes or loss-of-heterozygosityof HLA class I genes were detected in either sample from FL1018. However, coding mutationsin the histone acetyltransferase CREBBP, recently reported to be associated with HLA classII downregulation [220], were detected in all samples from both patients, providing a possibleexplanation for the lower HLA class II expression observed among malignant cells.154HLA−BB2MHLA−A HLA−DRB1HLA−DRAHLA−C CD740100200300−3 −2 −1 0 1 2 3log2 (fold change)−log 10 (q−value)Significance P <= 0.05 P > 0.05FL1018aB2MHLA−AHLA−CHLA−DRACD74HLA−BHLA−DRB1050100150200−2 −1 0 1 2log2 (fold change)−log 10 (q−value)Significance P <= 0.05 P > 0.05FL2001bFigure 4.21: Differentially expressed genes between malignant B cells from T2 vs. T1 in (a) FL1018and (b) FL2001. Genes upregulated in T2 have log fold change values > 0. HLA class I genes andselect HLA class II genes are highlighted. P -values were adjusted with the Benjamini-Hochbergmethod.155Interferon gamma signalingAttenuation phaseHSF1−dependent transactivation Cytokine Signaling in Immune systemInterferon SignalingAntigen Presentation: Folding, assembly and peptide loading of class I MHCFL1018, malignant B cells0.001 0.003 0.010Adjusted p value Genes5 10Figure 4.22: Significantly enriched Reactome pathways (BH-adjusted P -value ≤ 0.05) amongthe top 50 most downregulated genes (ranked by log fold change) in FL1018. No pathways weresignificantly downregulated in FL2001.1564.4 DiscussionWe developed a computational method to automatically annotate single cell RNA sequencingdata into cell types based on pre-defined marker gene information. Our approach systematicallydetermines cell type expression patterns and assignment probabilities based solely on theassumption that marker genes are highly expressed in their respective cell types, eliminatingthe need for manual cluster annotation or existing training data for cell type mapping methods.In simulations and on real scRNA-seq data from purified populations, CellAssign’s accuracywas comparable or superior to state-of-the-art workflows based on unsupervised clusteringand mapping methods, and ran in a minute on datasets of tens of thousands of cells. Weadditionally demonstrate how bulk RNA-seq data can enable marker gene identification foraccurate discrimination of phenotypically similar cell types with CellAssign.We subsequently applied CellAssign to dissect the microenvironment composition of spatially-and temporally-collected samples from HGSC and follicular lymphoma. We show how CellAssigncan not only delineate multiple malignant and nonmalignant epithelial, stromal, and immunecell types, but also identify subpopulations defined by arbitrary marker genes, uncoveringIGKC:IGLC ratios among nonmalignant B cells in follicular lymphoma consistent with thosefor normal lymphoid structures [215]. While these analyses are constrained by restricted cohortsize, they provide first-of-kind examples of spatiotemporal dynamics and microenvironmentinterplay interpreted through leveraging prior knowledge of cell types in a prinicipled statisticalapproach.We note that CellAssign is intended for scenarios where well understood marker genes exist.Poorly characterized cell types (or unknown cell types or cell states) may be invisible tothe CellAssign approach. Furthermore, we make no a priori distinction between “medium”or “high” expression of the same marker in two different cell types, though these could beincorporated by extending the model to accommodate constraints between different δ parameters.Nevertheless, we suggest a large proportion of clinical applications profiling complex tissuesstart with hypotheses relating the composition of known cell types to disease states. As such,CellAssign fills an important role in the scRNA-seq analysis toolbox, providing interpretableoutput from biologically motivated prior knowledge that is immune to common issues plaguingunsupervised clustering approaches [183].The volume of scRNA-seq data will increase over time in two important ways: (i) the numberof cell types profiled will increase, thereby expanding databases of known marker genes and(ii) scRNA-seq data will become more widely available in research and clinical settings [221].CellAssign is therefore poised to provide scalable, systematic and automated classification of157cells based on known parameters of interest, such as cell type, clone-specific markers, or genesassociated with drug response. Furthermore, by appropriately modifying the observation modelCellAssign can easily be extended to annotate cell types in data generated by other single-cellmeasurement technologies such as mass cytometry. We anticipate the CellAssign approach willhelp unlock the potential for large scale population-wide studies of cell composition of humandisease and other complex tissues through encoding biological prior knowledge in a robustprobabilistic framework.158Chapter 5Conclusions and Future DirectionsDespite extensive efforts to find effective therapies, prognosis for patients with high-grade serousovarian cancer remains bleak. High-grade serous ovarian cancer patients often present withclonally heterogeneous disease with spread to peritoneal sites including the contralateral ovary,omentum, and pelvic wall. The extent of intratumoural heterogeneity in HGSC is thoughtto contribute to the prevalence of recurrence following initial treatment with standard-of-carecombination platinum-taxane chemotherapy, with treatment resistant-clones escaping elimination.Nevertheless, the presence of TILs is associated with superior outcomes, hinting at the tantalizingpossibility that the immune system may be able to contend with intratumoural heterogeneity inHGSC.Consequently, the primary goal of this thesis was to profile the immune microenvironment, andmore broadly, the tumour microenvironment in HGSC. This would improve understanding of theunderlying spatial characteristics driving differential patterns of clonal seeding and proliferationin HGSC. In addition, given the recent success of immunotherapies in other cancer types,including checkpoint inhibitors in melanoma [222, 223] and CAR T-cells in B cell lymphomas[224], this work may help inform immunotherapeutic strategies for HGSC. The broad implicationsof this work are summarized below.Assembly of an extensively profiled and largest published multi-site cohort ofHGSC cases to date. Chapter 2 described the construction of our multi-site HGSC co-hort which I led, encompassing clinical case identification and sample processing for genomic,imaging, single cell, and patient-derived xenograft studies. The patient-derived xenografts willbe used to study clonal evolution in response to drug perturbation with cytotoxic and DNAdamaging agents such as platinum compounds and PARP inhibitors, and the single cell materialswill set the foundation for biological studies that leverage the CellAssign method describedin Chapter 4. This resource acts as the bedrock for the multimodal profiling study describedin Chapter 3, the single cell RNA-sequencing analysis of HGSC described in Chapter 4, andplanned future work (Section 5.1).Deciphering multiple interfaces of evolutionary interplay between tumour and im-159mune cells in HGSC. To investigate the interplay between immune and malignant cellsunderlying the spatial distribution of clones in HGSC, we profiled 212 multi-site HGSC samplesfrom 38 patients, including cases collected as part of Chapter 2 with whole-genome sequencing,targeted sequencing and clonal deconvolution, Nanostring expression assays, histologic imageanalysis, immunohistochemistry, and T- and B-cell receptor sequencing. From this work, Iuncovered 3 major subtypes of HGSC based on the spatial distribution of lymphocytes betweentumour epithelium and stroma, with imaging and sequencing-based evidence of tumour-immuneinteraction in extensively infiltrated (ES-TIL) samples. I extended and applied probabilisticmethods for clonal decomposition, revealing lower intratumoural heterogeneity among highlyinfiltrated samples, with some cases exhibiting loss-of-heterozygosity of HLA class I loci. Thefindings from this work include the novel discovery of immune escape mechanisms in HGSC (HLAloss-of-heterozygosity) despite T cell tracking of tumour clones among highly infiltrated tumoursamples, providing an explanation for the poor responses to immune checkpoint inhibitorsobserved in HGSC to date.Discovery of prognostically relevant associations between mutational processes andimmunologic signatures. In Chapter 3, I also applied a novel topic model-based approachdeveloped by [26] to quantify mutational signatures associated with defective DNA damage repairin HGSC. Building on first-of-kind work by [3] identifying 2 major genomic subtypes in HGSC, weidentified 4 major subtypes of HGSC, further subdividing the homologous recombination-deficientsubtype and introducing a new tandem duplicator subtype with distinct survival outcomesand immunologic infiltration patterns [26] previously classified as foldback inversion-type [1].Additionally, I describe a novel association between low immunologic infiltration and foldbackinversion-harbouring tumours, along with a prognostic association between immune infiltrationand mutational processes whereby immune infiltration is associated with superior overall survivalin HRD but not FBI cases. In light of recent success with PARP inhibitor maintenance therapyin HGSC [17], these findings provide context for combination immunotherapy-PARP inhibitortherapies in clinical trials.Developing an automated, scalable method for microenvironmental cell type iden-tification from single cell transcriptomic data. Extending the work I led on exploring theimmune composition of HGSC in Chapter 3, I developed a novel marker gene-based probabilisticapproach to identifying known cell types from single cell RNA-seq data (Chapter 4). TheCellAssign model uses a hierarchical Bayesian framework to systematically assign input cellsto known cell types while accounting for any additional experimental covariates such as batcheffects in a scalable, automated fashion. CellAssign only requires binary marker gene informa-tion, rather than the purified single cell RNA-seq data required by supervised methods, and160performs superiorly or comparably to state-of-the-art methods. I apply CellAssign to high-gradeserous ovarian cancer and follicular lymphoma, demonstrating the utility of CellAssign formicroenvironmental deconvolution and revealing interplay between malignant and immune cells.CellAssign sets the stage for scalable inference of cell types in large-scale single cell RNA-seqstudies of cancer that are beginning to emerge, including our ongoing single cell RNA-seq workin high-grade serous ovarian cancer.5.1 Future Directions5.1.1 Understanding the molecular basis of immunologic infiltration patternsin HGSCChapter 3 identified 3 major patterns of immunologic infiltration in HGSC characterized bythe absence of TIL (N-TIL), stromally-restricted TIL (S-TIL), and epithelial and stromal TIL(ES-TIL). While our work identified potential mechanisms by which ES-TIL tumours mayescape from immune recognition, it does not explain the lack of epithelial immune infiltration inN-TIL and S-TIL tumours. Tumours with lower immune infiltration generally harbour highersubclonal neoantigen loads, suggesting that antigen deficiency does not underlie the lack ofobserved immune infiltrate. Furthermore, anatomic location was not significantly associated withparticular patterns of immune infiltration. How S-TIL tumours manage to exclude TIL fromepithelial areas remains unknown. One possibility is that S-TIL tumours contain stromal barriersthat physically prevent TIL entry into epithelial regions. The predominance of fibroblasts inS-TIL tumours may inhibit T cell effector function through TGFβ production [225]. Recent workby [226] proposes an alternative cancer cell-mediated mechanism for epithelial TIL exclusion intriple-negative breast cancer, whereby immune infiltration in S-TIL patterns is more consistentwith the presence of an TIL repellent produced by tumour cells than with physical blockadeby desmoplastic elements. Moreover, the mechanisms that N-TIL tumours employ to escapefrom immune recognition remain a mystery. Laser capture microdissection-assisted RNA-seqand single cell RNA-sequencing may provide crucial insights into malignant or nonmalignantphenotypes that contribute to immune cell exclusion. These methods enable investigation ofsite-specific, clone-level phenotypes associated with N-TIL and S-TIL patterns. They also allowfor deeper investigation of other microenvironmental cell types classically associated with immuneregulation, including macrophages and fibroblasts. Multi-site cohorts, which intrinsically controlfor patient-specific factors and minimize batch effects associated with single cell processing,provide ideal substrate for these types of studies. Understanding the mechanisms underlying161immune exclusion in antigen-harbouring tumours may open therapeutic avenues for improvingthe delivery and efficacy of cancer immunotherapies.5.1.2 Deciphering the mechanisms of treatment resistance in HGSCThe patterns of clonal dynamics across space explored in Chapter 3 and [29] highlight theextensive intra- and intertumoural heterogeneity that exists at the time of diagnosis in HGSC.While it is believed that clonal heterogeneity underlies treatment resistance by providing geneticsubstrate for evolution to act upon, the molecular mechanisms that are ultimately responsiblefor recurrent disease in the context of treatment with standard-of-care chemotherapy, apartfrom secondary mutations in BRCA genes [15], remain largely unknown. In Chapter 3, wepresent evidence for subclonal immune escape, including neoantigen depletion and HLA loss-of-heterozygosity, that may allow certain clones to escape immune-mediated elimination. This mayhelp explain some cases of resistance, where platinum-taxane chemotherapy leads to cell death,inducing a systemic abscopal-like anticancer response. More generally however, phenotypicalterations in cancer cells that affect drug influx and efflux, metabolism, and proliferationmay lead to chemotherapeutic resistance. Drug challenge studies in model systems, such aspatient-derived xenografts (PDXs), provide controlled systems in which to study clonal dynamicsin response to treatment. Chapter 2 describes the construction of a first-of-kind multi-siteHGSC PDX cohort intended for this purpose.Broadly speaking, resistance mechanisms can be classified as intrinsic—encoded in the genome—or adaptive—mediated by epigenetic or context-specific factors, such as the microenvironment,and reflected in the transcriptome and proteome. The advent of scalable, low-bias single cellwhole-genome sequencing technologies for clonal reconstruction and single cell transcriptomesequencing technologies for malignant cell phenotyping can help identify rare resistant clones anddistinguish these two categories of resistance. This can be accomplished through experimentalintegration of single cell data from multiple modalities through combined genome-transcriptome-proteome sequencing or computational integration with methods such as clonealign [227]. Clonalpopulations that persist in drug-treated PDX models across multiple replicates and modelsderived from multiple diagnostic metastases may contain variants that can be profiled with thesemethods. Candidate variants could be further studied in cell lines with CRISPR perturbation.Emerging methods for fitness modeling from the population genetics literature can be leveragedto quantify the selective advantage conferred by each variant and genotype and predict futureclonal dynamics. Humanized mouse models can also be used to transplant human hematopoieticstem cells in PDXs, enabling study of clonal dynamics in similar microenvironmental contexts[228, 229].1625.1.3 Deconvolution of the HGSC microenvironmentDespite extensive literature profiling lymphocytes in HGSC [36, 38, 41] including our own(Chapter 3, [1]), few studies have characterized the cell type composition of the entire HGSCmicroenvironment. Other cell types, including macrophages and fibroblasts, are major compo-nents of the tumour microenvironment but remain largely uncharacterized in HGSC. One recentstudy attempted to address this question in a small cohort of epithelial ovarian cancers including5 HGSC cases [230], but this sample size was insufficient to capture single cell phenotypesreflecting the complete diversity of microenvironmental profiles according to bulk expressionprofiling [101]. Indeed, our initial investigation of the HGSC microenvironment using single cellRNA-seq revealed several additional populations, such as pericytes and non-collagen expressingovarian stromal cells (Chapter 4). Larger cohorts are needed to phenotypically and prognos-tically characterize these rare microenvironmental subpopulations. Given mounting evidenceimplicating the immune and non-immune microenvironment along with interactions betweenmicroenvironmental cell types in HGSC disease progression [1, 59, 147], treatment decisionsinformed by mathematical modeling of cancer clones (Section 5.1.2) will have to be interpretedin the context of the microenvironment. In Chapter 4, we developed an automated method forsystematically identifying cell types from single cell RNA-seq data. As single cell transcriptomestudies on large cohorts of HGSC tumours begin to emerge, our method will enable cell typeidentification at scale while controlling for batch effects.Furthermore, the roles of cell types that have been profiled in HGSC, particularly B cells andplasma cells, are unclear. While the presence of B and plasma cells is associated with superioroutcomes in tumours that contain T cells [38, 41], the mechanisms by which these cell types actare largely unknown. In a preprint related to Chapter 3 [231], I explored the phylogeographicsof B cell clones across space in multi-site HGSC, uncovering evidence suggesting that B cells areactive participants in the anticancer immune response. B cells may act as antigen presentingcells for T cells, produce antibodies against tumour antigens, or exert direct cytotoxic effects[42]. Single cell transcriptomics may help to resolve the role of individual B cell clones. The 10XChromium technology enables paired recovery of single cell transcriptomes and B cell receptorsequences, allowing B cell clones to be matched up to cellular phenotypes. This will enablephenotypic profiling of tumour-reactive B cell clones identified through clonal frequency analysisor antibody reactivity assays.Finally, molecular subtyping from bulk expression data has yielded 4 major subtypes of HGSCthat appear to be associated with microenvironmental features [14, 101]. Certain subtypes (C1[101], mesenchymal [14]) are generally associated with worse outcomes, though the significance163of this finding varies between studies. Single cell transcriptomics will uncover the cell types andcell type-specific phenotypes that ultimately underlie each subtype. Furthermore, single cellRNA-seq applied to large cohort studies will decode phenotypically-distinct subtypes of eachcell type, including cancer cells.5.1.4 In situ profiling of the tumour microenvironmentMulti-site studies of HGSC have reproducibly shown that the properties of both malignantcells and the surrounding microenvironment can differ appreciably between peritoneal foci,even within the same macroscopic tumour [1, 27, 29, 147, 232]. Tumours may contain bothimmune-privileged niches capable of supporting diverse clonal populations and highly infiltratedareas in which clonal pruning has occurred [1]. Beyond microenvironmental cell type composition(Chapter 3, Chapter 4), in situ spatial profiling of single cell phenotypes may yield importantinsights into microenvironment-malignant cell interactions. For example, the chemotherapeuticresistance properties associated with fibroblasts and abrogated by T cells in vitro [59] may bedependent on spatial proximity between fibroblasts, cancer cells, and T cells. Likewise, clonesbearing HLA loss-of-heterozygosity may reside in adjacent regions to cytotoxic T cells. Thesestudies may also help decipher the mechanisms behind TIL exclusion in N-TIL and S-TILtumours. Spatial transcriptomic profiling can be performed with techniques such as merFISH[233]. A recent study has described 3D intact-tissue spatial transcriptomics with a novel methodcalled STAR-MAP that leverages DNA barcoding with SEDAL sequencing to simultaneouslyobtain readouts of up to 1000 genes [234]. These technologies allow for deep investigation ofspatial interactions between microenvironmental cell types.Alternatively, spatial patterns in the tumour microenvironment can be studied at lower depthbut at scale. Cell type identification and pattern recognition from histologic images has beenused to prognostically stratify cancers [145, 167]. In Chapter 3, we profiled histologic imagesto identify cancer-immune cell hotspots associated with highly infiltrated tumours, suggestingdirect cell type interaction. Cell type interactions can be further studied using spatial statistics.For instance, Gibbs point process models model pairwise interactions between collections ofpoints, inferring the relative attractive or repulsive force between each pair of cell types. Immunerecognition of cancer cells may be read out as attractive forces between immune and cancercells. In summary, spatial profiling of histologic images from large cohorts may yield additionalinsights beyond those associated with lymphocyte abundance or TIL cluster (N-TIL, S-TIL,ES-TIL).1645.1.5 Guiding precision immunotherapies for HGSCCell-based immunotherapies, such as CAR T cell therapies, rely on selective expansion of immune-reactive T cell clones. Understanding the properties of these clones prior to perturbation, suchas immune exhaustion marker expression, may be important to stratifying patients likely torespond to cell-based immunotherapies. Reactive T cell clones can be isolated with traditionalMHC multimer assays, but transcriptome sequencing can only be done post-expansion whenphenotypes have likely changed substantially. Paired single cell RNA-sequencing and TCR-seqpresents a unique opportunity for phenotypic profiling of tumour-reactive T cells at diagnosis.Reactive T cell clones identified by MHC multimer or ELISPOT assays can be mapped to singlecell RNA-seq data using the cell-specific barcodes in the 10X Chromium protocol. This mayreveal unique properties of tumour-reactive T cells that can be deconvolved from bulk RNA-seqdata to stratify patients and inform immunotherapeutic options at the time of diagnosis.5.2 Concluding RemarksThe work presented in this thesis advances our understanding of clonal dynamics and the tumourmicroenvironment in HGSC. At the beginning of my thesis, I set out to understand the factorsinfluencing evolutionary dynamics across space in ovarian cancer. These studies leverage first-of-kind multimodal experimental design with spatial sampling to reveal some of the first evidence ofimmune-cancer evolution in ovarian cancer, and set the stage for further systematic investigationof the microenvironment in HGSC and other cancer types. The next major advances will beenabled by in situ methods that can provide spatial evidence of interaction within individualtumour samples, and temporal sampling to profile malignant-immune evolutionary dynamicsin recurrent disease. Resolving the tumour-microenvironment interface in HGSC will pavethe way for therapeutic options that exploit non-malignant cell types to overcome the clonalheterogeneity pervasive to the disease.165Bibliography[1] Allen W. Zhang, Andrew McPherson, Katy Milne, David R. Kroeger, Phineas T. Hamilton,Alex Miranda, Tyler Funnell, Nicole Little, Camila P.E. de Souza, Sonya Laan, StaceyLeDoux, Dawn R. Cochrane, Jamie L.P. Lim, Winnie Yang, Andrew Roth, Maia A. Smith,Julie Ho, Kane Tse, Thomas Zeng, Inna Shlafman, Michael R. Mayo, Richard Moore,Henrik Failmezger, Andreas Heindl, Yi Kan Wang, Ali Bashashati, Diljot S. Grewal,Scott D. Brown, Daniel Lai, Adrian N.C. Wan, Cydney B. Nielsen, Curtis Huebner, BasileTessier-Cloutier, Michael S. Anglesio, Alexandre Bouchard-Coˆte´, Yinyin Yuan, Wyeth W.Wasserman, C. Blake Gilks, Anthony N. Karnezis, Samuel Aparicio, Jessica N. McAlpine,David G. Huntsman, Robert A. Holt, Brad H. Nelson, and Sohrab P. Shah. Interfaces ofMalignant and Immunologic Clonal Dynamics in Ovarian Cancer. Cell, 173(7):1755–1769,6 2018.[2] Allen W Zhang, Ciara O’Flanagan, Elizabeth Chavez, Jamie LP Lim, Andrew McPherson,Matt Wiens, Pascale Walters, Tim Chan, Brittany Hewitson, Daniel Lai, Anja Mottok,Clementine Sarkozy, Lauren Chong, Tomohiro Aoki, Xuehai Wang, Andrew P Weng,Jessica N McAlpine, Samuel Aparicio, Christian Steidl, Kieran R Campbell, and Sohrab PShah. Probabilistic cell type assignment of single-cell transcriptomic data reveals spa-tiotemporal microenvironment dynamics in human cancers. bioRxiv, page 521914, 12019.[3] Yi Kan Wang, Ali Bashashati, Michael S Anglesio, Dawn R Cochrane, Diljot S Grewal,Gavin Ha, Andrew McPherson, Hugo M Horlings, Janine Senz, Leah M Prentice, An-thony N Karnezis, Daniel Lai, Mohamed R Aniba, Allen W Zhang, Karey Shumansky,Celia Siu, Adrian Wan, Melissa K McConechy, Hector Li-Chang, Alicia Tone, DianeProvencher, Manon de Ladurantaye, Hubert Fleury, Aikou Okamoto, Satoshi Yanagida,Nozomu Yanaihara, Misato Saito, Andrew J Mungall, Richard Moore, Marco A Marra,C Blake Gilks, Anne-Marie Mes-Masson, Jessica N McAlpine, Samuel Aparicio, David GHuntsman, and Sohrab P Shah. Genomic consequences of aberrant DNA repair mechanismsstratify ovarian cancer histotypes. Nature Genetics, 4 2017.[4] Roger Collier. Half of Canadians can expect cancer diagnosis during lifetime. CMAJ :166Canadian Medical Association journal = journal de l’Association medicale canadienne,189(27):E920, 7 2017.[5] Mel Greaves and Carlo C. Maley. Clonal evolution in cancer. Nature, 481(7381):306–313,1 2012.[6] Nicholas A Saunders, Fiona Simpson, Erik W Thompson, Michelle M Hill, Liliana Endo-Munoz, Graham Leggatt, Rodney F Minchin, and Alexander Guminski. Role of intra-tumoural heterogeneity in cancer drug resistance: molecular and clinical perspectives.EMBO molecular medicine, 4(8):675–84, 8 2012.[7] Christian Frantz, Kathleen M Stewart, and Valerie M Weaver. The extracellular matrixat a glance. Journal of cell science, 123(Pt 24):4195–200, 12 2010.[8] Padmanee Sharma and James P Allison. Immune checkpoint targeting in cancer therapy:toward combination strategies with curative potential. Cell, 161(2):205–14, 4 2015.[9] Stephanie C Casey, Amedeo Amedei, Katia Aquilano, Asfar S Azmi, Fabian Benencia,Dipita Bhakta, Alan E Bilsland, Chandra S Boosani, Sophie Chen, Maria Rosa Ciriolo,Sarah Crawford, Hiromasa Fujii, Alexandros G Georgakilas, Gunjan Guha, Dorota Halicka,William G Helferich, Petr Heneberg, Kanya Honoki, W Nicol Keith, Sid P Kerkar, Sulma IMohammed, Elena Niccolai, Somaira Nowsheen, H P Vasantha Rupasinghe, Abbas Samadi,Neetu Singh, Wamidh H Talib, Vasundara Venkateswaran, Richard L Whelan, XujuanYang, and Dean W Felsher. Cancer prevention and therapy through the modulation ofthe tumor microenvironment. Seminars in cancer biology, 35 Suppl:199–223, 12 2015.[10] Rebecca L. Siegel, Kimberly D. Miller, and Ahmedin Jemal. Cancer statistics, 2016. CA:A Cancer Journal for Clinicians, 66(1):7–30, 1 2016.[11] Martine J. Piccart. Response: Re: Randomized Intergroup Trial of CisplatinPaclitaxelVersus CisplatinCyclophosphamide in Women With Advanced Epithelial Ovarian Cancer:Three-Year Results. JNCI: Journal of the National Cancer Institute, 92(17):1446–1447, 92000.[12] Rosemary D. Cress, Yingjia S. Chen, Cyllene R. Morris, Megan Petersen, and Gary S. Leis-erowitz. Characteristics of Long-Term Survivors of Epithelial Ovarian Cancer. Obstetrics& Gynecology, 126(3):491–497, 9 2015.167[13] Mirjana Kessler, Christina Fotopoulou, and Thomas Meyer. The molecular fingerprint ofhigh grade serous ovarian cancer reflects its fallopian tube origin. International journal ofmolecular sciences, 14(4):6571–96, 3 2013.[14] D. Bell, A. Berchuck, M. Birrer, J. Chien, D. W. Cramer, F. Dao, R. Dhir, P. DiSaia,H. Gabra, P. Glenn, A. K. Godwin, J. Gross, L. Hartmann, M. Huang, D. G. Huntsman,M. Iacocca, M. Imielinski, S. Kalloger, B. Y. Karlan, D. A. Levine, G. B. Mills, C. Morrison,D. Mutch, N. Olvera, S. Orsulic, K. Park, N. Petrelli, B. Rabeno, J. S. Rader, B. I. Sikic,K. Smith-McCune, A. K. Sood, D. Bowtell, R. Penny, J. R. Testa, K. Chang, H. H. Dinh,J. A. Drummond, G. Fowler, P. Gunaratne, A. C. Hawes, C. L. Kovar, L. R. Lewis, M. B.Morgan, I. F. Newsham, J. Santibanez, J. G. Reid, L. R. Trevino, Y.-Q. Wu, M. Wang,D. M. Muzny, D. A. Wheeler, R. A. Gibbs, G. Getz, M. S. Lawrence, K. Cibulskis, A. Y.Sivachenko, C. Sougnez, D. Voet, J. Wilkinson, T. Bloom, K. Ardlie, T. Fennell, J. Baldwin,S. Gabriel, E. S. Lander, L. Ding, R. S. Fulton, D. C. Koboldt, M. D. McLellan, T. Wylie,J. Walker, M. OLaughlin, D. J. Dooling, L. Fulton, R. Abbott, N. D. Dees, Q. Zhang,C. Kandoth, M. Wendl, W. Schierding, D. Shen, C. C. Harris, H. Schmidt, J. Kalicki, K. D.Delehaunty, C. C. Fronick, R. Demeter, L. Cook, J. W. Wallis, L. Lin, V. J. Magrini, J. S.Hodges, J. M. Eldred, S. M. Smith, C. S. Pohl, F. Vandin, B. J. Raphael, G. M. Weinstock,E. R. Mardis, R. K. Wilson, M. Meyerson, W. Winckler, G. Getz, R. G. W. Verhaak, S. L.Carter, C. H. Mermel, G. Saksena, H. Nguyen, R. C. Onofrio, M. S. Lawrence, D. Hubbard,S. Gupta, A. Crenshaw, A. H. Ramos, K. Ardlie, L. Chin, A. Protopopov, Juinhua Zhang,T. M. Kim, I. Perna, Y. Xiao, H. Zhang, G. Ren, N. Sathiamoorthy, R. W. Park, E. Lee,P. J. Park, R. Kucherlapati, D. M. Absher, L. Waite, G. Sherlock, J. D. Brooks, J. Z. Li,J. Xu, R. M. Myers, P. W. Laird, L. Cope, J. G. Herman, H. Shen, D. J. Weisenberger,H. Noushmehr, F. Pan, T. Triche Jr, B. P. Berman, D. J. Van Den Berg, J. Buckley, S. B.Baylin, P. T. Spellman, E. Purdom, P. Neuvial, H. Bengtsson, L. R. Jakkula, S. Durinck,J. Han, S. Dorton, H. Marr, Y. G. Choi, V. Wang, N. J. Wang, J. Ngai, J. G. Conboy,B. Parvin, H. S. Feiler, T. P. Speed, J. W. Gray, D. A. Levine, N. D. Socci, Y. Liang, B. S.Taylor, N. Schultz, L. Borsu, A. E. Lash, C. Brennan, A. Viale, C. Sander, M. Ladanyi,K. A. Hoadley, S. Meng, Y. Du, Y. Shi, L. Li, Y. J. Turman, D. Zang, E. B. Helms,S. Balu, X. Zhou, J. Wu, M. D. Topal, D. N. Hayes, C. M. Perou, G. Getz, D. Voet,G. Saksena, Junihua Zhang, H. Zhang, C. J. Wu, S. Shukla, K. Cibulskis, M. S. Lawrence,A. Sivachenko, R. Jing, R. W. Park, Y. Liu, P. J. Park, M. Noble, L. Chin, H. Carter,D. Kim, R. Karchin, P. T. Spellman, E. Purdom, P. Neuvial, H. Bengtsson, S. Durinck,J. Han, J. E. Korkola, L. M. Heiser, R. J. Cho, Z. Hu, B. Parvin, T. P. Speed, J. W. Gray,168N. Schultz, E. Cerami, B. S. Taylor, A. Olshen, B. Reva, Y. Antipin, R. Shen, P. Mankoo,R. Sheridan, G. Ciriello, W. K. Chang, J. A. Bernanke, L. Borsu, D. A. Levine, M. Ladanyi,C. Sander, D. Haussler, C. C. Benz, J. M. Stuart, S. C. Benz, J. Z. Sanborn, C. J. Vaske,J. Zhu, C. Szeto, G. K. Scott, C. Yau, K. A. Hoadley, Y. Du, S. Balu, D. N. Hayes,C. M. Perou, M. D. Wilkerson, N. Zhang, R. Akbani, K. A. Baggerly, W. K. Yung, G. B.Mills, J. N. Weinstein, R. Penny, T. Shelton, D. Grimm, M. Hatfield, S. Morris, P. Yena,P. Rhodes, M. Sherman, J. Paulauskis, S. Millis, A. Kahn, J. M. Greene, R. Sfeir, M. A.Jensen, J. Chen, J. Whitmore, S. Alonso, J. Jordan, A. Chu, Jinghui Zhang, A. Barker,C. Compton, G. Eley, M. Ferguson, P. Fielding, D. S. Gerhard, R. Myles, C. Schaefer,K. R. Mills Shaw, J. Vaught, J. B. Vockley, P. J. Good, M. S. Guyer, B. Ozenberger,J. Peterson, and E. Thomson. Integrated genomic analyses of ovarian carcinoma. Nature,474(7353):609–615, 6 2011.[15] Elizabeth M Swisher, Wataru Sakai, Beth Y Karlan, Kaitlyn Wurz, Nicole Urban, andToshiyasu Taniguchi. Secondary BRCA1 mutations in BRCA1-mutated ovarian carcinomaswith platinum resistance. Cancer research, 68(8):2581–6, 4 2008.[16] Mansoor R. Mirza, Bradley J. Monk, Jrn Herrstedt, Amit M. Oza, Sven Mahner, AndrsRedondo, Michel Fabbro, Jonathan A. Ledermann, Domenica Lorusso, Ignace Vergote,Noa E. Ben-Baruch, Christian Marth, Radosaw Ma¸dry, Ren D. Christensen, Jonathan S.Berek, Anne Dørum, Anna V. Tinker, Andreas du Bois, Antonio Gonza´lez-Mart´ın, PhilippeFollana, Benedict Benigno, Per Rosenberg, Lucy Gilbert, Bobbie J. Rimel, Joseph Buscema,John P. Balser, Shefali Agarwal, and Ursula A. Matulonis. Niraparib Maintenance Therapyin Platinum-Sensitive, Recurrent Ovarian Cancer. New England Journal of Medicine,375(22):2154–2164, 12 2016.[17] Kathleen Moore, Nicoletta Colombo, Giovanni Scambia, Byoung-Gie Kim, Ana Oaknin,Michael Friedlander, Alla Lisyanskaya, Anne Floquet, Alexandra Leary, Gabe S. Sonke,Charlie Gourley, Susana Banerjee, Amit Oza, Antonio Gonza´lez-Mart´ın, Carol Aghajanian,William Bradley, Cara Mathews, Joyce Liu, Elizabeth S. Lowe, Ralph Bloomfield, andPaul DiSilvestro. Maintenance Olaparib in Patients with Newly Diagnosed AdvancedOvarian Cancer. New England Journal of Medicine, page NEJMoa1810858, 10 2018.[18] Robert E. (Robert Edward) Scully, Robert H. (Robert Henry) Young, Philip B. Clement,Armed Forces Institute of Pathology (U.S.), and Universities Associated for Research andEducation in Pathology. Tumors of the ovary, maldeveloped gonads, fallopian tube, andbroad ligament. Armed Forces Institute of Pathology, 1998.169[19] Joseph W. Carlson, Alexander Miron, Elke A. Jarboe, Mana M. Parast, Michelle S.Hirsch, Yonghee Lee, Michael G. Muto, David Kindelberger, and Christopher P. Crum.Serous Tubal Intraepithelial Carcinoma: Its Potential Role in Primary Peritoneal SerousCarcinoma and Serous Cancer Prevention. Journal of Clinical Oncology, 26(25):4160–4165,9 2008.[20] Michael H. Roh, David Kindelberger, and Christopher P. Crum. Serous Tubal Intraep-ithelial Carcinoma and the Dominant Ovarian Mass. The American Journal of SurgicalPathology, 33(3):376–383, 3 2009.[21] S. Intidhar Labidi-Galy, Eniko Papp, Dorothy Hallberg, Noushin Niknafs, Vilmos Adleff,Michael Noe, Rohit Bhattacharya, Marian Novak, Sin Jones, Jillian Phallen, Carolyn A.Hruban, Michelle S. Hirsch, Douglas I. Lin, Lauren Schwartz, Cecile L. Maire, Jean-Christophe Tille, Michaela Bowden, Ayse Ayhan, Laura D. Wood, Robert B. Scharpf,Robert Kurman, Tian-Li Wang, Ie-Ming Shih, Rachel Karchin, Ronny Drapkin, andVictor E. Velculescu. High grade serous ovarian carcinomas originate in the fallopian tube.Nature Communications, 8(1):1093, 12 2017.[22] Ahmed Ashour Ahmed, Dariush Etemadmoghadam, Jillian Temple, Andy G Lynch,Mohamed Riad, Raghwa Sharma, Colin Stewart, Sian Fereday, Carlos Caldas, AnnaDefazio, David Bowtell, and James D Brenton. Driver mutations in TP53 are ubiquitousin high grade serous carcinoma of the ovary. The Journal of pathology, 221(1):49–56, 52010.[23] Geoff Macintyre, Teodora E. Goranova, Dilrini De Silva, Darren Ennis, Anna M. Piskorz,Matthew Eldridge, Daoud Sie, Liz-Anne Lewsley, Aishah Hanif, Cheryl Wilson, SuzanneDowson, Rosalind M. Glasspool, Michelle Lockley, Elly Brockbank, Ana Montes, AxelWalther, Sudha Sundar, Richard Edmondson, Geoff D. Hall, Andrew Clamp, CharlieGourley, Marcia Hall, Christina Fotopoulou, Hani Gabra, James Paul, Anna Supernat,David Millan, Aoisha Hoyle, Gareth Bryson, Craig Nourse, Laura Mincarelli, Luis NavarroSanchez, Bauke Ylstra, Mercedes Jimenez-Linan, Luiza Moore, Oliver Hofmann, FlorianMarkowetz, Iain A. McNeish, and James D. Brenton. Copy number signatures andmutational processes in ovarian carcinoma. Nature Genetics, 50(9):1262–1270, 9 2018.[24] Peter J. Campbell, Shinichi Yachida, Laura J. Mudie, Philip J. Stephens, Erin D. Pleasance,Lucy A. Stebbings, Laura A. Morsberger, Calli Latimer, Stuart McLaren, Meng-Lay Lin,David J. McBride, Ignacio Varela, Serena A. Nik-Zainal, Catherine Leroy, Mingming Jia,170Andrew Menzies, Adam P. Butler, Jon W. Teague, Constance A. Griffin, John Burton,Harold Swerdlow, Michael A. Quail, Michael R. Stratton, Christine Iacobuzio-Donahue,and P. Andrew Futreal. The patterns and dynamics of genomic instability in metastaticpancreatic cancer. Nature, 467(7319):1109–1113, 10 2010.[25] Tatiana Popova, Elodie Manie´, Valentina Boeva, Aude Battistella, Oumou Goundiam,Nicholas K. Smith, Christopher R. Mueller, Virginie Raynal, Odette Mariani, XavierSastre-Garau, and Marc-Henri Stern. Ovarian Cancers Harboring Inactivating Mutationsin CDK12 Display a Distinct Genomic Instability Pattern Characterized by Large TandemDuplications. Cancer Research, 76(7):1882–1891, 4 2016.[26] Tyler Funnell, Allen Zhang, Yu-Jia Shiah, Diljot Grewal, Robert Lesurf, Steven McKinney,Ali Bashashati, Yi Kan Wang, Paul Boutros, and Sohrab Shah. Integrated single-nucleotideand structural variation signatures of DNA-repair deficient human cancers. bioRxiv, 267500,2 2018.[27] Ali Bashashati, Gavin Ha, Alicia Tone, Jiarui Ding, Leah M Prentice, Andrew Roth,Jamie Rosner, Karey Shumansky, Steve Kalloger, Janine Senz, Winnie Yang, MelissaMcConechy, Nataliya Melnyk, Michael Anglesio, Margaret T Y Luk, Kane Tse, ThomasZeng, Richard Moore, Yongjun Zhao, Marco A Marra, Blake Gilks, Stephen Yip, David GHuntsman, Jessica N McAlpine, and Sohrab P Shah. Distinct evolutionary trajectories ofprimary high-grade serous ovarian cancers revealed through spatial mutational profiling.The Journal of pathology, 231(1):21–34, 9 2013.[28] Roland F. Schwarz, Charlotte K. Y. Ng, Susanna L. Cooke, Scott Newman, Jillian Temple,Anna M. Piskorz, Davina Gale, Karen Sayal, Muhammed Murtaza, Peter J. Baldwin,Nitzan Rosenfeld, Helena M. Earl, Evis Sala, Mercedes Jimenez-Linan, Christine A.Parkinson, Florian Markowetz, James D. Brenton, PC Nowell, DL Dexter, HM Kowalski,BA Blazar, Z Fligiel, R Vogel, L Khalique, A Ayhan, ME Weale, IJ Jacobs, SJ Ramus,L Khalique, A Ayhan, JC Whittaker, N Singh, IJ Jacobs, SP Shah, RD Morin, J Khattra,L Prentice, T Pugh, N Navin, A Krasnitz, L Rodgers, K Cook, J Meth, PJ Campbell,S Yachida, LJ Mudie, PJ Stephens, ED Pleasance, N Navin, J Kendall, J Troge, P Andrews,L Rodgers, A Marusyk, V Almendro, K Polyak, JS Vermaat, IJ Nijman, MJ Koudijs,FL Gerritse, SJ Scherer, X Wu, PA Northcott, A Dubuc, AJ Dupuy, DJH Shih, M Gerlinger,AJ Rowan, S Horswell, J Larkin, D Endesfelder, SP Shah, A Roth, R Goya, A Oloumi,G Ha, EC de Bruin, N McGranahan, R Mitter, M Salm, DC Wedge, S Nik-Zainal, P VanLoo, DC Wedge, LB Alexandrov, CD Greenman, LMF Merlo, JW Pepper, BJ Reid,171CC Maley, M Greaves, CC Maley, K Anderson, C Lutz, FW van Delft, CM Bateman,Y Guo, DA Landau, SL Carter, P Stojanov, A McKenna, K Stevenson, AA Ahmed,D Etemadmoghadam, J Temple, AG Lynch, M Riad, KL Gorringe, S Jacobs, ER Thompson,A Sridhar, W Qiu, N Sangha, R Wu, R Kuick, S Powers, D Mu, SL Carter, K Cibulskis,E Helman, A McKenna, H Shen, A Bashashati, G Ha, A Tone, J Ding, LM Prentice,J Zhang, Y Shi, E Lalonde, L Li, L Cavallone, M Hoogstraat, MS de Pagter, GA Cirkel,MJ van Roosmalen, TT Harkins, SL Cooke, CKY Ng, N Melnyk, MJ Garcia, T Hardcastle,SL Cooke, JD Brenton, ZC Wang, NJ Birkbak, AC Culhane, R Drapkin, A Fatima,PA Cowin, J George, S Fereday, E Loehrer, P Van Loo, RF Schwarz, A Trinh, B Sipos,JD Brenton, N Goldman, CD Greenman, G Bignell, A Butler, S Edkins, J Hinton, CKYNg, SL Cooke, K Howe, S Newman, J Xian, H Li, R Durbin, A Untergasser, I Cutcutache,T Koressaar, J Ye, BC Faircloth, T Forshew, M Murtaza, C Parkinson, D Gale, DWY Tsui,JT Robinson, H Thorvaldsdo´ttir, W Winckler, M Guttman, ES Lander, KM Archibald,H Kulbe, J Kwong, P Chakravarty, J Temple, E Sala, MY Kataoka, AN Priest, AB Gill,MA McLean, DJ McBride, D Etemadmoghadam, SL Cooke, K Alsop, J George, D Bryant,V Moulton, M Castellarin, K Milne, T Zeng, K Tse, and M Mayo. Spatial and TemporalHeterogeneity in High-Grade Serous Ovarian Cancer: A Phylogenetic Analysis. PLOSMedicine, 12(2):e1001789, 2 2015.[29] Andrew McPherson, Andrew Roth, Emma Laks, Tehmina Masud, Ali Bashashati, Allen WZhang, Gavin Ha, Justina Biele, Damian Yap, Adrian Wan, Leah M Prentice, JaswinderKhattra, Maia A Smith, Cydney B Nielsen, Sarah C Mullaly, Steve Kalloger, AnthonyKarnezis, Karey Shumansky, Celia Siu, Jamie Rosner, Hector Li Chan, Julie Ho, NataliyaMelnyk, Janine Senz, Winnie Yang, Richard Moore, Andrew J Mungall, Marco A Marra,Alexandre Bouchard-Coˆte´, C Blake Gilks, David G Huntsman, Jessica N McAlpine, SamuelAparicio, and Sohrab P Shah. Divergent modes of clonal spread and intraperitoneal mixingin high-grade serous ovarian cancer. Nature Genetics, 48(7):758–767, 5 2016.[30] Yuliya Klymenko, Jeffrey Johnson, Brandi Bos, Rachel Lombard, Leigh Campbell, Eliza-beth Loughran, and M. Sharon Stack. Heterogeneous Cadherin Expression and Multicel-lular Aggregate Dynamics in Ovarian Cancer Dissemination. Neoplasia, 19(7):549–563, 72017.[31] Sara Al Habyan, Christina Kalos, Joseph Szymborski, and Luke McCaffrey. Multicellu-lar detachment generates metastatic spheroids during intra-abdominal dissemination inepithelial ovarian cancer. Oncogene, 37(37):5127–5135, 9 2018.172[32] Daniela F Quail and Johanna A Joyce. Microenvironmental regulation of tumor progressionand metastasis. Nature medicine, 19(11):1423–37, 11 2013.[33] Costas A Lyssiotis and Alec C Kimmelman. Metabolic Interactions in the Tumor Microen-vironment. Trends in cell biology, 27(11):863–875, 11 2017.[34] Brad H Nelson, Philip D Greenberg, and Hans Schreiber. New insights into tumorimmunity revealed by the unique genetic and genomic aspects of ovarian cancer. CurrentOpinion in Immunology, 33:93–100, 2015.[35] Jean M. Hansen, Robert L. Coleman, and Anil K. Sood. Targeting the tumour microenvi-ronment in ovarian cancer. European Journal of Cancer, 56:131–143, 3 2016.[36] Lin Zhang, Jose R Conejo-Garcia, Dionyssios Katsaros, Phyllis A Gimotty, Marco Masso-brio, Giorgia Regnani, Antonis Makrigiannakis, Heidi Gray, Katia Schlienger, Michael NLiebman, Stephen C Rubin, and George Coukos. Intratumoral T cells, recurrence, andsurvival in epithelial ovarian cancer. The New England journal of medicine, 348(3):203–13,1 2003.[37] Sine Hadrup, Marco Donia, and Per Thor Straten. Effector CD4 and CD8 T cells andtheir role in the tumor microenvironment. Cancer microenvironment : official journal ofthe International Cancer Microenvironment Society, 6(2):123–33, 8 2013.[38] Julie S. Nielsen, Rob A. Sahota, Katy Milne, Sara E. Kost, Nancy J. Nesslinger, Peter H.Watson, and Brad H. Nelson. CD20+ Tumor-Infiltrating Lymphocytes Have an AtypicalCD27- Memory Phenotype and Together with CD8+ T Cells Promote Favorable Prognosisin Ovarian Cancer. Clinical Cancer Research, 18(12), 2012.[39] Darin A Wick, John R Webb, Julie S Nielsen, Spencer D Martin, David R Kroeger, KatyMilne, Mauro Castellarin, Kwame Twumasi-Boateng, Peter H Watson, Rob A Holt, andBrad H Nelson. Surveillance of the tumor mutanome by T cells during progression fromprimary to recurrent ovarian cancer. Clinical cancer research : an official journal of theAmerican Association for Cancer Research, 20(5):1125–34, 3 2014.[40] J. R. Webb, K. Milne, P. Watson, R. J. deLeeuw, and B. H. Nelson. Tumor-InfiltratingLymphocytes Expressing the Tissue Resident Memory Marker CD103 Are Associatedwith Increased Survival in High-Grade Serous Ovarian Cancer. Clinical Cancer Research,20(2):434–444, 1 2014.173[41] David R Kroeger, Katy Milne, and Brad H Nelson. Tumor-Infiltrating Plasma Cells AreAssociated with Tertiary Lymphoid Structures, Cytolytic T-Cell Responses, and SuperiorPrognosis in Ovarian Cancer. Clinical cancer research : an official journal of the AmericanAssociation for Cancer Research, 22(12):3005–15, 6 2016.[42] Brad H Nelson. CD20+ B cells: the other tumor-infiltrating lymphocytes. Journal ofimmunology (Baltimore, Md. : 1950), 185(9):4977–82, 11 2010.[43] Tyler J Curiel, George Coukos, Linhua Zou, Xavier Alvarez, Pui Cheng, Peter Mottram,Melina Evdemon-Hogan, Jose R Conejo-Garcia, Lin Zhang, Matthew Burow, Yun Zhu,Shuang Wei, Ilona Kryczek, Ben Daniel, Alan Gordon, Leann Myers, Andrew Lackner,Mary L Disis, Keith L Knutson, Lieping Chen, and Weiping Zou. Specific recruitmentof regulatory T cells in ovarian carcinoma fosters immune privilege and predicts reducedsurvival. Nature Medicine, 10(9):942–949, 9 2004.[44] Claudia C. Preston, Matthew J. Maurer, Ann L. Oberg, Daniel W. Visscher, Kimberly R.Kalli, Lynn C. Hartmann, Ellen L. Goode, and Keith L. Knutson. The Ratios of CD8+T Cells to CD4+CD25+ FOXP3+ and FOXP3- T Cells Correlate with Poor ClinicalOutcome in Human Serous Ovarian Cancer. PLoS ONE, 8(11):e80063, 11 2013.[45] Y Jiang, Y Li, and B Zhu. T-cell exhaustion in the tumor microenvironment. Cell Death& Disease, 6(6):e1792–e1792, 6 2015.[46] Fenne L Komdeur, Maartje C A Wouters, Hagma H Workel, Aline M Tijans, AnoukL J Terwindt, Kim L Brunekreeft, Annechien Plat, Harry G Klip, Florine A Eggink,Ninke Leffers, Wijnand Helfrich, Douwe F Samplonius, Edwin Bremer, G Bea A Wisman,Toos Daemen, Evelien W Duiker, Harry Hollema, Hans W Nijman, and Marco de Bruyn.CD103+ intraepithelial T cells in high-grade serous ovarian cancer are phenotypicallydiverse TCRαβ+ CD8αβ+ T cells that can be targeted for cancer immunotherapy.Oncotarget, 7(46):75130–75144, 11 2016.[47] John R. Webb, Katy Milne, David R. Kroeger, and Brad H. Nelson. PD-L1 expressionis associated with tumor-infiltrating T cells and favorable prognosis in high-grade serousovarian cancer. Gynecologic Oncology, 141(2):293–302, 5 2016.[48] Dung T. Le, Jennifer N. Uram, Hao Wang, Bjarne R. Bartlett, Holly Kemberling, Aleksan-dra D. Eyring, Andrew D. Skora, Brandon S. Luber, Nilofer S. Azad, Dan Laheru, BarbaraBiedrzycki, Ross C. Donehower, Atif Zaheer, George A. Fisher, Todd S. Crocenzi, James J.174Lee, Steven M. Duffy, Richard M. Goldberg, Albert de la Chapelle, Minori Koshiji, FeriylBhaijee, Thomas Huebner, Ralph H. Hruban, Laura D. Wood, Nathan Cuka, Drew M.Pardoll, Nickolas Papadopoulos, Kenneth W. Kinzler, Shibin Zhou, Toby C. Cornish,Janis M. Taube, Robert A. Anders, James R. Eshleman, Bert Vogelstein, and Luis A.Diaz. PD-1 Blockade in Tumors with Mismatch-Repair Deficiency. New England Journalof Medicine, 372(26):2509–2520, 6 2015.[49] Krisztian Homicsko and George Coukos. Targeting Programmed Cell Death 1 in OvarianCancer. Journal of Clinical Oncology, 33(34):3987–3989, 12 2015.[50] Stphanie L. Gaillard, Angeles A. Secord, and Bradley Monk. The role of immune checkpointinhibition in the treatment of ovarian cancer. Gynecologic Oncology Research and Practice,3(1):11, 12 2016.[51] Thomas F Gajewski, Hans Schreiber, and Yang-Xin Fu. Innate and adaptive immune cellsin the tumor microenvironment. Nature Immunology, 14(10):1014–1022, 9 2013.[52] Nicholas McGranahan, Rachel Rosenthal, Crispin T. Hiley, Andrew J. Rowan, Thomas B.K.Watkins, Gareth A. Wilson, Nicolai J. Birkbak, Selvaraju Veeriah, Peter Van Loo, JavierHerrero, Charles Swanton, Charles Swanton, Mariam Jamal-Hanjani, Selvaraju Veeriah,Seema Shafi, Justyna Czyzewska-Khan, Diana Johnson, Joanne Laycock, Leticia Bosshard-Carter, Rachel Rosenthal, Pat Gorman, Robert E. Hynds, Gareth Wilson, Nicolai J.Birkbak, Thomas B.K. Watkins, Nicholas McGranahan, Stuart Horswell, Richard Mitter,Mickael Escudero, Aengus Stewart, Peter Van Loo, Andrew Rowan, Hang Xu, SamraTurajlic, Crispin Hiley, Christopher Abbosh, Jacki Goldman, Richard Kevin Stone, TamaraDenner, Nik Matthews, Greg Elgar, Sophia Ward, Marta Costa, Sharmin Begum, BenPhillimore, Tim Chambers, Emma Nye, Sofia Graca, Maise Al Bakir, Kroopa Joshi, An-drew Furness, Assma Ben Aissa, Yien Ning Sophia Wong, Andy Georgiou, Sergio Quezada,John A. Hartley, Helen L. Lowe, Javier Herrero, David Lawrence, Martin Hayward, Niko-laos Panagiotopoulos, Shyam Kolvekar, Mary Falzon, Elaine Borg, Teresa Marafioti, CeliaSimeon, Gemma Hector, Amy Smith, Marie Aranda, Marco Novelli, Dahmane Oukrif,Sam M. Janes, Ricky Thakrar, Martin Forster, Tanya Ahmad, Siow Ming Lee, DionysisPapadatos-Pastos, Dawn Carnell, Ruheena Mendes, Jeremy George, Neal Navani, AsiaAhmed, Magali Taylor, Junaid Choudhary, Yvonne Summers, Raffaele Califano, Paul Tay-lor, Rajesh Shah, Piotr Krysiak, Kendadai Rammohan, Eustace Fontaine, Richard Booton,Matthew Evison, Phil Crosbie, Stuart Moss, Faiza Idries, Leena Joseph, Paul Bishop, An-shuman Chaturved, Anne Marie Quinn, Helen Doran, Angela Leek, Phil Harrison, Katrina175Moore, Rachael Waddington, Juliette Novasio, Fiona Blackhall, Jane Rogan, Elaine Smith,Caroline Dive, Jonathan Tugwood, Ged Brady, Dominic G. Rothwell, Francesca Chemi,Jackie Pierce, Sakshi Gulati, Babu Naidu, Gerald Langman, Simon Trotter, Mary Bellamy,Hollie Bancroft, Amy Kerr, Salma Kadiri, Joanne Webb, Gary Middleton, Madava Djeara-man, Dean Fennell, Jacqui A. Shaw, John Le Quesne, David Moore, Apostolos Nakas,Sridhar Rathinam, William Monteiro, Hilary Marshall, Louise Nelson, Jonathan Bennett,Joan Riley, Lindsay Primrose, Luke Martinson, Girija Anand, Sajid Khan, Anita Amadi,Marianne Nicolson, Keith Kerr, Shirley Palmer, Hardy Remmen, Joy Miller, Keith Buchan,Mahendran Chetty, Lesley Gomersall, Jason Lester, Alison Edwards, Fiona Morgan, HaydnAdams, Helen Davies, Malgorzata Kornaszewska, Richard Attanoos, Sara Lock, AzminaVerjee, Mairead MacKenzie, Maggie Wilcox, Harriet Bell, Allan Hackshaw, Yenting Ngai,Sean Smith, Nicole Gower, Christian Ottensmeier, Serena Chee, Benjamin Johnson, AimanAlzetani, Emily Shaw, Eric Lim, Paulo De Sousa, Monica Tavares Barbosa, Alex Bowman,Simon Jordan, Alexandra Rice, Hilgardt Raubenheimer, Chiara Proli, Maria Elena Cufari,John Carlo Ronquillo, Angela Kwayie, Harshil Bhayani, Morag Hamilton, Yusura Bakar,Natalie Mensah, Lyn Ambrose, Anand Devaraj, Silviu Buderi, Jonathan Finch, LeireAzcarate, Hema Chavan, Sophie Green, Hillaria Mashinga, Andrew G. Nicholson, KelvinLau, Michael Sheaff, Peter Schmid, John Conibear, Veni Ezhil, Babikir Ismail, MelanieIrvin-sellers, Vineet Prakash, Peter Russell, Teresa Light, Tracey Horey, Sarah Danson,Jonathan Bury, John Edwards, Jennifer Hill, Sue Matthews, Yota Kitsanta, Kim Suvarna,Patricia Fisher, Allah Dino Keerio, Michael Shackcloth, John Gosney, Pieter Postmus,Sarah Feeney, Julius Asante-Siaw, Hugo J.W.L. Aerts, Stefan Dentro, and ChristopheDessimoz. Allele-Specific HLA Loss and Immune Escape in Lung Cancer Evolution. Cell,171(6):1259–1271, 11 2017.[53] Yulei Chen and Xiaobo Zhang. Pivotal regulators of tissue homeostasis and cancer:macrophages. Experimental Hematology & Oncology, 6(1):23, 12 2017.[54] Florian Finkernagel, Silke Reinartz, Sonja Lieber, Till Adhikary, Annika Wortmann,Nathalie Hoffmann, Tim Bieringer, Andrea Nist, Thorsten Stiewe, Julia M Jansen, UweWagner, Sabine Mu¨ller-Bru¨sselbach, and Rolf Mu¨ller. The transcriptional signature of hu-man ovarian carcinoma macrophages is associated with extracellular matrix reorganization.Oncotarget, 7(46):75339–75352, 11 2016.[55] Till Adhikary, Annika Wortmann, Florian Finkernagel, Sonja Lieber, Andrea Nist, ThorstenStiewe, Uwe Wagner, Sabine Mu¨ller-Bru¨sselbach, Silke Reinartz, and Rolf Mu¨ller. Interferon176signaling in ascites-associated macrophages is linked to a favorable clinical outcome in asubgroup of ovarian carcinoma patients. BMC Genomics, 18(1):243, 12 2017.[56] Samuel F. Bakhoum, Bryan Ngo, Ashley M. Laughney, Julie-Ann Cavallo, Charles J.Murphy, Peter Ly, Pragya Shah, Roshan K. Sriram, Thomas B. K. Watkins, Neil K. Taunk,Mercedes Duran, Chantal Pauli, Christine Shaw, Kalyani Chadalavada, Vinagolu K.Rajasekhar, Giulio Genovese, Subramanian Venkatesan, Nicolai J. Birkbak, NicholasMcGranahan, Mark Lundquist, Quincey LaPlant, John H. Healey, Olivier Elemento,Christine H. Chung, Nancy Y. Lee, Marcin Imielenski, Gouri Nanjangud, Dana Peer,Don W. Cleveland, Simon N. Powell, Jan Lammerding, Charles Swanton, and Lewis C.Cantley. Chromosomal instability drives metastasis through a cytosolic DNA response.Nature, 553(7689):467–472, 1 2018.[57] Seng-Ryong Woo, Leticia Corrales, and Thomas F. Gajewski. Innate Immune Recognitionof Cancer. Annual Review of Immunology, 33(1):445–474, 3 2015.[58] Fei Xing, Jamila Saidou, and Kounosuke Watabe. Cancer associated fibroblasts (CAFs) intumor microenvironment. Frontiers in bioscience (Landmark edition), 15:166–79, 1 2010.[59] Weimin Wang, Ilona Kryczek, Lubomr Dosta´l, Heng Lin, Lijun Tan, Lili Zhao, Fujia Lu,Shuang Wei, Tomasz Maj, Dongjun Peng, Gong He, Linda Vatan, Wojciech Szeliga, RorkKuick, Jan Kotarski, Rafa Tarkowski, Yali Dou, Ramandeep Rattan, Adnan Munkarah,J Rebecca Liu, and Weiping Zou. Effector T Cells Abrogate Stroma-Mediated Chemore-sistance in Ovarian Cancer. Cell, 165(5):1092–105, 5 2016.[60] Naoyo Nishida, Hirohisa Yano, Takashi Nishida, Toshiharu Kamura, and Masamichi Kojiro.Angiogenesis in cancer. Vascular health and risk management, 2(3):213–9, 2006.[61] Andrew C Dudley. Tumor endothelial cells. Cold Spring Harbor perspectives in medicine,2(3):a006536, 3 2012.[62] Stuart C. Williamson, Robert L. Metcalf, Francesca Trapani, Sumitra Mohan, JennyAntonello, Benjamin Abbott, Hui Sun Leong, Christopher P. E. Chester, Nicole Simms,Radoslaw Polanski, Daisuke Nonaka, Lynsey Priest, Alberto Fusi, Fredrika Carlsson,Anders Carlsson, Mary J. C. Hendrix, Richard E. B. Seftor, Elisabeth A. Seftor, Dominic G.Rothwell, Andrew Hughes, James Hicks, Crispin Miller, Peter Kuhn, Ged Brady, Kathryn L.Simpson, Fiona H. Blackhall, and Caroline Dive. Vasculogenic mimicry in small cell lungcancer. Nature Communications, 7:13322, 11 2016.177[63] K L Eales, K E R Hollinshead, and D A Tennant. Hypoxia and metabolic adaptation ofcancer cells. Oncogenesis, 5(1):e190, 1 2016.[64] Ying Zhang and Hildegund C. J. Ertl. Starved and Asphyxiated: How Can CD8+ T Cellswithin a Tumor Microenvironment Prevent Tumor Progression. Frontiers in Immunology,7:32, 2 2016.[65] Dong Chul Lee, Hyun Ahm Sohn, Zee-Yong Park, Sangho Oh, Yun Kyung Kang, Kyoung-Min Lee, Minho Kang, Ye Jin Jang, Suk-Jin Yang, Young Ki Hong, Hanmi Noh, Jung-AeKim, Dong Joon Kim, Kwang-Hee Bae, Dong Min Kim, Sang J Chung, Hyang Sook Yoo,Dae-Yeul Yu, Kyung Chan Park, and Young Il Yeom. A lactate-induced response tohypoxia. Cell, 161(3):595–609, 4 2015.[66] Emma Williams, Stewart Martin, Robert Moss, Lindy Durrant, and Suha Deen. Co-expression of VEGF and CA9 in ovarian high-grade serous carcinoma and relationship tosurvival. Virchows Archiv, 461(1):33–39, 7 2012.[67] James M. Heather and Benjamin Chain. The sequence of sequencers: The history ofsequencing DNA. Genomics, 107(1):1–8, 1 2016.[68] Yong Wang and Nicholas E Navin. Advances and applications of single-cell sequencingtechnologies. Molecular cell, 58(4):598–609, 5 2015.[69] Noemi Andor, Trevor A Graham, Marnix Jansen, Li C Xia, C Athena Aktipis, ClaudiaPetritsch, Hanlee P Ji, and Carlo C Maley. Pan-cancer analysis of the extent andconsequences of intratumor heterogeneity. Nature Medicine, 22(1):105–113, 11 2015.[70] Luc G T Morris, Nadeem Riaz, Alexis Desrichard, Yasin S¸enbabaog˘lu, A Ari Hakimi,Vladimir Makarov, Jorge S Reis-Filho, and Timothy A Chan. Pan-cancer analysis ofintratumor heterogeneity as a prognostic determinant of survival. Oncotarget, 7(9):10051–63, 3 2016.[71] Andrew Roth, Jaswinder Khattra, Damian Yap, Adrian Wan, Emma Laks, Justina Biele,Gavin Ha, Samuel Aparicio, Alexandre Bouchard-Coˆte´, and Sohrab P Shah. PyClone:statistical inference of clonal population structure in cancer. Nature Methods, 11(4):396–398,3 2014.[72] Samuel Aparicio and Carlos Caldas. The Implications of Clonal Genome Evolution forCancer Medicine. New England Journal of Medicine, 368(9):842–851, 2 2013.178[73] Christopher A Miller, Brian S White, Nathan D Dees, Malachi Griffith, John S Welch,Obi L Griffith, Ravi Vij, Michael H Tomasson, Timothy A Graubert, Matthew J Walter,Matthew J Ellis, William Schierding, John F DiPersio, Timothy J Ley, Elaine R Mardis,Richard K Wilson, and Li Ding. SciClone: inferring clonal architecture and trackingthe spatial and temporal patterns of tumor evolution. PLoS computational biology,10(8):e1003665, 8 2014.[74] Salem Malikic, Andrew W. McPherson, Nilgun Donmez, and Cenk S. Sahinalp. Clonalityinference in multiple tumor samples using phylogeny. Bioinformatics, 31(9):1349–1356, 52015.[75] Victoria Popic, Raheleh Salari, Iman Hajirasouliha, Dorna Kashef-Haghighi, Robert BWest, and Serafim Batzoglou. Fast and scalable inference of multi-sample cancer lineages.Genome Biology, 16(1):91, 12 2015.[76] Amit G Deshwar, Shankar Vembu, Christina K Yung, Gun Jang, Lincoln Stein, andQuaid Morris. PhyloWGS: Reconstructing subclonal composition and evolution fromwhole-genome sequencing of tumors. Genome Biology, 16(1):35, 2 2015.[77] Sohrab Salehi, Adi Steif, Andrew Roth, Samuel Aparicio, Alexandre Bouchard-Coˆte´, andSohrab P. Shah. ddClone: joint statistical inference of clonal populations from single celland bulk tumour sequencing data. Genome Biology, 18(1):44, 12 2017.[78] E L Reinherz and S F Schlossman. The differentiation and function of human T lympho-cytes. Cell, 19(4):821–7, 4 1980.[79] Kathrin Pieper, Bodo Grimbacher, and Hermann Eibel. B-cell biology and development.Journal of Allergy and Clinical Immunology, 131(4):959–971, 4 2013.[80] Craig H Bassing, Wojciech Swat, and Frederick W Alt. The Mechanism and Regulationof Chromosomal V(D)J Recombination. Cell, 109(2):S45–S55, 4 2002.[81] Heinz Jacobs and Linda Bross. Towards an understanding of somatic hypermutation.Current Opinion in Immunology, 13(2):208–218, 4 2001.[82] Daniel J Laydon, Charles R M Bangham, and Becca Asquith. Estimating T-cell repertoirediversity: limitations of classical estimators and a new approach. Philosophical transactionsof the Royal Society of London. Series B, Biological sciences, 370(1675), 8 2015.179[83] Ren L Warren, J Douglas Freeman, Thomas Zeng, Gina Choe, Sarah Munro, RichardMoore, John R Webb, and Robert A Holt. Exhaustive T-cell repertoire sequencing ofhuman peripheral blood samples reveals signatures of antigen selection and a directlymeasured repertoire size of at least 1 million clonotypes. Genome research, 21(5):790–7, 52011.[84] Daniel J Woodsworth, Mauro Castellarin, and Robert A Holt. Sequence analysis of T-cellrepertoires in health and disease. Genome medicine, 5(10):98, 2013.[85] Elisa Rosati, C Marie Dowds, Evaggelia Liaskou, Eva Kristine Klemsdal Henriksen, Tom HKarlsen, and Andre Franke. Overview of methodologies for T-cell receptor repertoireanalysis. BMC biotechnology, 17(1):61, 7 2017.[86] Dmitriy A Bolotin, Stanislav Poslavsky, Igor Mitrophanov, Mikhail Shugay, Ilgar ZMamedov, Ekaterina V Putintseva, and Dmitriy M Chudakov. MiXCR: software forcomprehensive adaptive immunity profiling. Nature Methods, 12(5):380–381, 4 2015.[87] Yaxuan Yu, Rhodri Ceredig, and Cathal Seoighe. LymAnalyzer: a tool for comprehensiveanalysis of next generation sequencing data of T cell receptors and immunoglobulins.Nucleic Acids Research, 44(4):e31–e31, 2 2016.[88] Leon Kuchenbecker, Mikalai Nienen, Jochen Hecht, Avidan U. Neumann, Nina Babel, KnutReinert, and Peter N. Robinson. IMSEQa fast and error aware approach to immunogeneticsequence analysis. Bioinformatics, 31(18):2963–2971, 9 2015.[89] Duncan K. Ralph and Frederick A. Matsen. Likelihood-Based Inference of B Cell ClonalFamilies. PLOS Computational Biology, 12(10):e1005086, 10 2016.[90] Duncan K. Ralph, Frederick A. Matsen, BJ Huntly, R Rance, GS Vassiliou, and GA Follows.Consistency of VDJ Rearrangement and Substitution Parameters Enables Accurate B CellReceptor Sequence Annotation. PLOS Computational Biology, 12(1):e1004409, 1 2016.[91] Bryan Howie, Anna M Sherwood, Ashley D Berkebile, Jan Berka, Ryan O Emerson,David W Williamson, Ilan Kirsch, Marissa Vignali, Mark J Rieder, Christopher S Carlson,and Harlan S Robins. High-throughput pairing of T cell receptor α and β sequences.Science translational medicine, 7(301):301ra131, 8 2015.[92] Andrew Roth, Andrew McPherson, Emma Laks, Justina Biele, Damian Yap, AdrianWan, Maia A Smith, Cydney B Nielsen, Jessica N McAlpine, Samuel Aparicio, Alexandre180Bouchard-Coˆte´, and Sohrab P Shah. Clonal genotype and population structure inferencefrom single-cell tumor sequencing. Nature methods, 13(7):573–6, 7 2016.[93] Anna K. Casasent, Aislyn Schalck, Ruli Gao, Emi Sei, Annalyssa Long, William Pang-burn, Tod Casasent, Funda Meric-Bernstam, Mary E. Edgerton, and Nicholas E. Navin.Multiclonal Invasion in Breast Tumors Identified by Topographic Single Cell Sequencing.Cell, 172(1-2):205–217, 1 2018.[94] Charissa Kim, Ruli Gao, Emi Sei, Rachel Brandt, Johan Hartman, Thomas Hatschek,Nicola Crosetto, Theodoros Foukakis, and Nicholas E. Navin. Chemoresistance Evolution inTriple-Negative Breast Cancer Delineated by Single-Cell Sequencing. Cell, 173(4):879–893,5 2018.[95] Charles Gawad, Winston Koh, and Stephen R. Quake. Single-cell genome sequencing:current state of the science. Nature Reviews Genetics, 17(3):175–188, 3 2016.[96] Konstantin A. Blagodatskikh, Vladimir M. Kramarov, Ekaterina V. Barsova, Alexey V.Garkovenko, Dmitriy S. Shcherbo, Andrew A. Shelenkov, Vera V. Ustinova, Maria R.Tokarenko, Simon C. Baker, Tatiana V. Kramarova, and Konstantin B. Ignatov. ImprovedDOP-PCR (iDOP-PCR): A robust and simple WGA method for efficient amplification oflow copy number genomic DNA. PLOS ONE, 12(9):e0184507, 9 2017.[97] Hans Zahn, Adi Steif, Emma Laks, Peter Eirew, Michael VanInsberghe, Sohrab P Shah,Samuel Aparicio, and Carl L Hansen. Scalable whole-genome single-cell library preparationwithout preamplification. Nature Methods, 14(2):167–173, 2 2017.[98] Emma Laks, Hans Zahn, Daniel Lai, Andrew McPherson, Adi Steif, Jazmine Brimhall,Justina Biele, Beixi Wang, Tehmina Masud, Diljot Grewal, Cydney Nielsen, SamanthaLeung, Viktoria Bojilova, Maia Smith, Oleg Golovko, Steven Poon, Peter Eirew, FarhiaKabeer, Teresa Ruiz de Algara, So Ra Lee, M. Jafar Taghiyar, Curtis Huebner, Jessica Ngo,Tim Chan, Spencer Vatrt-Watts, Pascale Walters, Nafis Abrar, Sophia Chan, Matt Wiens,Lauren Martin, R. Wilder Scott, Michael T. Underhill, Elizabeth Chavez, Christian Steidl,Daniel Da Costa, Yusanne Ma, Robin J. N. Coope, Richard Corbett, Stephen Pleasance,Richard Moore, Andy J. Mungall, CRUK IMAXT Consortium, Marco A. Marra, CarlHansen, Sohrab Shah, and Samuel Aparicio. Resource: Scalable whole genome sequencingof 40,000 single cells identifies stochastic aneuploidies, genome replication states and clonalrepertoires. bioRxiv, page 411058, 9 2018.[99] Single Cell CNV - 10x Genomics.181[100] Zhong Wang, Mark Gerstein, and Michael Snyder. RNA-Seq: a revolutionary tool fortranscriptomics. Nature Reviews Genetics, 10(1):57–63, 1 2009.[101] R. W. Tothill, A. V. Tinker, J. George, R. Brown, S. B. Fox, S. Lade, D. S. Johnson,M. K. Trivett, D. Etemadmoghadam, B. Locandro, N. Traficante, S. Fereday, J. A. Hung,Y.-E. Chiew, I. Haviv, D. Gertig, A. deFazio, D. D.L. Bowtell, and David D L Bowtell.Novel Molecular Subtypes of Serous and Endometrioid Ovarian Cancer Linked to ClinicalOutcome. Clinical Cancer Research, 14(16):5198–5208, 8 2008.[102] Daniel C. Koboldt, Robert S. Fulton, Michael D. McLellan, Heather Schmidt, JoelleKalicki-Veizer, Joshua F. McMichael, Lucinda L. Fulton, David J. Dooling, Li Ding,Elaine R. Mardis, Richard K. Wilson, Adrian Ally, Miruna Balasundaram, Yaron S. N.Butterfield, Rebecca Carlsen, Candace Carter, Andy Chu, Eric Chuah, Hye-Jung E.Chun, Robin J. N. Coope, Noreen Dhalla, Ranabir Guin, Carrie Hirst, Martin Hirst,Robert A. Holt, Darlene Lee, Haiyan I. Li, Michael Mayo, Richard A. Moore, Andrew J.Mungall, Erin Pleasance, A. Gordon Robertson, Jacqueline E. Schein, Arash Shafiei, PayalSipahimalani, Jared R. Slobodan, Dominik Stoll, Angela Tam, Nina Thiessen, Richard J.Varhol, Natasja Wye, Thomas Zeng, Yongjun Zhao, Inanc Birol, Steven J. M. Jones,Marco A. Marra, Andrew D. Cherniack, Gordon Saksena, Robert C. Onofrio, Nam H. Pho,Scott L. Carter, Steven E. Schumacher, Barbara Tabak, Bryan Hernandez, Jeff Gentry,Huy Nguyen, Andrew Crenshaw, Kristin Ardlie, Rameen Beroukhim, Wendy Winckler,Gad Getz, Stacey B. Gabriel, Matthew Meyerson, Lynda Chin, Peter J. Park, RajuKucherlapati, Katherine A. Hoadley, J. Todd Auman, Cheng Fan, Yidi J. Turman, YanShi, Ling Li, Michael D. Topal, Xiaping He, Hann-Hsiang Chao, Aleix Prat, Grace O. Silva,Michael D. Iglesia, Wei Zhao, Jerry Usary, Jonathan S. Berg, Michael Adams, JessicaBooker, Junyuan Wu, Anisha Gulabani, Tom Bodenheimer, Alan P. Hoyle, Janae V.Simons, Matthew G. Soloway, Lisle E. Mose, Stuart R. Jefferys, Saianand Balu, Joel S.Parker, D. Neil Hayes, Charles M. Perou, Simeen Malik, Swapna Mahurkar, Hui Shen,Daniel J. Weisenberger, Timothy Triche Jr, Phillip H. Lai, Moiz S. Bootwalla, Dennis T.Maglinte, Benjamin P. Berman, David J. Van Den Berg, Stephen B. Baylin, Peter W.Laird, Chad J. Creighton, Lawrence A. Donehower, Gad Getz, Michael Noble, Doug Voet,Gordon Saksena, Nils Gehlenborg, Daniel DiCara, Juinhua Zhang, Hailei Zhang, Chang-Jiun Wu, Spring Yingchun Liu, Michael S. Lawrence, Lihua Zou, Andrey Sivachenko, PeiLin, Petar Stojanov, Rui Jing, Juok Cho, Raktim Sinha, Richard W. Park, Marc-DanieNazaire, Jim Robinson, Helga Thorvaldsdottir, Jill Mesirov, Peter J. Park, Lynda Chin,Sheila Reynolds, Richard B. Kreisberg, Brady Bernard, Ryan Bressler, Timo Erkkila, Jake182Lin, Vesteinn Thorsson, Wei Zhang, Ilya Shmulevich, Giovanni Ciriello, Nils Weinhold,Nikolaus Schultz, Jianjiong Gao, Ethan Cerami, Benjamin Gross, Anders Jacobsen, RileenSinha, B. Arman Aksoy, Yevgeniy Antipin, Boris Reva, Ronglai Shen, Barry S. Taylor,Marc Ladanyi, Chris Sander, Pavana Anur, Paul T. Spellman, Yiling Lu, Wenbin Liu, RoelR. G. Verhaak, Gordon B. Mills, Rehan Akbani, Nianxiang Zhang, Bradley M. Broom,Tod D. Casasent, Chris Wakefield, Anna K. Unruh, Keith Baggerly, Kevin Coombes,John N. Weinstein, David Haussler, Christopher C. Benz, Joshua M. Stuart, Stephen C.Benz, Jingchun Zhu, Christopher C. Szeto, Gary K. Scott, Christina Yau, Evan O. Paull,Daniel Carlin, Christopher Wong, Artem Sokolov, Janita Thusberg, Sean Mooney, Sam Ng,Theodore C. Goldstein, Kyle Ellrott, Mia Grifford, Christopher Wilks, Singer Ma, BrianCraft, Chunhua Yan, Ying Hu, Daoud Meerzaman, Julie M. Gastier-Foster, Jay Bowen,Nilsa C. Ramirez, Aaron D. Black, Robert E. XPATH ERROR: unknown variable ”tname”.,Peter White, Erik J. Zmuda, Jessica Frick, Tara M. Lichtenberg, Robin Brookens, Myra M.George, Mark A. Gerken, Hollie A. Harper, Kristen M. Leraas, Lisa J. Wise, Teresa R.Tabler, Cynthia McAllister, Thomas Barr, Melissa Hart-Kothari, Katie Tarvin, CharlesSaller, George Sandusky, Colleen Mitchell, Mary V. Iacocca, Jennifer Brown, BrendaRabeno, Christine Czerwinski, Nicholas Petrelli, Oleg Dolzhansky, Mikhail Abramov, OlgaVoronina, Olga Potapova, Jeffrey R. Marks, Wiktoria M. Suchorska, Dawid Murawa,Witold Kycler, Matthew Ibbs, Konstanty Korski, Arkadiusz Spycha la, Pawe Murawa,Jacek J. Brzezin´ski, Hanna Perz, Radosaw  Laz´niak, Marek Teresiak, Honorata Tatka,Ewa Leporowska, Marta Bogusz-Czerniewicz, Julian Malicki, Andrzej Mackiewicz, MaciejWiznerowicz, Xuan Van Le, Bernard Kohl, Nguyen Viet Tien, Richard Thorp, NguyenVan Bang, Howard Sussman, Bui Duc Phu, Richard Hajek, Nguyen Phi Hung, Tran VietThe Phuong, Huynh Quyet Thang, Khurram Zaki Khan, Robert Penny, David Mallery,Erin Curley, Candace Shelton, Peggy Yena, James N. Ingle, Fergus J. Couch, Wilma L.Lingle, Tari A. King, Ana Maria Gonzalez-Angulo, Gordon B. Mills, Mary D. Dyer,Shuying Liu, Xiaolong Meng, Modesto Patangan, Frederic Waldman, Hubert Sto¨ppler,W. Kimryn Rathmell, Leigh Thorne, Mei Huang, Lori Boice, Ashley Hill, Carl Morrison,Carmelo Gaudioso, Wiam Bshara, Kelly Daily, Sophie C. Egea, Mark D. Pegram, CarmenGomez-Fernandez, Rajiv Dhir, Rohit Bhargava, Adam Brufsky, Craig D. Shriver, Jeffrey A.Hooke, Jamie Leigh Campbell, Richard J. Mural, Hai Hu, Stella Somiari, Caroline Larson,Brenda Deyarmin, Leonid Kvecher, Albert J. Kovatich, Matthew J. Ellis, Tari A. King,Hai Hu, Fergus J. Couch, Richard J. Mural, Thomas Stricker, Kevin White, OlufunmilayoOlopade, James N. Ingle, Chunqing Luo, Yaqin Chen, Jeffrey R. Marks, Frederic Waldman,Maciej Wiznerowicz, Ron Bose, Li-Wei Chang, Andrew H. Beck, Ana Maria Gonzalez-183Angulo, Todd Pihl, Mark Jensen, Robert Sfeir, Ari Kahn, Anna Chu, Prachi Kothiyal,Zhining Wang, Eric Snyder, Joan Pontius, Brenda Ayala, Mark Backus, Jessica Walton,Julien Baboud, Dominique Berton, Matthew Nicholls, Deepak Srinivasan, Rohini Raman,Stanley Girshik, Peter Kigonya, Shelley Alonso, Rashmi Sanbhadti, Sean Barletta, DavidPot, Margi Sheth, John A. Demchok, Kenna R. Mills Shaw, Liming Yang, Greg Eley,Martin L. Ferguson, Roy W. Tarnuzzer, Jiashan Zhang, Laura A. L. Dillon, KennethBuetow, Peter Fielding, Bradley A. Ozenberger, Mark S. Guyer, Heidi J. Sofia, andJacqueline D. Palchik. Comprehensive molecular portraits of human breast tumours.Nature, 490(7418):61–70, 9 2012.[103] The Cancer Genome Atlas Network. Comprehensive molecular characterization of humancolon and rectal cancer. Nature, 487(7407):330–337, 7 2012.[104] Aaron M Newman, Chih Long Liu, Michael R Green, Andrew J Gentles, Weiguo Feng,Yue Xu, Chuong D Hoang, Maximilian Diehn, and Ash A Alizadeh. Robust enumerationof cell subsets from tissue expression profiles. Nature Methods, 12(5):453–457, 3 2015.[105] Aleksandra A Kolodziejczyk, Jong Kyoung Kim, Valentine Svensson, John C Marioni, andSarah A Teichmann. The technology and biology of single-cell RNA sequencing. Molecularcell, 58(4):610–20, 5 2015.[106] D P Bartel, M Sheng, L F Lau, and M E Greenberg. Growth factors and membranedepolarization activate distinct programs of early response gene expression: dissociationof fos and jun induction. Genes & development, 3(3):304–13, 3 1989.[107] Benjamin Lacar, Sara B Linker, Baptiste N Jaeger, Suguna R Krishnaswami, Jerika JBarron, Martijn J E Kelder, Sarah L Parylak, Apu C M Paquola, Pratap Venepally, MarkNovotny, Carolyn O’Connor, Conor Fitzpatrick, Jennifer A Erwin, Jonathan Y Hsu, DavidHusband, Michael J McConnell, Roger Lasken, and Fred H Gage. Nuclear RNA-seq ofsingle neurons reveals molecular signatures of activation. Nature communications, 7:11022,2016.[108] Alexander B Rosenberg, Charles M Roco, Richard A Muscat, Anna Kuchina, PaulSample, Zizhen Yao, Lucas T Graybuck, David J Peeler, Sumit Mukherjee, Wei Chen,Suzie H Pun, Drew L Sellers, Bosiljka Tasic, and Georg Seelig. Single-cell profiling of thedeveloping mouse brain and spinal cord with split-pool barcoding. Science (New York,N.Y.), 360(6385):176–182, 3 2018.184[109] Simone Picelli, Omid R Faridani, sa K Bjo¨rklund, Gsta Winberg, Sven Sagasser, andRickard Sandberg. Full-length RNA-seq from single cells using Smart-seq2. NatureProtocols, 9(1):171–181, 1 2014.[110] Grace X. Y. Zheng, Jessica M. Terry, Phillip Belgrader, Paul Ryvkin, Zachary W. Bent,Ryan Wilson, Solongo B. Ziraldo, Tobias D. Wheeler, Geoff P. McDermott, Junjie Zhu,Mark T. Gregory, Joe Shuga, Luz Montesclaros, Jason G. Underwood, Donald A. Masque-lier, Stefanie Y. Nishimura, Michael Schnall-Levin, Paul W. Wyatt, Christopher M.Hindson, Rajiv Bharadwaj, Alexander Wong, Kevin D. Ness, Lan W. Beppu, H. JoachimDeeg, Christopher McFarland, Keith R. Loeb, William J. Valente, Nolan G. Ericson,Emily A. Stevens, Jerald P. Radich, Tarjei S. Mikkelsen, Benjamin J. Hindson, andJason H. Bielas. Massively parallel digital transcriptional profiling of single cells. NatureCommunications, 8:14049, 1 2017.[111] Po-Yuan Tung, John D. Blischak, Chiaowen Joyce Hsiao, David A. Knowles, Jonathan E.Burnett, Jonathan K. Pritchard, and Yoav Gilad. Batch effects and the effective design ofsingle-cell gene expression studies. Scientific Reports, 7(1):39921, 12 2017.[112] Tomislav Ilicic, Jong Kyoung Kim, Aleksandra A Kolodziejczyk, Frederik Otzen Bagger,Davis James McCarthy, John C Marioni, and Sarah A Teichmann. Classification of lowquality cells from single-cell RNA-seq data. Genome biology, 17:29, 2 2016.[113] Marlon Stoeckius, Shiwei Zheng, Brian Houck-Loomis, Stephanie Hao, Bertrand Yeung,Peter Smibert, and Rahul Satija. Cell &quot;hashing&quot; with barcoded antibodiesenables multiplexing and doublet detection for single cell genomics. bioRxiv, page 237693,12 2017.[114] Matthew D Young and Sam Behjati. SoupX removes ambient RNA contamination fromdroplet based single cell RNA sequencing data. bioRxiv, page 303727, 4 2018.[115] Davide Risso, Fanny Perraudeau, Svetlana Gribkova, Sandrine Dudoit, and Jean-PhilippeVert. A general and flexible method for signal extraction from single-cell RNA-seq data.Nature Communications, 9(1):284, 12 2018.[116] David van Dijk, Roshan Sharma, Juozas Nainys, Kristina Yim, Pooja Kathail, Ambrose JCarr, Cassandra Burdziak, Kevin R Moon, Christine L Chaffer, Diwakar Pattabiraman,Brian Bierie, Linas Mazutis, Guy Wolf, Smita Krishnaswamy, and Dana Pe’er. RecoveringGene Interactions from Single-Cell Data Using Data Diffusion. Cell, 174(3):716–729, 72018.185[117] M. E. Ritchie, B. Phipson, D. Wu, Y. Hu, C. W. Law, W. Shi, and G. K. Smyth. limmapowers differential expression analyses for RNA-sequencing and microarray studies. NucleicAcids Research, 43(7):e47–e47, 4 2015.[118] Laleh Haghverdi, Aaron T L Lun, Michael D Morgan, and John C Marioni. Batch effectsin single-cell RNA-sequencing data are corrected by matching mutual nearest neighbors.Nature Biotechnology, 36(5):421–427, 4 2018.[119] Brian L Hie, Bryan Bryson, and Bonnie Berger. Panoramic stitching of heterogeneoussingle-cell transcriptomic data. bioRxiv, page 371179, 7 2018.[120] Jong-Eun Park, Krzysztof Polanski, Kerstin Meyer, and Sarah A Teichmann. Fast BatchAlignment of Single Cell Transcriptomes Unifies Multiple Mouse Cell Atlases into anIntegrated Landscape. bioRxiv, page 397042, 8 2018.[121] Evan Z Macosko, Anindita Basu, Rahul Satija, James Nemesh, Karthik Shekhar, MelissaGoldman, Itay Tirosh, Allison R Bialas, Nolan Kamitaki, Emily M Martersteck, John JTrombetta, David A Weitz, Joshua R Sanes, Alex K Shalek, Aviv Regev, Steven AMcCarroll, A.D. Amir, K.L. Davis, M.D. Tadmor, E.F. Simonds, J.H. Levine, S.C. Ben-dall, D.K. Shenfeld, S. Krishnaswamy, G.P. Nolan, D. Peer, N.R. Beer, E.K. Wheeler,L. Lee-Houghton, N. Watkins, S. Nasarabadi, N. Hebert, P. Leung, D.W. Arnold, C.G.Bailey, B.W. Colston, G.J. Berman, D.M. Choi, W. Bialek, J.W. Shaevitz, P. Brennecke,S. Anders, J.K. Kim, A.A. Ko lodziejczyk, X. Zhang, V. Proserpio, B. Baying, V. Benes,S.A. Teichmann, J.C. Marioni, M.G. Heisler, R.J. Britten, D.E. Kohne, N.C. Chung, J.D.Storey, F.J. Descamps, E. Martens, P. Proost, S. Starckx, P.E. Van den Steen, J. VanDamme, G. Opdenakker, M. Ester, H.P. Kriegel, J. Sander, X. Xu, E.V. Famiglietti, S.J.Sundquist, A. Feigenspan, B. Teubner, K. Willecke, R. Weiler, T. Hashimshony, F. Wagner,N. Sher, I. Yanai, S. Haverkamp, H. Wa¨ssle, B.J. Hindson, K.D. Ness, D.A. Masquelier,P. Belgrader, N.J. Heredia, A.J. Makarewicz, I.J. Bright, M.Y. Lucero, A.L. Hiddessen,T.C. Legler, et al., S. Islam, A. Zeisel, S. Joost, G. La Manno, P. Zajac, M. Kasper,P. Lo¨nnerberg, S. Linnarsson, D.A. Jaitin, E. Kenigsberg, H. Keren-Shaul, N. Elefant,F. Paul, I. Zaretsky, A. Mildner, N. Cohen, S. Jung, A. Tanay, I. Amit, C.J. Jeon,E. Strettoi, R.H. Masland, J.N. Kay, P.E. Voinescu, M.W. Chu, J.R. Sanes, T. Kivioja,A. Va¨ha¨rautio, K. Karlsson, M. Bonke, M. Enge, S. Linnarsson, J. Taipale, A.M. Klein,L. Mazutis, I. Akartuna, N. Tallapragada, A. Veres, V. Li, L. Peshkin, D.A. Weitz, M.W.Kirschner, L. Luo, E.M. Callaway, K. Svoboda, R.H. Masland, A. McDavid, G. Finak,P.K. Chattopadyay, M. Dominguez, L. Lamoreaux, S.S. Ma, M. Roederer, R. Gottardo,186G.A. Ascoli, L. Alonso-Nanclares, S.A. Anderson, G. Barrionuevo, R. Benavides-Piccione,A. Burkhalter, G. Buzsa´ki, B. Cauli, J. Defelipe, A. Faire´n, Petilla Interneuron Nomen-clature Group, et al., S. Picelli, A.K. Bjo¨rklund, O.R. Faridani, S. Sagasser, G. Winberg,R. Sandberg, J.R. Sanes, R.H. Masland, J.R. Sanes, S.L. Zipursky, R. Satija, J.A. Farrell,D. Gennert, A.F. Schier, A. Regev, A.K. Shalek, R. Satija, X. Adiconis, R.S. Gertner, J.T.Gaublomme, R. Raychowdhury, S. Schwartz, N. Yosef, C. Malboeuf, D. Lu, et al., A.K.Shalek, R. Satija, J. Shuga, J.J. Trombetta, D. Gennert, D. Lu, P. Chen, R.S. Gertner,J.T. Gaublomme, N. Yosef, et al., K. Shekhar, P. Brodin, M.M. Davis, A.K. Chakraborty,S. Siegert, E. Cabuy, B.G. Scherf, H. Kohler, S. Panda, Y.Z. Le, H.J. Fehling, D. Gaidatzis,M.B. Stadler, B. Roska, N.T. Sweeney, H. Tierney, D.A. Feldheim, F. Tang, C. Barbacioru,Y. Wang, E. Nordman, C. Lee, N. Xu, X. Wang, J. Bodeau, B.B. Tuch, A. Siddiqui, et al.,T. Thorsen, R.W. Roberts, F.H. Arnold, S.R. Quake, P.B. Umbanhowar, V. Prasad, D.A.Weitz, A.S. Utada, A. Fernandez-Nieves, H.A. Stone, D.A. Weitz, L. van der Maaten,G. Hinton, B. Vogelstein, K.W. Kinzler, J.G. Wetmur, N. Davidson, M.L. Whitfield,G. Sherlock, A.J. Saldanha, J.I. Murray, C.A. Ball, K.E. Alexander, J.C. Matese, C.M.Perou, M.M. Hurt, P.O. Brown, D. Botstein, Y. Yang, A. Cvekl, Y.Y. Zhu, E.M. Machleder,A. Chenchik, R. Li, and P.D. Siebert. Highly Parallel Genome-wide Expression Profilingof Individual Cells Using Nanoliter Droplets. Cell, 161(5):1202–14, 5 2015.[122] Florian Buettner, Kedar N Natarajan, F Paolo Casale, Valentina Proserpio, AntonioScialdone, Fabian J Theis, Sarah A Teichmann, John C Marioni, and Oliver Stegle.Computational analysis of cell-to-cell heterogeneity in single-cell RNA-sequencing datareveals hidden subpopulations of cells. Nature Biotechnology, 33(2):155–160, 1 2015.[123] Angelo Duo`, Mark D. Robinson, and Charlotte Soneson. A systematic performanceevaluation of clustering methods for single-cell RNA-seq data. F1000Research, 7:1141, 92018.[124] Elham Azizi, Ambrose J. Carr, George Plitas, Andrew E. Cornish, Catherine Konopacki,Sandhya Prabhakaran, Juozas Nainys, Kenmin Wu, Vaidotas Kiseliovas, Manu Setty,Kristy Choi, Rachel M. Fromme, Phuong Dao, Peter T. McKenney, Ruby C. Wasti,Krishna Kadaveru, Linas Mazutis, Alexander Y. Rudensky, and Dana Peer. Single-CellMap of Diverse Immune Phenotypes in the Breast Tumor Microenvironment. Cell, 0(0), 62018.[125] Woosung Chung, Hye Hyeon Eum, Hae-Ock Lee, Kyung-Min Lee, Han-Byoel Lee, Kyu-Tae Kim, Han Suk Ryu, Sangmin Kim, Jeong Eon Lee, Yeon Hee Park, Zhengyan Kan,187Wonshik Han, and Woong-Yang Park. Single-cell RNA-seq enables comprehensive tumourand immune cell profiling in primary breast cancer. Nature Communications, 8:15081, 52017.[126] Diether Lambrechts, Els Wauters, Bram Boeckx, Sara Aibar, David Nittner, OliverBurton, Ayse Bassez, Herbert Decaluwe´, Andreas Pircher, Kathleen Van den Eynde,Birgit Weynand, Erik Verbeken, Paul De Leyn, Adrian Liston, Johan Vansteenkiste, PeterCarmeliet, Stein Aerts, and Bernard Thienpont. Phenotype molding of stromal cells inthe lung tumor microenvironment. Nature Medicine, 24(8):1277–1289, 8 2018.[127] Cole Trapnell, Davide Cacchiarelli, Jonna Grimsby, Prapti Pokharel, Shuqiang Li, MichaelMorse, Niall J Lennon, Kenneth J Livak, Tarjei S Mikkelsen, and John L Rinn. Thedynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering ofsingle cells. Nature Biotechnology, 32(4):381–386, 4 2014.[128] Manu Setty, Michelle D Tadmor, Shlomit Reich-Zeliger, Omer Angel, Tomer Meir Salame,Pooja Kathail, Kristy Choi, Sean Bendall, Nir Friedman, and Dana Pe’er. Wishboneidentifies bifurcating developmental trajectories from single-cell data. Nature Biotechnology,34(6):637–645, 6 2016.[129] Zhicheng Ji and Hongkai Ji. TSCAN: Pseudo-time reconstruction and evaluation insingle-cell RNA-seq analysis. Nucleic Acids Research, 44(13):e117–e117, 7 2016.[130] Kieran R Campbell, Christopher Yau, and Inanc Birol. A descriptive marker gene approachto single-cell pseudotime inference. Bioinformatics, 6 2018.[131] Mireya Plass, Jordi Solana, F Alexander Wolf, Salah Ayoub, Aristotelis Misios, PetarGlazˇar, Benedikt Obermayer, Fabian J Theis, Christine Kocks, and Nikolaus Rajewsky.Cell type atlas and lineage tree of a whole complex animal by single-cell transcriptomics.Science (New York, N.Y.), 360(6391):eaaq1723, 4 2018.[132] Hirotaka Matsumoto, Hisanori Kiryu, Chikara Furusawa, Minoru S H Ko, Shigeru B H Ko,Norio Gouda, Tetsutaro Hayashi, Itoshi Nikaido, and Ziv Bar-Joseph. SCODE: an efficientregulatory network inference algorithm from single-cell RNA-Seq during differentiation.Bioinformatics, 33(15):2314–2321, 8 2017.[133] Jonathan Ronen and Altuna Akalin. netSmooth: Network-smoothing based imputationfor single cell RNA-seq. F1000Research, 7:8, 2018.188[134] Nicolette G. Alkema, Tushar Tomar, Evelien W. Duiker, Gert Jan Meersma, Harry Klip,Ate G. J. van der Zee, G. Bea A. Wisman, and Steven de Jong. Biobanking of patientand patient-derived xenograft ovarian tumour tissue: efficient preservation with low andhigh fetal calf serum based methods. Scientific Reports, 5(1):14495, 11 2015.[135] Peter Eirew, Adi Steif, Jaswinder Khattra, Gavin Ha, Damian Yap, Hossein Farahani,Karen Gelmon, Stephen Chia, Colin Mar, Adrian Wan, Emma Laks, Justina Biele, KareyShumansky, Jamie Rosner, Andrew McPherson, Cydney Nielsen, Andrew J. L. Roth,Calvin Lefebvre, Ali Bashashati, Camila de Souza, Celia Siu, Radhouane Aniba, JazmineBrimhall, Arusha Oloumi, Tomo Osako, Alejandra Bruna, Jose L. Sandoval, Teresa Algara,Wendy Greenwood, Kaston Leung, Hongwei Cheng, Hui Xue, Yuzhuo Wang, Dong Lin,Andrew J. Mungall, Richard Moore, Yongjun Zhao, Julie Lorette, Long Nguyen, DavidHuntsman, Connie J. Eaves, Carl Hansen, Marco A. Marra, Carlos Caldas, Sohrab P. Shah,and Samuel Aparicio. Dynamics of genomic clones in breast cancer patient xenografts atsingle-cell resolution. Nature, 518(7539):422–426, 11 2014.[136] Bjoern Chapuy, Hongwei Cheng, Akira Watahiki, Matthew D Ducar, Yuxiang Tan,Linfeng Chen, Margaretha G M Roemer, Jing Ouyang, Amanda L Christie, Liye Zhang,Daniel Gusenleitner, Ryan P Abo, Pedro Farinha, Frederike von Bonin, Aaron R Thorner,Heather H Sun, Randy D Gascoyne, Geraldine S Pinkus, Paul van Hummelen, Gerald GWulf, Jon C Aster, David M Weinstock, Stefano Monti, Scott J Rodig, Yuzhuo Wang,and Margaret A Shipp. Diffuse large B-cell lymphoma patient-derived xenograft modelscapture the molecular and biological heterogeneity of the disease. Blood, 127(18):2203–13,2016.[137] Library Prep - Single Cell Gene Expression - Official 10x Genomics Support.[138] Els M J J Berns and David D Bowtell. The Changing View of High-Grade Serous OvarianCancer. 2012.[139] Daniela Luvero, Andrea Milani, and Jonathan A Ledermann. Treatment options inrecurrent ovarian cancer: latest evidence and clinical potential. Therapeutic advances inmedical oncology, 6(5):229–39, 9 2014.[140] Wataru Sakai, Elizabeth M. Swisher, Beth Y. Karlan, Mukesh K. Agarwal, Jake Higgins,Cynthia Friedman, Emily Villegas, Cline Jacquemont, Daniel J. Farrugia, Fergus J. Couch,Nicole Urban, and Toshiyasu Taniguchi. Secondary mutations as a mechanism of cisplatinresistance in BRCA2-mutated cancers. Nature, 451(7182):1116–1120, 2 2008.189[141] Eun Jin Heo, Young Jae Cho, William Chi Cho, Ji Eun Hong, Hye-Kyung Jeon, Doo-Yi Oh,Yoon-La Choi, Sang Yong Song, Jung-Joo Choi, Duk-Soo Bae, Yoo-Young Lee, Chel HunChoi, Tae-Joong Kim, Woong-Yang Park, Byoung-Gie Kim, and Jeong-Won Lee. Patient-Derived Xenograft Models of Epithelial Ovarian Cancer for Preclinical Studies. Cancerresearch and treatment : official journal of Korean Cancer Association, 49(4):915–926, 102017.[142] Clare L Scott, Helen J Mackay, Paul Haluska, and Jr. Patient-derived xenograft modelsin gynecologic malignancies. American Society of Clinical Oncology educational book.American Society of Clinical Oncology. Annual Meeting, pages 258–66, 2014.[143] Ruifen Dong, Wenan Qiang, Haiyang Guo, Xiaofei Xu, J Julie Kim, Andrew Mazar,Beihua Kong, and Jian-Jun Wei. Histologic and molecular analysis of patient derivedxenografts of high-grade serous ovarian carcinoma. Journal of hematology & oncology,9(1):92, 9 2016.[144] Wei-Ting Hwang, Sarah F. Adams, Emin Tahirovic, Ian S. Hagemann, and George Coukos.Prognostic significance of tumor-infiltrating T cells in ovarian cancer: A meta-analysis.Gynecologic Oncology, 124(2):192–198, 2 2012.[145] Andreas Heindl, Chunyan Lan, Daniel Nava Rodrigues, Konrad Koelble, Yinyin Yuan,Andreas Heindl, Chunyan Lan, Daniel Nava Rodrigues, Konrad Koelble, and Yinyin Yuan.Similarity and diversity of the tumor microenvironment in multiple metastases: criticalimplications for overall and progression-free survival of high-grade serous ovarian cancer.Oncotarget, 7(44):71123–71135, 9 2016.[146] Katharina Auer, Anna Bachmayr-Heyda, Nyamdelger Sukhbaatar, Stefanie Aust, Klaus G.Schmetterer, Samuel M. Meier, Christopher Gerner, Christoph Grimm, Reinhard Horvat,Dietmar Pils, Katharina Auer, Anna Bachmayr-Heyda, Nyamdelger Sukhbaatar, StefanieAust, Klaus G. Schmetterer, Samuel M. Meier, Christopher Gerner, Christoph Grimm,Reinhard Horvat, and Dietmar Pils. Role of the immune system in the peritoneal tumorspread of high grade serous ovarian cancer. Oncotarget, 7(38):61336–61354, 9 2016.[147] Alejandro Jime´nez-Sa´nchez, Danish Memon, Stephane Pourpe, Harini Veeraraghavan,Yanyun Li, Hebert Alberto Vargas, Michael B Gill, Kay J Park, Oliver Zivanovic, JasonKonner, Jacob Ricca, Dmitriy Zamarin, Tyler Walther, Carol Aghajanian, Jedd D Wolchok,Evis Sala, Taha Merghoub, Alexandra Snyder, and Martin L Miller. Heterogeneous Tumor-190Immune Microenvironments among Differentially Growing Metastases in an OvarianCancer Patient. Cell, 170(5):927–938, 8 2017.[148] Charlotte S. Lo, Sanaz Sanii, David R. Kroeger, Katy Milne, Aline Talhouk, Derek S.Chiu, Kurosh Rahimi, Patricia A. Shaw, Blaise A. Clarke, and Brad H. Nelson. Neoadju-vant Chemotherapy of Ovarian Cancer Results in Three Patterns of Tumor-InfiltratingLymphocyte Response with Distinct Implications for Immunotherapy. Clinical CancerResearch, 23(4):925–934, 2 2017.[149] Alessandra Cesano. nCounter( R©) PanCancer Immune Profiling Panel (NanoString Tech-nologies, Inc., Seattle, WA). Journal for immunotherapy of cancer, 3:42, 2015.[150] Huei San Leong, Laura Galletta, Dariush Etemadmoghadam, Joshy George, MartinKo¨bel, Susan J Ramus, David Bowtell, and David Bowtell. Efficient molecular subtypeclassification of high-grade serous ovarian cancer. The Journal of Pathology, 236(3):272–277,7 2015.[151] H. Li and R. Durbin. Fast and accurate short read alignment with Burrows-Wheelertransform. Bioinformatics, 25(14):1754–1760, 7 2009.[152] Christopher T. Saunders, Wendy S. W. Wong, Sajani Swamy, Jennifer Becq, Lisa J.Murray, and R. Keira Cheetham. Strelka: accurate somatic small-variant calling fromsequenced tumor-normal sample pairs. Bioinformatics, 28(14):1811–1817, 7 2012.[153] Jiarui Ding, Ali Bashashati, Andrew Roth, Arusha Oloumi, Kane Tse, Thomas Zeng,Gholamreza Haffari, Martin Hirst, Marco A. Marra, Anne Condon, Samuel Aparicio, andSohrab P. Shah. Feature-based classifiers for somatic mutation detection in tumour-normalpaired sequencing data. Bioinformatics, 28(2):167–175, 1 2012.[154] Andrew McPherson, Sohrab P Shah, and S Cenk Sahinalp. deStruct: Accurate Rearrange-ment Detection using Breakpoint Specific Realignment. bioRxiv, 117523, 3 2017.[155] Ryan M Layer, Colby Chiang, Aaron R Quinlan, and Ira M Hall. LUMPY: a probabilisticframework for structural variant discovery. Genome Biology, 15(6):R84, 6 2014.[156] Andrew W. McPherson, Andrew Roth, Gavin Ha, Cedric Chauve, Adi Steif, Camila P. E.de Souza, Peter Eirew, Alexandre Bouchard-Coˆte´, Sam Aparicio, S. Cenk Sahinalp, andSohrab P. Shah. ReMixT: clone-specific genomic structure estimation in cancer. GenomeBiology, 18(1):140, 12 2017.191[157] Gavin Ha, Andrew Roth, Jaswinder Khattra, Julie Ho, Damian Yap, Leah M. Prentice,Nataliya Melnyk, Andrew McPherson, Ali Bashashati, Emma Laks, Justina Biele, JiaruiDing, Alan Le, Jamie Rosner, Karey Shumansky, Marco A. Marra, C. Blake Gilks, David G.Huntsman, Jessica N. McAlpine, Samuel Aparicio, and Sohrab P. Shah. TITAN: inferenceof copy number architectures in clonal cell populations from tumor whole-genome sequencedata. Genome Research, 24(11):1881–1893, 11 2014.[158] Ann-Marie Patch, Elizabeth L. Christie, Dariush Etemadmoghadam, Dale W. Garsed,Joshy George, Sian Fereday, Katia Nones, Prue Cowin, Kathryn Alsop, Peter J. Bailey,Karin S. Kassahn, Felicity Newell, Michael C. J. Quinn, Stephen Kazakoff, Kelly Quek,Charlotte Wilhelm-Benartzi, Ed Curry, Huei San Leong, Anne Hamilton, Linda Mileshkin,George Au-Yeung, Catherine Kennedy, Jillian Hung, Yoke-Eng Chiew, Paul Harnett,Michael Friedlander, Michael Quinn, Jan Pyman, Stephen Cordner, Patricia O?Brien,Jodie Leditschke, Greg Young, Kate Strachan, Paul Waring, Walid Azar, Chris Mitchell,Nadia Traficante, Joy Hendley, Heather Thorne, Mark Shackleton, David K. Miller,Gisela Mir Arnau, Richard W. Tothill, Timothy P. Holloway, Timothy Semple, IvonHarliwong, Craig Nourse, Ehsan Nourbakhsh, Suzanne Manning, Senel Idrisoglu, TimothyJ. C. Bruxner, Angelika N. Christ, Barsha Poudel, Oliver Holmes, Matthew Anderson,Conrad Leonard, Andrew Lonie, Nathan Hall, Scott Wood, Darrin F. Taylor, QinyingXu, J. Lynn Fink, Nick Waddell, Ronny Drapkin, Euan Stronach, Hani Gabra, RobertBrown, Andrea Jewell, Shivashankar H. Nagaraj, Emma Markham, Peter J. Wilson, JasonEllul, Orla McNally, Maria A. Doyle, Ravikiran Vedururu, Collin Stewart, Ernst Lengyel,John V. Pearson, Nicola Waddell, Anna DeFazio, Sean M. Grimmond, David D. L. Bowtell,and David D. L. Bowtell. Whole-genome characterization of chemoresistant ovarian cancer.Nature, 521(7553):489–494, 5 2015.[159] Serena Nik-Zainal, Helen Davies, Johan Staaf, Manasa Ramakrishna, Dominik Glodzik,Xueqing Zou, Inigo Martincorena, Ludmil B Alexandrov, Sancha Martin, David C Wedge,Peter Van Loo, Young Seok Ju, Marcel Smid, Arie B Brinkman, Sandro Morganella,Miriam R Aure, Ole Christian Lingjærde, Anita Langerød, Markus Ringne´r, Sung-MinAhn, Sandrine Boyault, Jane E Brock, Annegien Broeks, Adam Butler, Christine Desmedt,Luc Dirix, Serge Dronov, Aquila Fatima, John A Foekens, Moritz Gerstung, Gerrit K JHooijer, Se Jin Jang, David R Jones, Hyung-Yong Kim, Tari A King, Savitri Krishnamurthy,Hee Jin Lee, Jeong-Yeon Lee, Yilong Li, Stuart McLaren, Andrew Menzies, Ville Mustonen,Sarah O’Meara, Iris Pauporte´, Xavier Pivot, Colin A Purdie, Keiran Raine, KamnaRamakrishnan, F Germn Rodr´ıguez-Gonza´lez, Gilles Romieu, Anieta M Sieuwerts, Peter T192Simpson, Rebecca Shepherd, Lucy Stebbings, Olafur A Stefansson, Jon Teague, StefaniaTommasi, Isabelle Treilleux, Gert G Van den Eynden, Peter Vermeulen, Anne Vincent-Salomon, Lucy Yates, Carlos Caldas, Laura van’t Veer, Andrew Tutt, Stian Knappskog,Benita Kiat Tee Tan, Jos Jonkers, ke Borg, Naoto T Ueno, Christos Sotiriou, Alain Viari,P Andrew Futreal, Peter J Campbell, Paul N Span, Steven Van Laere, Sunil R Lakhani,Jorunn E Eyfjord, Alastair M Thompson, Ewan Birney, Hendrik G Stunnenberg, Marc Jvan de Vijver, John W M Martens, Anne-Lise Børresen-Dale, Andrea L Richardson,Gu Kong, Gilles Thomas, and Michael R Stratton. Landscape of somatic mutations in560 breast cancer whole-genome sequences. Nature, 534(7605):47–54, 2016.[160] H. M. Li, T. Hiroi, Y. Zhang, A. Shi, G. Chen, S. De, E. J. Metter, W. H. Wood, A. Sharov,J. D. Milner, K. G. Becker, M. Zhan, and N.-p. Weng. TCRB repertoire of CD4+ andCD8+ T cells is distinct in richness, distribution, and CDR3 amino acid composition.Journal of Leukocyte Biology, 99(3):505–513, 3 2016.[161] Ryan Emerson, Anna Sherwood, Cindy Desmarais, Sachin Malhotra, Deborah Phippard,and Harlan Robins. Estimating the ratio of CD4+ to CD8+ T cells using high-throughputsequence data. Journal of Immunological Methods, 391(1):14–21, 2013.[162] Paul L Klarenbeek, Marieke E Doorenspleet, Rebecca E E Esveldt, Barbera D C vanSchaik, Neubury Lardy, Antoine H C van Kampen, Paul P Tak, Robert M Plenge, FrankBaas, Paul I W de Bakker, and Niek de Vries. Somatic Variation of T-Cell Receptor GenesStrongly Associate with HLA Class Restriction. PloS one, 10(10):e0140815, 2015.[163] Andrs Szolek, Benjamin Schubert, Christopher Mohr, Marc Sturm, Magdalena Feldhahn,and Oliver Kohlbacher. OptiType: precision HLA typing from next-generation sequencingdata. Bioinformatics (Oxford, England), 30(23):3310–6, 12 2014.[164] Nicola Ternette, Hongbing Yang, Thomas Partridge, Anuska Llano, Samandhy Ceden˜o,Roman Fischer, Philip D. Charles, Nadine L. Dudek, Beatriz Mothe, Manuel Crespo,William M. Fischer, Bette T. M. Korber, Morten Nielsen, Persephone Borrow, Anthony W.Purcell, Christian Brander, Lucy Dorrell, Benedikt M. Kessler, and Tom Hanke. Definingthe HLA class I-associated viral antigen repertoire from HIV-1-infected human cells.European Journal of Immunology, 46(1):60–69, 1 2016.[165] MichaelS. Rooney, SachetA. Shukla, CatherineJ. Wu, Gad Getz, and Nir Hacohen. Molec-ular and Genetic Properties of Tumors Associated with Local Immune Cytolytic Activity.Cell, 160(1-2):48–61, 1 2015.193[166] Arthur Getis and J. K. Ord. The Analysis of Spatial Association by Use of DistanceStatistics. Geographical Analysis, 24(3):189–206, 9 1992.[167] Sidra Nawaz, Andreas Heindl, Konrad Koelble, and Yinyin Yuan. Beyond immune density:critical role of spatial heterogeneity in estrogen receptor-negative breast cancer. Modernpathology : an official journal of the United States and Canadian Academy of Pathology,Inc, 28(6):766–77, 6 2015.[168] Sayuri Yoshihama, Jason Roszik, Isaac Downs, Torsten B. Meissner, Saptha Vijayan,Bjoern Chapuy, Tabasum Sidiq, Margaret A. Shipp, Gregory A. Lizee, and Koichi S.Kobayashi. NLRC5/MHC class I transactivator is a target for immune evasion in cancer.Proceedings of the National Academy of Sciences, 113(21):5999–6004, 5 2016.[169] S. Spranger, R. M. Spaapen, Y. Zha, J. Williams, Y. Meng, T. T. Ha, and T. F. Gajewski.Up-Regulation of PD-L1, IDO, and Tregs in the Melanoma Tumor Microenvironment IsDriven by CD8+ T Cells. Science Translational Medicine, 5(200):116–200, 8 2013.[170] Marcel Smid, F Germn Rodr´ıguez-Gonza´lez, Anieta M Sieuwerts, Roberto Salgado, WendyJ C Prager-Van der Smissen, Michelle van der Vlugt-Daane, Anne van Galen, SerenaNik-Zainal, Johan Staaf, Arie B Brinkman, Marc J van de Vijver, Andrea L Richardson,Aquila Fatima, Kim Berentsen, Adam Butler, Sancha Martin, Helen R Davies, RenoDebets, Marion E Meijer-Van Gelder, Carolien H M van Deurzen, Gatan MacGrogan, GertG G M Van den Eynden, Colin Purdie, Alastair M Thompson, Carlos Caldas, Paul N Span,Peter T Simpson, Sunil R Lakhani, Steven Van Laere, Christine Desmedt, Markus Ringne´r,Stefania Tommasi, Jorunn Eyford, Annegien Broeks, Anne Vincent-Salomon, P AndrewFutreal, Stian Knappskog, Tari King, Gilles Thomas, Alain Viari, Anita Langerød, Anne-Lise Børresen-Dale, Ewan Birney, Hendrik G Stunnenberg, Mike Stratton, John A Foekens,and John W M Martens. Breast cancer genome and transcriptome integration implicatesspecific mutational signatures with immune cell infiltration. Nature communications,7:12910, 9 2016.[171] N. McGranahan, A. J. S. Furness, R. Rosenthal, S. Ramskov, R. Lyngaa, S. K. Saini,M. Jamal-Hanjani, G. A. Wilson, N. J. Birkbak, C. T. Hiley, T. B. K. Watkins, S. Shafi,N. Murugaesu, R. Mitter, A. U. Akarca, J. Linares, T. Marafioti, J. Y. Henry, E. M.Van Allen, D. Miao, B. Schilling, D. Schadendorf, L. A. Garraway, V. Makarov, N. A.Rizvi, A. Snyder, M. D. Hellmann, T. Merghoub, J. D. Wolchok, S. A. Shukla, C. J.Wu, K. S. Peggs, T. A. Chan, S. R. Hadrup, S. A. Quezada, and C. Swanton. Clonal194neoantigens elicit T cell immunoreactivity and sensitivity to immune checkpoint blockade.Science, 351(6280):1463–1469, 3 2016.[172] Ovarian Tumor Tissue Analysis Consortium, Ellen L Goode, Matthew Block, Kimberly RKalli, Robert A Vierkant, Wenqian Chen, Zachary Fogarty, Aleksandra Gentry-Maharaj,Aleksandra A To loczko, Alexander Hein, Aliecia L Bouligny, Allan Jensen, Ana Osorio,Andreas Hartkopf, Andy Ryan, Anita Chudecka-G laz, Anthony M Magliocco, ArndtHartmann, Audrey Y Jung, Bo Gao, Brenda Y Hernandez, Brooke L Fridley, Bryan MMcCauley, Catherine J Kennedy, Chen Wang, Chloe Karpinskyj, Christiani B de Sousa,Daniel G Tiezzi, David L Wachter, Esther Herpel, Florin Andrei Taran, Francesmary Mod-ugno, Gregg Nelson, Jan Lubin´ski, Janusz Menkiszak, Jennifer Alsop, Jenny Lester, JessGarc´ıa-Donas, Jill Nation, Jillian Hung, Jos Palacios, Joseph H Rothstein, Joseph L Kelley,Jurandyr M de Andrade, Luis Robles-Dı´az, Maria P Intermaggio, Martin Widschwendter,Matthias W Beckmann, Matthias Ruebner, Mercedes Jimenez-Linan, Naveena Singh,Oleg Oszurek, Paul R Harnett, Peter F Rambau, Peter Sinn, Philipp Wagner, PrafullGhatage, Raghwa Sharma, Robert P Edwards, Roberta B Ness, Sandra Orsulic, Sara YBrucker, Sharon E Johnatty, Teri A Longacre, Eilber Ursula, Valerie McGuire, WeivaSieh, Yanina Natanzon, Zheng Li, Alice S Whittemore, DeFazio Anna, Annette Staebler,Beth Y Karlan, Blake Gilks, David D Bowtell, Estrid Høgdall, Francisco J Candido dosReis, Helen Steed, Ian G Campbell, Jacek Gronwald, Javier Ben´ıtez, Jennifer M Koziak,Jenny Chang-Claude, Kirsten B Moysich, Linda E Kelemen, Linda S Cook, Marc TGoodman, Mara Jos Garc´ıa, Peter A Fasching, Stefan Kommoss, Suha Deen, Susanne KKjaer, Usha Menon, James D Brenton, Paul DP Pharoah, Georgia Chenevix-Trench,David G Huntsman, Stacey J Winham, Martin Ko¨bel, and Susan J Ramus. Dose-ResponseAssociation of CD8+ Tumor-Infiltrating Lymphocytes and Survival Time in High-GradeSerous Ovarian Cancer. JAMA oncology, 3(12):e173290, 2017.[173] Poorval M Joshi, Shari L Sutor, Catherine J Huntoon, and Larry M Karnitz. Ovariancancer-associated mutations disable catalytic activity of CDK12, a kinase that promoteshomologous recombination repair and resistance to cisplatin and poly(ADP-ribose) poly-merase inhibitors. The Journal of biological chemistry, 289(13):9247–53, 3 2014.[174] Roy S. Herbst, Jean-Charles Soria, Marcin Kowanetz, Gregg D. Fine, Omid Hamid,Michael S. Gordon, Jeffery A. Sosman, David F. McDermott, John D. Powderly, Scott N.Gettinger, Holbrook E. K. Kohrt, Leora Horn, Donald P. Lawrence, Sandra Rost, MayaLeabman, Yuanyuan Xiao, Ahmad Mokatrin, Hartmut Koeppen, Priti S. Hegde, Ira195Mellman, Daniel S. Chen, and F. Stephen Hodi. Predictive correlates of response to theanti-PD-L1 antibody MPDL3280A in cancer patients. Nature, 515(7528):563–567, 11 2014.[175] Kamil A Lipinski, Louise J Barber, Matthew N Davies, Matthew Ashenden, AndreaSottoriva, and Marco Gerlinger. Cancer Evolution and the Limits of Predictability inPrecision Cancer Medicine. Trends in cancer, 2(1):49–63, 1 2016.[176] Tabula Muris Consortium and others. Single-cell transcriptomics of 20 mouse organscreates a Tabula Muris. Nature, 2018.[177] Vladimir Yu Kiselev, Kristina Kirschner, Michael T Schaub, Tallulah Andrews, AndrewYiu, Tamir Chandra, Kedar N Natarajan, Wolf Reik, Mauricio Barahona, Anthony RGreen, and others. SC3: consensus clustering of single-cell RNA-seq data. Nature methods,14(5):483, 2017.[178] Andrew Butler, Paul Hoffman, Peter Smibert, Efthymia Papalexi, and Rahul Satija.Integrating single-cell transcriptomic data across different conditions, technologies, andspecies. Nature Biotechnology, 2018.[179] Justina Zurauskiene and Christopher Yau. pcaReduce: hierarchical clustering of singlecell transcriptional profiles. BMC bioinformatics, 17(1):140, 2016.[180] Jacob H Levine, Erin F Simonds, Sean C Bendall, Kara L Davis, D Amir El-ad, Michelle DTadmor, Oren Litvin, Harris G Fienberg, Astraea Jager, Eli R Zunder, and others. Data-driven phenotypic dissection of AML reveals progenitor-like cells that correlate withprognosis. Cell, 162(1):184–197, 2015.[181] Angelo Duo`, Mark D Robinson, and Charlotte Soneson. A systematic performanceevaluation of clustering methods for single-cell RNA-seq data. F1000Research, 7, 2018.[182] Saskia Freytag, Luyi Tian, Ingrid Lo¨nnstedt, Milica Ng, and Melanie Bahlo. Comparisonof clustering tools in R for medium-sized 10x Genomics single-cell RNA-sequencing data.F1000Research, 7:1297, 8 2018.[183] Vladimir Yu Kiselev, Tallulah S. Andrews, and Martin Hemberg. Challenges in unsu-pervised clustering of single-cell RNA-seq data. Nature Reviews Genetics, page 1, 12019.[184] Xinxin Zhang, Yujia Lan, Jinyuan Xu, Fei Quan, Erjie Zhao, Chunyu Deng, Tao Luo,Liwen Xu, Gaoming Liao, Min Yan, Yanyan Ping, Feng Li, Aiai Shi, Jing Bai, Tingting196Zhao, Xia Li, and Yun Xiao. CellMarker: a manually curated resource of cell markers inhuman and mouse. Nucleic Acids Research, 10 2018.[185] Vladimir Yu Kiselev, Andrew Yiu, and Martin Hemberg. scmap: projection of single-cellRNA-seq data across data sets. Nature methods, 15(5):359, 2018.[186] Stephanie C Hicks, F William Townes, Mingxiang Teng, and Rafael A Irizarry. Missingdata and technical variability in single-cell RNA-sequencing experiments. Biostatistics,2017.[187] Pang Wei Koh, Rahul Sinha, Amira A. Barkal, Rachel M. Morganti, Angela Chen, Irving L.Weissman, Lay Teng Ang, Anshul Kundaje, and Kyle M. Loh. An atlas of transcriptional,chromatin accessibility, and surface marker changes in human mesoderm development.Scientific Data, 3:160109, 12 2016.[188] Sylvia Richardson John C Marioni Catalina A Vallejos Nils Eling Arianne C. Richard.Correcting the Mean-Variance Dependency for Differential Variability Testing UsingSingle-Cell RNA Sequencing Data. Cell Systems, 7, 2018.[189] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXivpreprint arXiv:1412.6980, 2014.[190] Mart\’\in˜Abadi, Ashish˜Agarwal, Paul˜Barham, Eugene˜Brevdo, Zhifeng˜Chen,Craig˜Citro, Greg˜S.˜Corrado, Andy˜Davis, Jeffrey˜Dean, Matthieu˜Devin, Sanjay˜Ghe-mawat, Ian˜Goodfellow, Andrew˜Harp, Geoffrey˜Irving, Michael˜Isard, Yangqing Jia,Rafal˜Jozefowicz, Lukasz˜Kaiser, Manjunath˜Kudlur, Josh˜Levenberg, Dandelion˜Mane´,Rajat˜Monga, Sherry˜Moore, Derek˜Murray, Chris˜Olah, Mike˜Schuster, Jonathon˜Sh-lens, Benoit˜Steiner, Ilya˜Sutskever, Kunal˜Talwar, Paul˜Tucker, Vincent˜Vanhoucke,Vijay˜Vasudevan, Fernanda˜Vie´gas, Oriol˜Vinyals, Pete˜Warden, Martin˜Wattenberg,Martin˜Wicke, Yuan˜Yu, and Xiaoqiang˜Zheng. {TensorFlow}: Large-Scale MachineLearning on Heterogeneous Systems, 2015.[191] Debajyoti Sinha, Akhilesh Kumar, Himanshu Kumar, Sanghamitra Bandyopadhyay, andDebarka Sengupta. dropClust: efficient clustering of ultra-large scRNA-seq data. NucleicAcids Research, 46(6):e36–e36, 4 2018.[192] Aaron T.L. Lun, Davis J. McCarthy, and John C. Marioni. A step-by-step workflow forlow-level analysis of single-cell RNA-seq data with Bioconductor. F1000Research, 5:2122,10 2016.197[193] Jacob H. Levine, Erin F. Simonds, Sean C. Bendall, Kara L. Davis, El-ad D. Amir,Michelle D. Tadmor, Oren Litvin, Harris G. Fienberg, Astraea Jager, Eli R. Zunder,Rachel Finck, Amanda L. Gedman, Ina Radtke, James R. Downing, Dana Peer, andGarry P. Nolan. Data-Driven Phenotypic Dissection of AML Reveals Progenitor-like Cellsthat Correlate with Prognosis. Cell, 162(1):184–197, 7 2015.[194] Aaron M Newman, Chih Long Liu, Michael R Green, Andrew J Gentles, Weiguo Feng,Yue Xu, Chuong D Hoang, Maximilian Diehn, and Ash A Alizadeh. Robust enumerationof cell subsets from tissue expression profiles. Nature Methods, 12(5):453–457, 5 2015.[195] Monika Trzpis, Pamela M J McLaughlin, Lou M F H de Leij, and Martin C Harmsen.Epithelial cell adhesion molecule: more than a carcinoma marker and adhesion molecule.The American journal of pathology, 171(2):386–95, 8 2007.[196] Melissa G Mendez, Shin-Ichiro Kojima, and Robert D Goldman. Vimentin induceschanges in cell shape, motility, and adhesion during the epithelial to mesenchymal tran-sition. FASEB journal : official publication of the Federation of American Societies forExperimental Biology, 24(6):1838–51, 6 2010.[197] Jamie R Privratsky and Peter J Newman. PECAM-1: regulator of endothelial junctionalintegrity. Cell and tissue research, 355(3):607–19, 3 2014.[198] Mathias Uhlen, Cheng Zhang, Sunjae Lee, Evelina Sjo¨stedt, Linn Fagerberg, GholamrezaBidkhori, Rui Benfeitas, Muhammad Arif, Zhengtao Liu, Fredrik Edfors, Kemal Sanli,Kalle von Feilitzen, Per Oksvold, Emma Lundberg, Sophia Hober, Peter Nilsson, JohannaMattsson, Jochen M. Schwenk, Hans Brunnstro¨m, Bengt Glimelius, Tobias Sjo¨blom,Per-Henrik Edqvist, Dijana Djureinovic, Patrick Micke, Cecilia Lindskog, Adil Mardinoglu,and Fredrik Ponten. A pathology atlas of the human cancer transcriptome. Science,357(6352):eaan2507, 8 2017.[199] M Mura, R K Swain, X Zhuang, H Vorschmitt, G Reynolds, S Durant, J F J Beesley,J M J Herbert, H Sheldon, M Andre, S Sanderson, K Glen, N-T Luu, H M McGettrick,P Antczak, F Falciani, G B Nash, Z S Nagy, and R Bicknell. Identification and angiogenicrole of the novel tumor endothelial marker CLEC14A. Oncogene, 31(3):293–305, 1 2012.[200] Antonio Scialdone, Kedar N Natarajan, Luis R Saraiva, Valentina Proserpio, Sarah A Te-ichmann, Oliver Stegle, John C Marioni, and Florian Buettner. Computational assignmentof cell-cycle stage from single-cell transcriptome data. Methods, 85:54–61, 2015.198[201] Jiarui Ding, Anne Condon, and Sohrab P. Shah. Interpretable dimensionality reductionof single cell transcriptome data with deep generative models. Nature Communications,9(1):2002, 12 2018.[202] D Payne, S Drinkwater, R Baretto, M Duddridge, and M J Browning. Expression ofchemokine receptors CXCR4, CXCR5 and CCR7 on B and T lymphocytes from patientswith primary antibody deficiency. Clinical and experimental immunology, 156(2):254–62,5 2009.[203] Antonio Fabregat, Steven Jupe, Lisa Matthews, Konstantinos Sidiropoulos, Marc Gillespie,Phani Garapati, Robin Haw, Bijay Jassal, Florian Korninger, Bruce May, Marija Milacic,Corina Duenas Roca, Karen Rothfels, Cristoffer Sevilla, Veronica Shamovsky, SolomonShorser, Thawfeek Varusai, Guilherme Viteri, Joel Weiser, Guanming Wu, Lincoln Stein,Henning Hermjakob, and Peter D’Eustachio. The Reactome Pathway Knowledgebase.Nucleic Acids Research, 46(D1):D649–D655, 1 2018.[204] Arthur Liberzon, Chet Birger, Helga Thorvaldsdo´ttir, Mahmoud Ghandi, JillP. Mesirov,and Pablo Tamayo. The Molecular Signatures Database Hallmark Gene Set Collection.Cell Systems, 1(6):417–425, 12 2015.[205] Robert Kridel, Fong Chun Chan, Anja Mottok, Merrill Boyle, Pedro Farinha, King Tan,Barbara Meissner, Ali Bashashati, Andrew McPherson, Andrew Roth, Karey Shumansky,Damian Yap, Susana Ben-Neriah, Jamie Rosner, Maia A. Smith, Cydney Nielsen, EvaGine´, Adele Telenius, Daisuke Ennishi, Andrew Mungall, Richard Moore, Ryan D. Morin,Nathalie A. Johnson, Laurie H. Sehn, Thomas Tousseyn, Ahmet Dogan, Joseph M.Connors, David W. Scott, Christian Steidl, Marco A. Marra, Randy D. Gascoyne, andSohrab P. Shah. Histological Transformation and Progression in Follicular Lymphoma: AClonal Evolution Study. PLOS Medicine, 13(12):e1002197, 12 2016.[206] Jiarui Ding, Sohrab Shah, and Anne Condon. densityCut: an efficient and versatile topolog-ical approach for automatic clustering of biological data. Bioinformatics, 32(17):2567–2576,9 2016.[207] Peter Langfelder, Bin Zhang, and Steve Horvath. Defining clusters from a hierarchicalcluster tree: the Dynamic Tree Cut package for R. Bioinformatics, 24(5):719–720, 3 2008.[208] Fionnuala P. O’Connell, Jack L. Pinkus, and Geraldine S. Pinkus. CD138 (Syndecan-1), aPlasma Cell Marker Immunohistochemical Profile in Hematopoietic and NonhematopoieticNeoplasms. American Journal of Clinical Pathology, 121(2):254–263, 2 2004.199[209] R J Paul. The role of phospholamban and SERCA3 in regulation of smooth muscle-endothelial cell signalling mechanisms: evidence from gene-ablated mice. Acta physiologicaScandinavica, 164(4):589–97, 12 1998.[210] Henrik Lindskog, Elisabet Athley, Erik Larsson, and Samuel Lundin. New Insights toVascular Smooth Muscle Cell and Pericyte Differentiation of Mouse Embryonic Stem CellsIn Vitro. 2006.[211] O Skalli, M F Pelte, M C Peclet, G Gabbiani, P Gugliotta, G Bussolati, M Ravazzola,and L Orci. Alpha-smooth muscle actin, a differentiation marker of smooth musclecells, is present in microfilamentous bundles of pericytes. Journal of Histochemistry &Cytochemistry, 37(3):315–321, 3 1989.[212] G Gabbiani, E Schmid, S Winter, C Chaponnier, C de Ckhastonay, J Vandekerckhove,K Weber, and W W Franke. Vascular smooth muscle cells differ from other smooth musclecells: predominance of vimentin filaments and a specific alpha-type actin. Proceedings ofthe National Academy of Sciences of the United States of America, 78(1):298–302, 1 1981.[213] Aline Lopes Ribeiro and Oswaldo Keith Okamoto. Combined effects of pericytes in thetumor microenvironment. Stem cells international, 2015:868475, 2015.[214] Leland McInnes and John Healy. Umap: Uniform manifold approximation and projectionfor dimension reduction. arXiv preprint arXiv:1802.03426, 2018.[215] Roy Jefferis and Marie-Paule Lefranc. Human immunoglobulin allotypes: possible implica-tions for immunogenicity. mAbs, 1(4):332–8, 2009.[216] O Hermine, C Haioun, E Lepage, M F d’Agay, J Briere, C Lavignac, G Fillet, G Salles,J P Marolleau, J Diebold, F Reyas, and P Gaulard. Prognostic significance of bcl-2 proteinexpression in aggressive non-Hodgkin’s lymphoma. Groupe d’Etude des Lymphomes del’Adulte (GELA). Blood, 87(1):265–72, 1 1996.[217] Keni Gu, Kai Fu, Smrati Jain, Zhongfen Liu, Javeed Iqbal, Min Li, Warren G Sanger,Dennis D Weisenburger, Timothy C Greiner, Patricia Aoun, Bhavana J Dave, and Wing CChan. t(14;18)-negative follicular lymphomas are associated with a high frequency of BCL6rearrangement at the alternative breakpoint region. Modern Pathology, 22(9):1251–1257,9 2009.200[218] Katerina Hatzi and Ari Melnick. Breaking bad in the germinal center: how deregulationof BCL6 contributes to lymphomagenesis. Trends in molecular medicine, 20(6):343–52, 62014.[219] Bailey E Freeman, Erika Hammarlund, Hans-Peter Raue´, and Mark K Slifka. Regulationof innate CD8 + T-cell activation mediated by cytokines.[220] Michael R. Green, Shingo Kihira, Chih Long Liu, Ramesh V. Nair, Raheleh Salari,Andrew J. Gentles, Jonathan Irish, Henning Stehr, Carolina Vicente-Duen˜as, IsabelRomero-Camarero, Isidro Sanchez-Garcia, Sylvia K. Plevritis, Daniel A. Arber, SerafimBatzoglou, Ronald Levy, and Ash A. Alizadeh. Mutations in early follicular lymphomaprogenitors are associated with suppressed antigen presentation. Proceedings of theNational Academy of Sciences, 112(10):E1116–E1125, 3 2015.[221] Byungjin Hwang, Ji Hyun Lee, and Duhee Bang. Single-cell RNA sequencing technologiesand bioinformatics pipelines. Experimental & Molecular Medicine, 50(8):96, 8 2018.[222] Julie R. Brahmer, Scott S. Tykodi, Laura Q.M. Chow, Wen-Jen Hwu, Suzanne L. Topalian,Patrick Hwu, Charles G. Drake, Luis H. Camacho, John Kauh, Kunle Odunsi, Henry C.Pitot, Omid Hamid, Shailender Bhatia, Renato Martins, Keith Eaton, Shuming Chen,Theresa M. Salay, Suresh Alaparthy, Joseph F. Grosso, Alan J. Korman, Susan M.Parker, Shruti Agrawal, Stacie M. Goldberg, Drew M. Pardoll, Ashok Gupta, and Jon M.Wigginton. Safety and Activity of Anti-PD-L1 Antibody in Patients with AdvancedCancer. New England Journal of Medicine, 366(26):2455–2465, 6 2012.[223] Jedd D Wolchok, Harriet Kluger, Margaret K Callahan, Michael A Postow, Naiyer A Rizvi,Alexander M Lesokhin, Neil H Segal, Charlotte E Ariyan, Ruth-Ann Gordon, KathleenReed, Matthew M Burke, Anne Caldwell, Stephanie A Kronenberg, Blessing U Agunwamba,Xiaoling Zhang, Israel Lowy, Hector David Inzunza, William Feely, Christine E Horak,Quan Hong, Alan J Korman, Jon M Wigginton, Ashok Gupta, and Mario Sznol. Nivolumabplus ipilimumab in advanced melanoma. The New England journal of medicine, 369(2):122–33, 7 2013.[224] Stephen J. Schuster, Michael R. Bishop, Constantine S. Tam, Edmund K. Waller, PeterBorchmann, Joseph P. McGuirk, Ulrich Ja¨ger, Samantha Jaglowski, Charalambos An-dreadis, Jason R. Westin, Isabelle Fleury, Veronika Bachanova, S. Ronan Foley, P. JoyHo, Stephan Mielke, John M. Magenau, Harald Holte, Serafino Pantano, Lida B. Pacaud,201Rakesh Awasthi, Jufen Chu, zlem Anak, Gilles Salles, and Richard T. Maziarz. Tisagenle-cleucel in Adult Relapsed or Refractory Diffuse Large B-Cell Lymphoma. New EnglandJournal of Medicine, page NEJMoa1804980, 12 2018.[225] Jennifer L Barnas, Michelle R Simpson-Abelson, Sandra J Yokota, Raymond J Kelleher,Richard B Bankert, and Richard B. Bankert. T cells and stromal fibroblasts in human tumormicroenvironments represent potential therapeutic targets. Cancer microenvironment :official journal of the International Cancer Microenvironment Society, 3(1):29–47, 3 2010.[226] Xuefei Li, Tina Gruosso, Dongmei Zuo, Atilla Omeroglu, Sarkis Meterissian, Marie-Christine Guiot, Adam Salazar, Morag Park, and Herbert Levine. Infiltration of CD8+ Tcells into tumor-cell clusters in Triple Negative Breast Cancer. bioRxiv, page 430413, 102018.[227] Kieran R Campbell, Adi Steif, Emma Laks, Hans Zahn, Daniel Lai, Andrew McPherson,Hossein Farahani, Farhia Kabeer, Ciara O’Flanagan, Justina Biele, Jazmine Brimhall,Beixi Wang, Pascale Walters, IMAXT Consortium, Alexandre Bouchard-Coˆte´, SamuelAparicio, and Sohrab P Shah. clonealign: statistical integration of independent single-cellRNA &amp; DNA-seq from human cancers. bioRxiv, page 344309, 6 2018.[228] Richard B. Bankert, Sathy V. Balu-Iyer, Kunle Odunsi, Leonard D. Shultz, Raymond J.Kelleher, Jennifer L. Barnas, Michelle Simpson-Abelson, Robert Parsons, and Sandra J.Yokota. Humanized Mouse Model of Ovarian Cancer Recapitulates Patient Solid TumorProgression, Ascites Formation, and Metastasis. PLoS ONE, 6(9):e24420, 9 2011.[229] Nicole C Walsh, Laurie L Kenney, Sonal Jangalwe, Ken-Edwin Aryee, Dale L Greiner,Michael A Brehm, and Leonard D Shultz. Humanized Mouse Models of Clinical Disease.Annual review of pathology, 12:187–215, 1 2017.[230] Andrew J. Shih, Andrew Menzin, Jill Whyte, John Lovecchio, Anthony Liew, HoumanKhalili, Tawfiqul Bhuiya, Peter K. Gregersen, and Annette T. Lee. Identification ofgrade and origin specific cell populations in serous epithelial ovarian cancer by single cellRNA-seq. PLOS ONE, 13(11):e0206785, 11 2018.[231] Allen W. Zhang, Andrew McPherson, Katy Milne, David R. Kroeger, Phineas T. Hamilton,Alex Miranda, Tyler Funnell, Sonya Laan, Dawn R. Cochrane, Jamie L. P. Lim, WinnieYang, Andrew Roth, Maia A. Smith, Camila de Souza, Julie Ho, Kane Tse, ThomasZeng, Inna Shlafman, Michael R. Mayo, Richard Moore, Henrik Failmezger, Andreas202Heindl, Yi Kan Wang, Ali Bashashati, Scott D. Brown, Daniel Lai, Adrian N. C. Wan,Cydney B. Nielsen, Alexandre Bouchard-Cote, Yinyin Yuan, Wyeth W. Wasserman,C. Blake Gilks, Anthony N. Karnezis, Samuel Aparicio, Jessica N. McAlpine, David G.Huntsman, Robert A. Holt, Brad H. Nelson, and Sohrab P. Shah. The interface ofmalignant and immunologic clonal dynamics in high-grade serous ovarian cancer. bioRxiv,page 198101, 10 2017.[232] Alejandro Jime´nez-Sa´nchez, Paulina Cybulska, Katherine Lavigne, Tyler Walther, InesNikolovski, Yousef Mazaheri, Britta Weigelt, Dennis S Chi, Kay J Park, Travis Hollmann,Dominique-Laurent Couturier, Alberto Vargas, James D Brenton, Evis Sala, AlexandraSnyder, and Martin L Miller. Unraveling Tumor-Immune Heterogeneity in AdvancedOvarian Cancer Uncovers Immunogenic Effect of Chemotherapy.[233] K. H. Chen, A. N. Boettiger, J. R. Moffitt, S. Wang, and X. Zhuang. Spatially resolved,highly multiplexed RNA profiling in single cells. Science, 348(6233):aaa6090–aaa6090, 42015.[234] Xiao Wang, William E Allen, Matthew A Wright, Emily L Sylwestrak, Nikolay Samusik,Sam Vesuna, Kathryn Evans, Cindy Liu, Charu Ramakrishnan, Jia Liu, Garry P Nolan,Felice-Alessio Bava, and Karl Deisseroth. Three-dimensional intact-tissue sequencing ofsingle-cell transcriptional states. Science (New York, N.Y.), 361(6400):eaat5691, 6 2018.203Appendices204Appendix AChapter 3 Supplementary MaterialsSupplementary Table A.1 Related to: Figure 3.10. Primers used for deep amplicon sequencingand clonal inference.Supplementary Table A.2 Related to: Figure 3.1, Figure 3.8, and Figure 3.4. TIL densities,TIL-based clusters, molecular subtypes, epithelial colocalization measures from histologic imageanalysis, somatic SNV and rearrangement counts, ITH measures, and TCR and BCR repertoirediversity. NA: cannot be computed/data not available.Supplementary Table A.3 Nonsynonymous SNVs and the highest predicted affinity neoepitopefor each neoantigen, after filtering for HLA LOH. Observed and expected subclonal neoantigen rates,and subclonal neoantigen depletion indices for each multisite HGSC sample.Supplementary Table A.4 Related to: Figure 3.7. HLA-A, HLA-B, and HLA-C germline callsand LOH predictions for multisite HGSC patients. The “clonality” column indicates whether theLOH event is clonal or subclonal.Supplementary Table A.5 Related to: Figure 3.12. Mutation signature proportions andmutational subtype assignments for multisite HGSC (labeled as ITH), OV-AU, and [3] (labeled asOV133) patients.Supplementary Table A.6 Related to: Figure 3.12. Differentially expressed genes between HRD(HRD-DUP + HRD-DEL), FBI, and TD groups in the OV-AU cohort.Supplementary Table A.7 Related to: Figure 3.12. Foldback-HLAMP status and cytotoxicityexpression values for TCGA ovarian serous cystadenocarcinoma samples.205Appendix BChapter 4 Supplementary MaterialsSupplementary Table B.1 Performance measures on simulated data.Supplementary Table B.2 Marker gene matrices used in analysis.Supplementary Table B.3 Pathway enrichment results for follicular lymphoma data, by celltype.206

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0379237/manifest

Comment

Related Items