UBC Faculty Research and Publications

Balony: a software package for analysis of data generated by synthetic genetic array experiments Young, Barry P; Loewen, Christopher J Dec 4, 2013

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

Item Metadata


52383-12859_2013_Article_6219.pdf [ 892.31kB ]
JSON: 52383-1.0223461.json
JSON-LD: 52383-1.0223461-ld.json
RDF/XML (Pretty): 52383-1.0223461-rdf.xml
RDF/JSON: 52383-1.0223461-rdf.json
Turtle: 52383-1.0223461-turtle.txt
N-Triples: 52383-1.0223461-rdf-ntriples.txt
Original Record: 52383-1.0223461-source.json
Full Text

Full Text

SOFTWARE Open AccessBalony: a software package for analysis of datagenerated by synthetic genetic array experimentsBarry P Young* and Christopher JR LoewenAbstractBackground: Synthetic Genetic Array (SGA) analysis is a procedure which has been developed to allow the systematicexamination of large numbers of double mutants in the yeast Saccharomyces cerevisiae. The aim of these experimentsis to identify genetic interactions between pairs of genes. These experiments generate a number of images of orderedarrays of yeast colonies which must be analyzed in order to quantify the extent of the genetic interactions. We havedesigned software that is able to analyze virtually any image of regularly arrayed colonies and allows the usersignificant flexibility over the analysis procedure.Results: “Balony” is freely available software which enables the extraction of quantitative data from array-based geneticscreens. The program follows a multi-step process, beginning with the optional preparation of plate images fromsingle or composite images. Next, the colonies are identified on a plate and the pixel area of each is measured.This is followed by a scoring module which normalizes data and pairs control and experimental data files. The final stepis analysis of the scored data, where the strength and reproducibility of genetic interactions can be visualized andcross-referenced with information on each gene to provide biological insights into the results of the screen.Conclusions: Analysis of SGA screens with Balony can be either automated or highly interactive, enabling the user tocustomize the process to their specific needs. Quantitative data can be extracted at each stage for external analysis ifrequired. Beyond SGA, this software can be used for analyzing many types of plate-based high-throughput screens.Keywords: Yeast, Array, Image analysis, SoftwareBackgroundThe development of high-throughput array-based tech-nologies such as SGA analysis [1,2] has led to a rapidincrease in the popularity of systematic genome-widegenetic interaction screening in yeast. In a typical SGAexperiment, a query yeast strain containing a singlespecified gene deletion is mated to the yeast haploiddeletion collection of ~4800 individual gene deletionmutants arrayed in colonies on agar plates. Followingdiploid selection, sporulation and selection of haploids,a complete set of double mutant strains is generated,which can be used to define the spectrum of geneticinteractions for the query gene, thus providing unbiasedinformation about its function in the cell. In recentyears, the availability of relatively low-cost roboticplatforms such as the Singer RoToR HDA (http://www.singerinstruments.com) has led to the uptake of thistechnology by an increasing number of non-specialistlaboratories. However, the lack of availability of specializedsoftware for the analysis and quantitation of array colonieshas hampered these efforts.In an SGA experiment, after completion of the roboticpinning steps the experimenter is presented with a sub-stantial number of agar plates containing ordered arrays ofdifferently sized single and double mutant yeast colonies.The relative size of the colonies represents the fitness ofeach strain, which can be used as measure of the strengthof a genetic interaction. In order derive meaningful geneticinteraction data from these arrays, the size of each colonyneeds to be precisely measured and the data normalizedand compared with an appropriate control. Given thata single SGA experiment can result in numerous (oftenup to 50) replicates of arrays, each containing up to1,536 colonies per array, it is desirable that such analysescan be carried out in a high-throughput manner withas much automation as possible. Nevertheless, most* Correspondence: barry.young@ubc.caDepartment of Cellular and Physiological Sciences, Life Sciences Institute,University of British Columbia, 2350 Health Sciences Mall, Vancouver, BritishColumbia V6T 1Z3, Canada© 2013 Young and Loewen; licensee BioMed Central Ltd. This is an open access article distributed under the terms of theCreative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use,distribution, and reproduction in any medium, provided the original work is properly cited.Young and Loewen BMC Bioinformatics 2013, 14:354http://www.biomedcentral.com/1471-2105/14/354experimenters will want some degree of control overthe measurement process, so providing a level ofinteractivity will improve overall confidence in thefinal results.One problem we encountered when we first attemptedto analyze images using existing software packages (e.g.“ScreenMill” [3]) was that they tended to be designedwith a particular image format in mind. Although theywere effective at analyzing sample images provided witheach program, we were unable to analyze images thatwe had obtained ourselves. Furthermore, if the softwarewas unable to identify the colonies on a plate, the programwould fail with little recourse available. While theseprograms present a simple interface to the user, it is notpossible to adjust the imaging parameters that mightenable successful analysis of an image.To that end, we sought to develop a program “Balony”,that would be able to analyze images regardless oftheir specific properties, and with the flexibility toutilize arrays of any possible format. Although we findthat the default settings used by Balony are suitablefor analyzing most plates, the ability to manually adjustimage analysis parameters allows users to quantify eventhe most troublesome images.As a demonstration of the flexibility of our image analysisengine, we were able to use Balony to successfully quantifythe example image plates provided with both ScreenMill[3] and SGAtools [4], while neither of these packageswere able to analyze the sample images provided by anyof the other programs.We also sought to design a program that would enablethe complete analysis of a screen, from scanned imagesof plates to an interactive display of genes of interest,all from a single interface. While both ScreenMill andSGAtools necessarily involve using external web servicesto carry out some or all portions of their data analysis,Balony operates as a single, stand-alone window makingit easy to switch between modules to monitor the effectsof adjusting settings. Although this software is primarilyaimed at analyzing high-throughput experiments in yeast,it could also be employed for use with any system thatutilizes high-density arrays of microbial colonies.ImplementationBalony is a stand-alone Java program, which uses librariesfrom various sources, most notably the ImageJ libraryfor image manipulation [5], and The Apache CommonsMathematics Libraries for statistical analysis. The programhas a modular structure, shown in Figure 1. Data filesare generated at each stage of the analysis and can beinspected at will. If a user so chooses, they can merelyuse parts of the Balony package to measure colonysizes and perhaps perform normalization, and then usetheir own scripts or programs to further score their data.The data flow starts with composite images of multipleplates or single images of individual plates. In the caseof composite images, the “Scan” module converts theseto images of single plates. This can speed up the imageacquisition stage of the analysis by allowing the userto capture images of up to four plates at a time. Theseimages are then analyzed by the “Image” module whichproduces a text file containing raw data listing the pixelarea of each colony in each image. These text files arethen used as inputs for the “Scoring” module whichpairs control and experimental data sets and performsnormalization to produce one or more tab-delimitedtext files containing the normalized colony sizes for theexperiment. This data can then be viewed directly (e.g.by loading into a spreadsheet) or analyzed using the“Analysis” module. This enables collation of multiplesets of data and further refinement, e.g. by removal ofgenes linked to the query gene in an SGA experiment.Cut-off values to determine “hits”, p-value thresholdsand reproducibility across data sets can also be definedto precisely determine “hit lists” of genes. The Analysismodule enables the direct inspection of individual datapoints, providing gene information from the SaccharomycesGenome Database (SGD) [6]. The main window of theprogram is divided into five tabs which are used tosequentially analyze data (Figure 2A).Image segmentation: the scan tabThe “Scan” section of Balony allows users to take com-posite images of multiple plates and subdivide them intoseparate images for analysis (Figure 2B). We find thatimages of plates are best captured using a flatbed scanneras the reduced depth of field of a scanner compared toa digital camera results in less optical distortion of theimages. It is advisable to scan plates with a black back-ground (e.g. card or cloth) to improve contrast betweenthe colonies and the agar.We find that a final resolution of 300 dots per inch(dpi) is sufficient for most applications, although forultra-high density experiments using arrays with 6144colonies per plate (cpp), higher resolutions may be required.In general, processing time increases with image resolution,and the extra information above 300 dpi is unlikely toprovide more robust data as the inherent variance inthe size of yeast colonies will be more significant thanany additional fine detail gained.When performing SGA experiments we use a varietyof terms to describe the components of an experiment.Each array consists of a number of agar plates. For example,the haploid yeast Deletion Mutant Array (DMA) consists offour plates when arrayed at a density of 1536 cpp, which wewould simply refer to as plates 1-4. Replicates of each plateare termed sets. Thus in a typical SGA experiment, oneYoung and Loewen BMC Bioinformatics 2013, 14:354 Page 2 of 17http://www.biomedcentral.com/1471-2105/14/354Composite imageScan ModuleAgar plates (1-4) Flatbed scannerImage ModuleSingle plate imagesRaw pixel area dataJPEG filesTab-delimited text files(with meta-info)Scoring ModuleAnalysis ModuleTab-delimited text files(with headers)Normalized relative colony area dataData tableInteractive scatter-plot:SVG and PNG imagesIndividual colony informationTab-delimited text files andXLS MS-Excel worksheetsFigure 1 (See legend on next page.)Young and Loewen BMC Bioinformatics 2013, 14:354 Page 3 of 17http://www.biomedcentral.com/1471-2105/14/354Figure 2 The balony program. A. Screenshot of the graphical user interface, in this case showing the Imaging module. B. A composite imageof four plates demonstrating how it would be divided into separate images. C. A portion of an inverted, thresholded image overlayed with thearray grid in green. D. A portion of the same image showing successfully quantified colonies. E. Analysis module.(See figure on previous page.)Figure 1 Data-flow through Balony. Images are acquired using a flatbed scanner, either singly or in composites of up to four plates; in thelatter case these images are segmented into single images. The imaging module measures raw pixel areas and saves this information as text files(one file per plate). These files are collated by the scoring module and normalized; the data saved as tab-delimited text files; either one per set oras an averaged file. The analysis module facilitates interrogation of the screen data. In addition to saving the quantitative data in formats for usein spreadsheet applications, images of the plot of colony area ratios can be saved in bitmap and structured variants.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 4 of 17http://www.biomedcentral.com/1471-2105/14/354might produce three replicates of a query strain crossedwith the DMA, i.e. 3 sets of 4 plates.To ease downstream processing of images it is importantthat files are named systematically (Table 1). Input filesshould be named according to the template “screen_setx.jpg”, which will result in output files with the format“screen_set-x_plate-y.jpg”. If there are more than fourplates per set, an offset value can be included in the filename which will be added to each plate number. Forexample, a file named “yfg1_set1_[4].jpg” will by defaultnumber plates starting at 5 (i.e. 4 + 1). However, platescan also be scanned individually, in which case each setof plates should be named in the format “screen_setx_-platey.jpg”.Plates can be rotated automatically to ensure thecorrect orientation with respect to gene index files asthe subsequent analysis steps assume that the top leftposition of a plate is identified as “row 1, column 1”.Images can also be resized and, if they are not already,converted to greyscale. By default, the individual plateimage files are named according to a scheme to enabletracking of the plate and set number, which is recom-mended to ensure easy identification of plates in theColony Measurement section (see below). However,the “base name” and each individual plate name canbe overridden if necessary.An entire folder of composite images can be analyzedin batch mode to reduce the amount of user input requiredat this stage. Processing time is dependent on a numberof factors, such as the resolution of the input imagesand the speed of the computer, but should not takemore than a few minutes for a typical screen.The user can define different mappings which definehow the position of a plate in the scanned image isconverted into a plate number based on an array index.The default setting is shown in Figure 2B and is ideallysuited to the deletion collection maintained at 1536cpp format as this consists of exactly four plates. It isimportant to remember that a scanned image will usuallybe reflected about its y-axis, so it is advisable that thefirst time a user uses this module that they visually checkthat the plates are correctly assigned and oriented in orderto prevent mishaps further downstream. If the mappingneeds to be changed from the default setting, the newassignments will be remembered for future analyses.Colony measurement: the image tabBalony uses a multi-step process to measure colony sizeson individual plate images. Each step can be customizedwith varying parameters which enables a high degree ofcompatibility with plates from a variety of sources. Themeasurement process identifies colonies as elliptical objects,measures the pixel area of each object, and assigns theobject to a grid position. The raw data (grid row, gridcolumn and colony area) are saved for subsequentnormalization, scoring and analysis. This process canbe automated completely, requiring little to no userinput, but if this approach is not proving fruitful, eachplate can be analyzed manually using a variety of toolsto give fine-grained control of the analysis process.This panel shows a list of image files in the currentlyselected folder. A colour-coding and suffixing schemeis used to help the user identify which files have beenanalyzed. If a file has not yet been analyzed, then thefilename will simply be displayed in black. If a file hasbeen successfully analyzed, then it will be displayed ingreen and suffixed with [*]. If a file has been analyzed,but with a certain number of colonies that could not becorrectly measured (“bad spots”) then it will be displayedin orange and suffixed with [?], as long as the number ofbad spots is below a threshold value. This indicates thata plate has a small number of imperfections that maywarrant closer inspection. If the number of bad spots isabove the threshold value the file name will be displayedin red and suffixed with [!]. This is usually indicative of aplate which had significant defects, or the program wasunable to analyze automatically. The steps necessary toanalyze an image are described below.Format correctionThe image measurement process requires yeast coloniesas black ellipses on a white background. Upon opening,images will be converted automatically to greyscale if theyare not already in this colour format. As the colonies areusually lighter than the background, the image then needsto be inverted. This can be carried out automatically foreach plate and is recommended.Upon loading an image, Balony will attempt to decipherthe file name to determine a name for that particularexperimental set as well as the corresponding set andplate number. This unique name will be the same forall plates across sets for an individual experiment and isgenerated by stripping away the “set” and “plate” parts ofthe image file name. It is important that this informationis consistent between all plates from a given experiment asTable 1 File name handling in balonyInput file name Output file namesYfg1_G418_set-1.jpg Ygf1_G418_set-1_plate-1.jpgYgf1_G418_set-1_plate-2.jpgYgf1_G418_set-1_plate-3.jpgYgf1_G418_set-1_plate-4.jpgYfg2_URA_set-3_[4].jpg Yfg2_URA_set-3_plate-5.jpgYfg2_URA_set-3_plate-6.jpgYfg2_URA_set-3_plate-7.jpgYfg2_URA_set-3_plate-8.jpgYoung and Loewen BMC Bioinformatics 2013, 14:354 Page 5 of 17http://www.biomedcentral.com/1471-2105/14/354the information is written into the meta-data saved aftereach image is analyzed and is used by the scoring moduleto identify plates of the same experiment. However, thisoption can be disabled if so required.ThresholdingImages are converted to black-and-white using a procedureknown as “thresholding”, which separates the agar platebackground from the yeast colonies. In this processeach pixel in the image is converted to either black orwhite depending on whether it falls below or above adefined grey level. The images at this stage are storedat a colour depth of eight bits per pixel, so the greylevel will have a value of between 0 and 255. As inother programs, Balony can automatically define thisthreshold level using an algorithm based on a digitalhistogram of the image. Additionally, we have providedan option for the user to manually specify this greylevel which can salvage the analysis of an image whichwould otherwise fail if automatic thresholding is notsuccessful. Generally, if the thresholding is not successful,it is because the grey level selected is too high, resulting inthe plate background merging with the yeast colonies.There is an option to automatically attempt to re-analyzethe image with decreasing threshold values until the plateis successfully analyzed. However, care must be taken withthis as if too low a value is used then there is a danger ofdiscarding colony size information.GriddingArrays of yeast are indexed by identifying each strainin terms of its row and column position in a grid, andoptionally, a plate number. Therefore, a gridding step isrequired to identify the region of the image that containsthe arrayed colonies. Balony contains extensive controlsnot found in other programs to assist in the correctplacement of the grid, so that even if an array containsunusual features that may make automatic grid identifi-cation difficult (such as blank rows or columns, or verysmall colonies) manual intervention can resolve this.The software is supplied with a number of grid presetscorresponding to the most commonly used formats inuse, namely 96, 384 and 1536 cpp. New presets can bedefined and calibrated from sample plates. Balony canusually automatically determine the grid position using aparticle identifying routine; however, if a plate is provingproblematic, the user can manually specify the positionof the grid.The first time a user runs the program and loads animage, they will be prompted to either use an existingpreset, or define a new one that matches their particularimage acquisition platform. The latter option is recom-mended as it allows for a more precise fit for differenttypes of imaging hardware. The user will be prompted toenter the dimensions of the array (i.e. the number ofrows and columns), and then draw a box on the plateimage to indicate the boundaries of the array. Followingthis they will be prompted to name and save the settingsderived from this. The most important values from thisare the mean spacing (in pixels) between colonies in thehorizontal (dx) and vertical (dy) dimensions.Should a user not wish to use the automatic griddingprocess, or find it ineffective for their plates, they canindicate the position of the array manually, either bydragging a box from one corner to an opposite diagonalcorner; or by positioning a grid of fixed dimensions.After the gridding process has been completed, a gridis drawn over the thresholded image in green (Figure 2C).If required, this grid can be moved into a different positionusing the cursor keys. The user can decide if the griddingprocess should proceed automatically after a plate hasbeen thresholded. This generally speeds up the quanti-fication process, but it may be advisable not to use itfor the first few images so that the user has an opportunityto observe the different steps involved.Colony assignment and measurementNext, colony sizes are measured by analyzing all particleswithin the grid array and mapping them to their nearest(row, column) position within the array. Parameterscan be set to ensure certain criteria are met for a particleto be identified as a colony, including minimum pixel size,circularity and deviation from the grid position. If a gridposition appears not to have a colony present in thatposition, the program will re-scan that position to lookfor the presence of an object. This is sometimes necessarybecause overgrown colonies can and merge with neigh-bouring colonies and no longer appear as a discrete entity.This process is usually sufficient to identify the colonieson a plate. However, if there are many grid positionsthat appear to contain something that does not satisfythe minimum criteria for a yeast colony, the softwarewill offer the option of a low-stringency pass that willattempt to quantify the amount of growth occurring ina position, regardless of its circularity. Care should betaken with this option as blemishes on the plate surfacemay then be counted as colonies.When this process is complete, the successfully quantifiedcolonies are shown by outlining each colony in green overthe original input image (Figure 2D). The user can togglebetween this final image and the gridded, thresholdedimage which can be useful to confirm that the griddingand thresholding processes were accurate.Upon satisfactory quantification of an image, a tab-delim-ited text file is saved which contains the area of each colony(in pixels) for each row and column position. The resultingdata file can be viewed from within the program if sodesired.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 6 of 17http://www.biomedcentral.com/1471-2105/14/354The colony measurement process can be performedin batch using a set of default parameters to process anentire folder of images. After doing this, a log file isgenerated which can be inspected to review any problemsthat occurred during analysis. On a typical test set up, a1536 cpp plate scanned at 300 dpi takes approximatelythree seconds to analyze.There are some additional features to aid with prob-lematic images. If an image requires rotation, this can beachieved manually by a process which uses the positionsof two colonies within the same row to determine theappropriate angle required to correct the orientation ofthe plate.Any existing quantitation of a colony can be manuallyoverridden by drawing an ellipse over the colony. Theimage can be zoomed to help with this. Additionally, ifthe particle finding algorithm rejected a particular gridlocation, it will be highlighted as a red square to drawattention to a position that may require this manualintervention. Colonies that have been manually definedare highlighted in magenta to differentiate them fromthose automatically quantified by the program.Data scoring: scoring tabBy default the scoring module will search the folderlast used by the Image module to load files, although adifferent folder can be selected. The saved quantitationdata files are analyzed to find sets of data correspondingto experimental plates. The user can then select whichsets will comprise the control and experimental dataand load them by selection from a drop-down menu.Upon loading, the software will normalize each plate ofdata and align the corresponding control and experimentaldata to produce paired sets.The scored data can be saved in a variety of ways,listed either by colony grid position (recommended as itmakes it easier to track specific colonies) or by the nameof the ORF pertinent to each strain. In the case of multiplereplicates of the same ORF, the mean area will be saved,along with the standard deviation. Additionally the usercan select between two methods for saving experimentscomprising multiple sets of plates. The recommendedoption is to save a separate file for each paired set ofdata as this retains the most information on individualcolony sizes. However, it is also possible to combinemultiple data sets, in which case the mean area andstandard deviation will also be saved.The scored data files also specify the gene at eachposition in the array. This requires a key file that mapsthe position of each colony in the array to a yeast ORF.The format of the key file is a tab delimited text filewith four or five columns, with each row containingthe following data: Column 1: plate number Column 2: row number Column 3: column number Column 4: systematic ORF name (e.g. “YLL040C”) Column 5 (optional): standard gene name(e.g. “VPS13”)An example key file, “UBC-1536.key” is included withthe program. If the gene name is not specified the softwarewill attempt to determine this from the file “SGD_fea-tures.tab” which is found in the same folder as theprogram files. This file can be updated with the latestinformation from SGD from within the analysis module(see below). If the file is over 100 days old, the programwill prompt to download a new version. The “RefreshData” button will force the program to reload and scorethe selected control and experimental data files.Screen analysisThe final component of Balony is the Analysis module(Figure 2E). This enables the scored data from an experi-ment to be interrogated to identify positive and negativegenetic interactions. The analysis module requires thateach paired data set (control and experiment) is saved asa separate file as this ensures that quantitative data issaved for each individual colony. This is necessary forstatistical analysis of colony sizes.Users can elect to open all or a limited sub-set ofscored data sets, which may be useful on occasions ifthere was a suspected technical problem with the platesof a particular set. After selecting files to load, the useris presented with a new window showing a table of thescored data. The data table will show averaged data foreach array position along with the systematic ORF nameand the standard gene name, the mean and standarddeviation of the sizes of the control and experimentalcolonies at this position. The ratio (normalized experimen-tal colony size/normalized control colony size) is displayedas an indication of the extent of any genetic interaction;this is followed by the number of replicates in whichthe ratio is either below or above a cut-off value.The difference in colony sizes (normalized experimentalcolony size minus normalized control colony size) is alsoshown, which is analogous to the standard multiplicativescore used in other protocols [7]. This is followed by thep-value obtained by performing a paired two-tailed t-testtesting for a difference between the normalized colonysizes of the experimental strains vs. the normalized colonysizes of the control strains. Finally, whether this position isdeemed to be a “hit” (see below) is indicated, followedby a column that will state if a position should be excludedfrom the analysis. These data can be sorted by any ofthese criteria.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 7 of 17http://www.biomedcentral.com/1471-2105/14/354In the case of SGA screens it is useful to define genelinkage at this stage and remove these genes from theanalysis. This is accessed from the “Filtering” panelwhich opens a separate window displaying a graphicalrepresentation of the colony sizes of double mutants ingenes that flank the query gene. Because of the reducedrate of recombination between linked genes (e.g. thequery gene and an adjacent gene deletion), fewer doublemutant cells are generated compared to unlinked genes,which leads to a characteristic decrease in colony size forthese flanking linked genes [8]. The users then specifiesthe range of genes which are to be excluded from theanalysis to prevent them from being reported as falsepositives. In addition, genes can be excluded manuallyby the user if they are known to be problematic or falsepositives for some other reason. Genes that have beenexcluded from the analysis can be shown or hiddenusing a toggle switch.Analyses may benefit from discarding data where thecontrol strain has very poor growth. In these cases itcan be difficult to be sure a genuine synthetic lethalinteraction is being observed when the growth of thecontrol strain is particularly slow. This filtering can bebased on the growth of the control strain, the experimentalstrain or both strains; this allows for flexibility if theread-out of a screen is something other than syntheticlethality. If the data is to be discarded, the size of bothcolonies for that paired set is set to zero to excludethem from further analysis.In addition, maximal and minimal values may beassigned to colony sizes. This is useful if the ratio betweencolony sizes is being used as a measure of fitness. In thecase of a strong aggravating (synthetic lethal) interaction,the double mutant may be essentially dead. However,there will still be a small amount of yeast present onthe plate from the original pinning step. As this amountwill vary between different colonies, this can lead to theimpression that one synthetic lethal interaction is strongerthan another, whereas in fact they are both reporting thesame phenotype, i.e. no growth in the double mutant. Byassigning a minimum value, all truly lethal interactionswill have a similar score.Following this, the user may wish to define cut-off valuesto define “hits”, which in the case of SGA experiments,are genetic interactions. While the difference betweenexperimental and control colony size is often used as ameasure for the strength of a genetic interaction, wehave found that calculating the ratio of experimental tocontrol colony size is a useful alternative. We find thatthis parameter is less influenced by the growth rate ofcontrol strains. For example, consider two strains Δxand Δy which grow with normalized colony sizes of1.0 (i.e. as wild-type) and 0.4 respectively. If a secondmutation is introduced so that the double mutantstrain ΔxΔz grows to a colony size of 0.8, and theΔyΔz strain grows to 0.2, it is clear that this mutationhas had a greater effect on the Δy strain than Δx as ithas led to a halving of the growth rate of the Δy strain.However, if we were to report the difference betweencolony sizes, then both double mutant strains wouldyield a difference value of 0.2. In contrast, by using aratio score, we find values of 0.8 and 0.5 for ΔxΔz andΔyΔz, respectively, reflecting the relative strength ofthe observed genetic interactions.Using the analysis function, two types of hits aredistinguished between; those where the ratio is belowa cut-off value (aggravating) and those where the ratiois above a second cut-off value (alleviating). If there isno genetic interaction, the ratio will be close to 1. Aplot showing the distribution of ratios can be displayedto aid in screen validation. When first loaded, data filesare sorted by array position, so that the distribution ofratios can be inspected to check for any systematiceffects. Normally, it is expected that the variance betweencolonies would be distributed randomly, so if any trendsare apparent, it is indicative of a systematic effect fromeither the pinning or imaging process. When arrangedby ascending ratio this plot forms a characteristic curvewith a steep initial portion representing aggravatinginteractions which levels off to a portion with a shallowgradient indicating no significant interactions, and thenonce again returns to a steep portion representing alleviatinginteractions. The software can estimate an appropriatecut-off value by extrapolating the linear central portionof the distribution and finding the y-intercepts at eitherend of the x-axis. Once these values have been determined,the data table will highlight “hits” in the screen based oncriteria chosen by the user. The three criteria for a hit are:1. The ratio is below the low cut-off value (p) or abovethe high cut-off value (q) as described above.2. The number of replicates in which criterion (1) ismet must be equal to or above a specified value(e.g. 3 replicates out of 3).3. The p-value from a paired t-test of the sizes of theexperimental colonies and the control colonies mustbe below a given value.If all three criteria are met, then hits are highlighted inthe table (green for those less than p, red for thosegreater than q). The table can be sorted to list these hitsfirst, sorted by ratio from strongest to weakest.From here, many users will find it useful to merelybrowse the list of genes. To aid in this, more detailedinformation can be readily accessed from the table. Foreach array position, a pop-up window can be displayedgiving more detailed information on the corresponding.The normalized colony area will be shown for eachYoung and Loewen BMC Bioinformatics 2013, 14:354 Page 8 of 17http://www.biomedcentral.com/1471-2105/14/354individual colony, both numerically in a table; and as agraphical representation showing the currently highlightedcontrol/experiment pair (as two solid circles) and theindividual areas of all control and experimental colonies(as superimposed concentric circles). The description ofthe gene as defined by the SGD database is shown andthis information can be kept up to date by downloadingthe latest database file from within the program. If theratio plot is open, then the currently viewed query canbe viewed on this plot to give an overview in the contextof the entire screen. This window also contains a linkto the corresponding page for the ORF in question onthe SGD web site. The user can also quickly switch topositions containing duplicates of the current ORFelsewhere within the array, or to any other ORF ofinterest if present.The context menu can be used from the table to selectgenes of interest to copy (either as ORFs, gene names,or both) for use in other applications or web sites. Forexample, a list of ORFs can be pasted into the geneontology analysis utility at the SGD web site (http://www.yeastgenome.org/cgi-bin/GO/goTermFinder.pl).Additionally, the entire data table can be exported ei-ther as tab-delimited text, or as Microsoft Excel .xlsfile. This latter option prevents some formatting errorsthat can occur when importing tab-delimited text filesinto Excel, such as the interpretation of gene names asdates.Users can also filter the list of genes displayed, basedon a text string. Only those genes whose descriptioncontains this string will be displayed. This provides aquick way to check for interactions between genesinvolved in a particular function or process.OptionsA fifth panel provides for setting of some basic optionsfor operation of the program. This includes the ability tochoose the type of user interface offered and a simpleprocedure to automatically update Balony.Results and discussionColony size measurementTo speed up the quantification of plates with Balony,an automated gridding step can be used which attemptsto automatically locate the position of the array, using aparticle analysis routine to identify objects on the platethat resemble yeast colonies. To avoid counting extra-neous plate features (e.g. bubbles, off-grid contami-nants) as colonies, only a limited rectangular portion ofthe plate is scanned at one time. This region is based onthe expected dimensions of the array from the grid preset.This routine generates a list of objects with defined x and ycoordinates. The program assigns the objects closest to thecorners of this region as the corners of the array, and theninterrogates the spacing of all the other objects to seeif they fit the criteria necessary to be colonies withinthe array. Specifically, if the x and y coordinates areboth within 30% of an expected grid position, they areadded to a list of valid colonies. After analyzing allthe objects, the mean horizontal and vertical devi-ation of objects from their expected position is calcu-lated and if each of these is within 5% of theirexpected values, then it is assumed that the array hasbeen correctly established. If parameters are not deter-mined from the initial analysis, the rectangle is progres-sively repositioned until parameters are correctlyestablished. Occasionally, failure to automatically locatethe grid may be due to a plate not being placedsquarely in the scanner. The program contains an op-tion to try a number of rotations to correct for this ifthe gridding process fails. This attempts to repeat thegridding process after rotating the plate by up to 3° in0.5° increments. The current version of the algorithm wasarrived at by repeated refinement using hundreds of testplates and we find it to be effective in for virtually allimages we have encountered.Once the grid has been successfully established, thesizes of colonies can be measured. This uses the sameparticle analyzer routine as used in the gridding. Thisgenerates a list of objects with x and y coordinates,areas and circularity values. The program iterates throughthis list, testing each object. A number of criteria mustbe successfully met for an object to be identified as acolony.First, the centre of an object must be close to the centreof a grid position (by default both x and y coordinatesmust be within ¼ of a grid cell length, but this can bechanged). Second, the colony must be within certain sizelimits, which again, can be specified. Finally, the colonymust meet a minimum value for circularity, a parameterdetermined by the algorithm with possible values between0 and 1, where 1 represents a perfectly circular object. Thedefault minimum circularity value is 0.8. If more than oneobject is potentially allocated to the same grid position,the software will select the colony which is closest to thecentre of the cell.After interrogation of this list, the program individu-ally analyzes the pixel content of any grid positions thatdid not have a colony allocated. If any position in thegrid exceeds a minimum pixel count–suggesting the pos-sibility of a colony that was not detected in the first pass–then the user is presented with the option to perform alow stringency second pass. In this case, the circularity ofthe particle finding algorithm is set to zero in order toidentify “non-ideal” colonies; we find this helps to identifycolonies that have, for example, become smeared duringthe pinning process.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 9 of 17http://www.biomedcentral.com/1471-2105/14/354When we compared the ability of our software tomeasure colony sizes with ScreenMill and SGAtools, wefound near-identical results. We analyzed a 1536-colonyexample plate provided with each program, and foundthe correlation between measured colony sizes gave aR2 values of 0.9996 and 0.9959 respectively (Figure 3A,B), indicating that our implementation of colony meas-uring algorithms is similarly effective.Colony normalizationTo demonstrate the fidelity of the normalization proce-dures employed by Balony, we constructed a test arrayplate of 1536 colonies of the same wild type yeast strainand scored colony sizes for eight replicates of this testarray. First, we quantified growth of all 1536 wild typecolonies across the eight plates, which should all havethe same fitness. The mean pixel area per colony wasR² = 0.999601002003004000 100 200 300 400ColonyArea(pixels)-ScreenMillColony Area (pixels) - BalonyAR² = 0.9959050010001500200025003000350040000 1000 2000 3000 4000ColonyArea(pixel s)-SGAto olsColony Area (pixels) - BalonyBCD050100150200250300Corner Edge Interior Centre Plate meanMe anPixelA rea00. Edge Interior Centre Plate meanNormalizedPixelAreaFigure 3 Validation of colony measurement algorithm. A. A plate image provided with ScreenMill was analyzed with Balony and ScreenMill andthe measured size for each colony was plotted. B. As A, but comparing an image provided with SGATools. C. Eight YPD plates each containing 1536colonies of wild type yeast (strain BY4741) were created and colony sizes measured by Balony. Colonies representing different positions on the platewere compared along with the mean colony size of each plate. Error bars show standard error. D. The same data as in C, but following normalizationand “Row/Column” correction.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 10 of 17http://www.biomedcentral.com/1471-2105/14/354determined to be 138.22 ± 1.52 pixels (Figure 3C). How-ever, as previously described [9], we found that coloniesat the corners and edges of the plate grew more quicklythan colonies in the interior and at the centre (Figure 3C).This increased growth on the edges of the array is due todecreased competition with neighbouring colonies [9].Thus, the non-normalized mean colony size measurementwas not a reliable measure of colony fitness because it didnot take into account systematic effects of array positionon colony growth rate. In addition to the effects of arrayposition, growth conditions can also vary substantiallyfrom plate to plate, further confounding attempts toquantify the fitness of individual colonies [9]. This wasnoticeable in our test array in which colonies in cornerpositions showed a wider range of areas, with a variationof up to 92 pixels across the eight replicates. In contrast,colonies in the centre of the array varied by only 17pixels.To circumvent these problems, we developed colonynormalization algorithms which correct for decreasedcompetition resulting from array position and for variabilitybetween plates. Colony normalization is an essential featureof all protocols developed for analysis of colony-basedgrowth assays [9,10]. The first step in the normalizationprocedure employed by Balony is to divide the pixelarea of each colony by the median colony size on eachplate. Hence, a colony that grows near the average ratefor that plate will have a normalized area of ~1. Thereare then three optional correction procedures that canbe applied to the data, as described previously [9].First, “Row/Column” correction can be applied. As weobserved in our test array, the colonies at the edges ofplates will often grow faster due to decreased competitionwith other colonies. To compensate for this, a correctionfactor can be applied based on the deviation of a givenrow or column compared to all other rows or columns.This is achieved by calculating the median pixel area ofthe spots in a particular row or column. If this value isgreater than 1, then each spot in this row or column isdivided by the median value to normalize that row orcolumn with respect to the rest of the array.The next type of correction is “Spatial” correction. Thiscan be necessary in plates where the thickness of themedia is variable because yeast colonies will grow atvarying rates depending on the thickness of the media.To account for this, we take the median colony size ofeach row and column and fit these to a smoothed dis-tribution using a LOESS algorithm [11]. This generatesa pair of distributions, corresponding to the horizontaland vertical axes of the plate, with each value in thedistribution expressed relative to the median colonysize on the plate. From these distributions we candetermine a correction factor for each position in thearray, as the product of the corresponding row andcolumn positions in each of the horizontal and verticalLOESS distributions.The final type of correction employed by Balony is“Competition”. This can be necessary when a colony hasa number of slow-growing colonies surrounding it. Dueto this reduced competition for nutrients, that colonymay then grow faster [9]. To control for this, we examinethe whole plate for colonies whose eight surroundingneighbours have a mean growth rate of <75% of themedian colony size. Using a simple linear regressionanalysis, we can determine if colony size correlates withthe size of surrounding colonies on a given plate. If thiscorrelation proves sufficiently robust (R2 > 0.1, slopebetween 0.1 and−1), then any colonies on the platewhich have reduced competition (again, neighbourswith a mean growth rate of <75% of the median) arecorrected by applying the parameters derived from thelinear regression to its colony size.To confirm the effectiveness of colony normalizationand Row/Column correction, we applied this algorithmto the raw pixel area data from the test array of wildtype yeast shown in Figure 3C. While the raw pixel areasshowed variations in colony size as high as 67% greaterthan the plate mean (for a corner colony), followingRow/Column correction, each of the representativecolonies reported a growth rate within 3% of the platemean (Figure 3D). The low standard error associatedwith the mean corrected value for the plate (<5%) indicatesthat this algorithm effectively deals with variationscaused by growth at the edges of plates. Thus, followingnormalization and Row/Column correction, we wereable to determine with high accuracy that all coloniesin the array had similar fitness, as would be expectedsince they are genetically identical.Now we compared the effects of each normalizationalgorithm using an actual 1536-density array plate of yeastsingle deletion mutants routinely used for SGA analysisin our lab. The effects of sequentially applying each typeof correction are shown in Figure 4 (A-E). Applying Row/Column correction had the expected effect of normalizingthe sizes of colonies in the outer rows and columns(compare Figure 4B & C). Subsequently applying Spatialand Competition correction resulted in much less dramaticcorrections, likely because of the fairly uniform growthrates of the individual deletion mutants in the array(Figure 4C). For this reason we suggest that users onlyneed to apply Row/Column correction unless they feelthat their images would specifically benefit from theadditional steps, such as in the case of unevenly pouredplates, or with arrays containing a large number ofslow-growing strains or empty spaces. As each correctionstep has the potential to distort the original data, wefeel that it is beneficial to minimize the number ofpost-processing steps where possible.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 11 of 17http://www.biomedcentral.com/1471-2105/14/354Genetic interactions of SCS2To demonstrate the utility of this software, we performedan SGA experiment using a strain deleted for the geneSCS2. The analysis steps are described in a more detailed,step-by-step tutorial online at http://code.google.com/p/balony/wiki/Tutorial1 where a link to the scanned imagesis available should a user wish to follow the stages ofanalyzing a typical screen from start to finish.The experimental approach is outlined in Figure 5.All robotic pinning steps were performed using a Singer0 1 2CBAD EColony SizeFigure 4 Normalization of quantified colony measurements. A. Scanned image of one of three replicates used in B-E. The shading of each cellrepresents the corrected colony size as indicated by the color key and each heatmap shows the mean of three biological replicates. B. Heatmap ofuncorrected colony sizes. C. Heatmap after applying row/column correction. D. As C, but with spatial correction also applied. E. As D, but withcompetition correction also applied.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 12 of 17http://www.biomedcentral.com/1471-2105/14/354Query: Y7092, Δscs2::URA3(1 x 1536)DMA(4 x 1536)SC-UraYPD+G418Mating(4 x 1536)YPDDiploid Selection 1(4 x 1536)SC-Ura+G418Diploid Selection 2(4 x 1536)YPD-G418Sporulation(12 x 1536)EnrichedSporulation“HRK”DMA Selection 2(12 x 1536)DMA Selection 1(12 x 1536)“HRK”+G418 “HRK”+5-FOA+G418 “HURK”+G418 “HURK”+G418Double Mutant Selection 1(12 x 1536)Double Mutant Selection 2(12 x 1536)MATa Selection(12 x 1536)Figure 5 Experimental procedure for an SGA screen for SCS2. SC–synthetic complete medium; “HRK”–SC minus histidine/arginine/lysine, pluscanavnine (100 mg/L) and thialysine (100 mg/L); “HURK”–as HRK, but minus uracil; 5-FOA–5-fluoroorotic acid.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 13 of 17http://www.biomedcentral.com/1471-2105/14/354RoToR HDA robot with colonies arrayed at a density of1536 cpp. An scs2::URA3 strain was constructed in theY7092 strain background [2] and arrayed on SC-Uraplates. This was then mated with the DMA on YPDmedium and diploids selected for on SC-Ura mediumsupplemented with 200 mg/L G418. After sporulation,MATa haploid cells were germinated. We then generateda set of double mutants by two successive rounds ofincubation on medium lacking uracil. Simultaneously,we generated single mutant control (DMA) strains by firstincubating on medium containing 1 g/L 5-fluorooroticacid (5-FOA) to counter-select for strains containingthe scs2::URA3 allele, and then incubating on mediumcontaining uracil.Each set of double and single mutant plates wasscanned at 300 dpi and Balony was then used to analyzethe images using the default image settings. Eachpaired set of data was scored using median Row/Columncorrection. We compared over 5500 single mutant controlspots to the corresponding double mutant experimentalspots in three biological replicates. We were able toidentify significant differences in spot size for 638experimental spots using a maximum p value of 0.05(Figure 6A; red dots). Our ability to detect such a largenumber of potential interactions is a strong indicationof the robustness of our methodology and of the highfidelity of the Balony colony scoring and normalizationsystem. As expected, as the difference between controland experimental spot size approached zero it becameincreasingly difficult to define potential interactions withconfidence (Figure 6A). As has been reported for previousanalysis methods [10], we also identified a substantialnumber of experimental spots that exhibited only smalldifferences in size from the corresponding control spot,but which were measured with unusually high accuracy(Figure 6A), which is likely an artifact of using only a smallnumber of replicates (n = 3). However, these differenceswere unlikely to represent true genetic interactions becausethey fell within the 95% confidence interval of the meandifference measurements.Using the analysis module of Balony to interrogategenetic interactions, we examined the ratios of spot sizesand used the ratio plot window to automatically definethe ‘low’ and ‘high’ cut-off values for hits. The cut-offvalues obtained were 0.895 and 1.106, respectively. Usingthis method we were able to score 255 experimental spotsthat met these cut-off values in three out of threereplicates that also had a maximum p value of 0.05compared to the corresponding control spot size(Figure 6A; red dots within blue circles). This eliminated383 experimental spots that showed only small, butsignificant differences from the corresponding control spot.Due to genetic linkage, a total of 68 spots corresponding togenes neighbouring SCS2 were excluded from the analysis.Additionally, the URA1, URA2, URA4, URA5 and FUR4genes were excluded as these mutants are involved inuracil metabolism and generate false “hits” due to theuse of the–Ura selection media.Using these ratios and a maximum p-value of 0.05 forthe difference between control and experimental spotsizes, we identified 169 aggravating genetic interactionsand 98 alleviating interactions in three out of three bio-logical replicates. The list of genes responsible for theseinteractions was copied from the table and pasted intothe FunSpec web site at http://funspec.med.utoronto.ca,to test for enrichment of this hit list for various geneontology terms. We noted enrichment in a number ofcategories including “protein retention in Golgi apparatus”(p < 10-4), and “nuclear migration along microtubule”(p < 10-4) giving clues to the potential roles of the SCS2gene in the cell.To determine how our analysis compared with previousSGA analyses we downloaded the data for the SCS2 SGAscreen performed by the Boone Lab [7] as part of theirhigh throughput series of SGA experiments. To comparethe data sets we used the “diff” value for the hits identifiedusing Balony from our SCS2 screen and the “epsilon”value from the Costanzo data set (Figure 6B). These areapproximately equivalent measures of the strength of agenetic interaction as they compare the difference betweenthe normalized colony size of a yeast double mutant andthe corresponding single mutant control (see Conclusionsfor a more detailed definition). We found a high degreeof overlap, with both aggravating and alleviating geneticinteractions being found in both experiments. Of the48 genetic interactions that were common to bothscreens (Figure 6B), 39 were found to have the identicaleffect, with 32 aggravating interactions and 7 alleviating.Using this information we were able to calculate valuesestimating the sensitivity and precision of our method,as follows:The following formulae are used to define the parameters“precision” and “sensitivity” [7]:precision ¼ TPTP þ FP and sensitivity ¼TPTP þ FNwhere TP represents the number of “true positives” in thedata set (genuine interactions correctly identified), and FPrepresents the number of “false positives” (interactionsfalsely identified).Given that precision has been determined experimentallyfor the Boone lab data set at 0.63, and we know this paperreported 124 hits for this screen, which comprise a numberof true positives and a number of false positives: Therefore,TP + FP = 124, so FP = 124-TPYoung and Loewen BMC Bioinformatics 2013, 14:354 Page 14 of 17http://www.biomedcentral.com/1471-2105/14/354A00. -0.88 -0.66 -0.44 -0.22 0 0.22 0.44 0.66 0.88 1.10diff value1 - (p value)BThis study (diff value)Costanzo et al. (epsilon value) -0.88 -0.66 -0.44 -0.22 0 0.22 0.44 0.66 0.88 1.10diff value1 - (p value)CFigure 6 (See legend on next page.)Young and Loewen BMC Bioinformatics 2013, 14:354 Page 15 of 17http://www.biomedcentral.com/1471-2105/14/354precision ¼ TPTP þ 124−TPSo TP = 78 and FP = 46.As the sensitivity of the Boone lab set has been estimatedat 0.35, we can estimate the total number of geneticinteractions for SCS2 as 78/0.35 = 223. Our data setidentified 32 genetic interactions in common with theBoone lab data set and these are likely to be genuineinteractions; yet because the sensitivity of this data setis 0.35, this indicates that our data set contained atotal of 32/0.35 = 91 true positive interactions. So ofthe 169 interactions identified, there are 169-91 = 78false positives. The number of false negatives in ourdata set, i.e. interactions that we did not identify, musttherefore be 223-91 = 132.Applying the above formula, we were able to determineparameters for our screen which are summarized forcomparison purposes in Table 2 alongside the valuesobtained from the Boone lab data set. The sensitivity ofour screen compares well with that obtained by theBoone lab (0.41 vs. 0.35), as does the value we obtainedfor precision (0.53 vs. 0.63). This indicates that ourprotocol is sufficiently robust for routine laboratoryusage. We speculate that these differences are largelydue to two factors. First, the criteria used to distinguishhits are slightly different between the two methods,with our protocol relying on the ratio of colony sizes,with the Boone lab using the difference. Second, in ourprotocol we generate a control data set with eachexperimental data set, while the Boone lab uses astandard reference control set. These differences arelikely to impact on the relative rates of false positiveand false negative results obtained.We also compared the Balony analysis method to anavailable method that uses a Bayesian framework for theanalysis of biological data, which can be applied to anydataset that utilizes paired control and experimentalmeasurements, and is particularly effective when thereare only a limited number of replicates [12]. We used theCyber-T program to analyze the colony size data for theSCS2 SGA screen that was quantified and normalizedusing Balony, using the suggested parameters of anaveraging window equal to 101, a Bayesian confidencevalue equal to 10, and a minimum p value of 0.01 [12].We plotted the difference between control and ex-perimental spot size versus the Bayesian p values andhighlighted points with a p value below 0.01 (Figure 6C).As expected, this method increased the significancethreshold for small difference measurements comparedto t-test alone, eliminating spots with unusually lowstandard deviations due to having a small number ofreplicates. By this method 302 spots were identifiedthat corresponded to 239 potential genetic interactions.We compared the genes identified by this method to theBoone lab data set and identified 14 genetic interactions incommon, compared to 48 using Balony. Thus, making theassumption that the genetic interactions identified by theBoone lab were “gold standard” true genetic interactions,the algorithms employed by Balony appeared to beparticularly well suited for the analysis of genetic inter-action data derived from colony size measurements ofhigh density yeast arrays.ConclusionsIn this paper we have described a software package thatmakes the analysis of SGA data both rapid and flexible.We believe we have devised a complete system that canbe employed at a relatively low cost, and in many caseswill involve the purchase of no additional equipment. Ifnecessary, the components for a dedicated imaging andcomputational platform (scanner, computer, high-resolutionmonitor) could be purchased for less than $2,000. It isour intention to continue development of the programin response to the needs of the community and to releaseregular updates offering new features.Using the analysis features in Balony it is possible todetermine parameters similar to those published for largescale data sets. Specifically, our diff measurement is analo-gous to the epsilon value. The epsilon value is defined asthe difference between the observed growth of a doublemutant strain and the predicted growth of the strainbased on the relative fitness of each single mutantstrain according to the multiplicative model for geneticinteractions [13]. For example, in cases where the querystrain has no associated fitness defect, then the differencemeasurement determined using Balony is equivalent toTable 2 Estimates of parameters for balonyCostanzo et al. [7] Our dataTrue positives 78 91False positives 46 78False negatives 144 131Sensitivity 0.35 0.41Precision 0.63 0.54(See figure on previous page.)Figure 6 Summary of our SCS2 SGA experiment. A. Correlation of diff value with p value. Gray: p value > 0.05; red: p value < 0.05; circled redpoints: p < 0.05, diff above or below threshold in three replicates. B. Difference values in our screen plotted on the x-axis against epsilon valuesfrom the Boone lab data set on the y-axis. C. Correlation of diff value with p value using Bayesian analysis with Cyber-T. Gray: p value > 0.01;circled points: p value <0.01.Young and Loewen BMC Bioinformatics 2013, 14:354 Page 16 of 17http://www.biomedcentral.com/1471-2105/14/354the epsilon value. However, should a query strain beused which does not grow as wild type, the differencemeasurement can be easily corrected to account forthis. If a deletion strain is present in the DMA thatcorresponds to the query gene, then its growth can beused to approximate the growth rate of the querystrain; otherwise the growth rate must be determinedindependently.As a result it is possible for users of Balony to directlycompare their results with the large resource of geneticinteraction data already available. In this paper we haveshown an example of this, comparing our SCS2 SGAscreen with the data available in public databases. Theextensive correlations between the two data sets provideevidence that the analysis methods we have describedhere are sufficiently robust for routine analysis of geneticinteraction data.Availabilty and requirementsProject name: BalonyProject home page: http://code.google.com/p/balony/Operating system(s): Platform-independentProgramming language: JavaOther requirements: Java 1.6 or higher, >1GB freememory.Licence: GNU GPLAny restrictions to use by non-academics: noneThis site also hosts the source and a wiki which servesas a reference manual and contains a tutorial which guidesa user through the analysis of a sample screen. This shouldalso be consulted for details of system requirements andinstallation instructions.Competing interestsThe authors’ declare that they have no competing interests.Authors’ contributionsBPY designed and developed Balony and wrote the paper. CJRL providedvaluable input into the design of Balony and assisted with the programvalidation and writing the manuscript. Both authors read and approved thefinal version of the manuscript.AcknowledgementsWe are grateful to members of the Boone Lab, University of Toronto, inparticular Huiming Ding and Bilal Sheikh for providing us with access to the“Colony” programs which provided the inspiration for the design of Balony.This research was supported by grants from the Canadian Institute of HealthResearch (CIHR), the Michael Smith Foundation for Health Research (MSFHR),the Natural Sciences and Engineering Research Council of Canada (NSERC)and the Canada Foundation for Innovation (CFI). C.J.R.L. is a CIHR NewInvestigator, a MSFHR Scholar, and a Tula Foundation Investigator.Received: 11 July 2013 Accepted: 28 November 2013Published: 4 December 2013References1. Tong AH, Evangelista M, Parsons AB, Xu H, Bader GD, Page N, Robinson M,Raghibizadeh S, Hogue CW, Bussey H, et al: Systematic genetic analysiswith ordered arrays of yeast deletion mutants. Science 2001,294(5550):2364–2368.2. Tong AH, Boone C: Synthetic genetic array analysis in Saccharomycescerevisiae. Methods Mol Biol 2006, 313:171–192.3. Dittmar JC, Reid RJ, Rothstein R: ScreenMill: a freely available software suitefor growth measurement, analysis and visualization of high-throughputscreen data. BMC Bioinforma 2010, 11:353.4. Wagih O, Usaj M, Baryshnikova A, Vandersluis B, Kuzmin E, Costanzo M,Myers CL, Andrews BJ, Boone CM, Parts L: SGAtools: one-stop analysis andvisualization of array-based genetic interaction screens. Nucleic Acids Res2013. http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3692131/.5. Rasband WS: ImageJ, U S National Institutes of Health. Bethesda Maryland,USA. http://imagej.nih.gov/ij/ 1997-2011.6. Cherry JM, Adler C, Ball C, Chervitz SA, Dwight SS, Hester ET, Jia Y, Juvik G,Roe T, Schroeder M, et al: SGD: Saccharomyces Genome Database. NucleicAcids Res 1998, 26(1):73–79.7. Costanzo M, Baryshnikova A, Bellay J, Kim Y, Spear ED, Sevier CS, Ding H,Koh JL, Toufighi K, Mostafavi S, et al: The genetic landscape of a cell.Science 2010, 327(5964):425–431.8. Jorgensen P, Nelson B, Robinson MD, Chen Y, Andrews B, Tyers M, Boone C:High-resolution genetic mapping with ordered arrays of Saccharomycescerevisiae deletion mutants. Genetics 2002, 162(3):1091–1099.9. Baryshnikova A, Costanzo M, Kim Y, Ding H, Koh J, Toufighi K, Youn JY, Ou J,San Luis BJ, Bandyopadhyay S, et al: Quantitative analysis of fitness andgenetic interactions in yeast on a genome scale. Nature methods 2010,7(12):1017–1024.10. Collins SR, Schuldiner M, Krogan NJ, Weissman JS: A strategy for extractingand analyzing large-scale quantitative epistatic interaction data.Genome Biol 2006, 7(7):R63.11. Cleveland WS: Robust locally weighted regression and smoothingscatterplots. J Am Stat Assoc 1979, 74(368):829–836.12. Baldi P, Long AD: A Bayesian framework for the analysis of microarrayexpression data: regularized t -test and statistical inferences of genechanges. Bioinformatics 2001, 17(6):509–519.13. Boone C, Bussey H, Andrews BJ: Exploring genetic interactions andnetworks with yeast. Nature Rev Genet 2007, 8(6):437–449.doi:10.1186/1471-2105-14-354Cite this article as: Young and Loewen: Balony: a software package foranalysis of data generated by synthetic genetic array experiments. BMCBioinformatics 2013 14:354.Submit your next manuscript to BioMed Centraland take full advantage of: • Convenient online submission• Thorough peer review• No space constraints or color figure charges• Immediate publication on acceptance• Inclusion in PubMed, CAS, Scopus and Google Scholar• Research which is freely available for redistributionSubmit your manuscript at www.biomedcentral.com/submitYoung and Loewen BMC Bioinformatics 2013, 14:354 Page 17 of 17http://www.biomedcentral.com/1471-2105/14/354


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



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"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items