comparison of embryonic expression within multigene families...

a SPECIAL ISSUE RESEARCH ARTICLE

Comparison of Embryonic Expression WithinMultigene Families Using the FlyExpressDiscovery Platform Reveals More SpatialThan Temporal DivergenceCharlotte E. Konikoff,† Timothy L. Karr, Michael McCutchan, Stuart J. Newfeld, and Sudhir Kumar*

Background: Overlaps in spatial patterns of gene expression are frequently an initial clue to geneticinteractions during embryonic development. However, manual inspection of images requires considerabletime and resources impeding the discovery of important interactions because tens of thousands of imagesexist. The FlyExpress discovery platform was developed to facilitate data-driven comparative analysis ofexpression pattern images from Drosophila embryos. Results: An image-based search of the BDGP andFly-FISH datasets conducted in FlyExpress yields fewer but more precise results than text-based search-ing when the specific goal is to find genes with overlapping expression patterns. We also provide an exam-ple of a FlyExpress contribution to scientific discovery: an analysis of gene expression patterns formultigene family members revealed that spatial divergence is far more frequent than temporal diver-gence, especially after the maternal to zygotic transition. This discovery provides a new clue to molecularmechanisms whereby duplicated genes acquire novel functions. Conclusions: The application of FlyEx-press to understanding the process by which new genes acquire novel functions is just one of a myriad ofways in which it can contribute to our understanding of developmental and evolutionary biology. Thisresource has many other potential applications, limited only by the investigator’s imagination. Develop-mental Dynamics 241:150–160, 2012. VC 2011 Wiley Periodicals, Inc.

Key words: Drosophila; FlyExpress; gene expression patterns; image analysis; multigene family

Key findings:� FlyExpress resource facilitates analysis of expression pattern images from Drosophila embryos.� Image-based pattern searching yields fewer but more precise results than text-based searching.� Gene family expression more frequently displays spatial divergence than temporal divergence.

Accepted 30 August 2011

INTRODUCTION

High throughput and individual labo-ratory efforts have made available avast collection of spatial patterns for alarge number of developmentally rele-

vant genes (Tomancak et al., 2002;Lecuyer et al., 2007). These data canbe a key to the discovery of previouslyunknown links between genes andnew components within developmen-tal networks. A common first step in

discovering gene interactions is toidentify genes with overlappingexpression patterns. However, thestandard practice of manual inspec-tion of images is not efficient giventhe extraordinary number of images

Dev

elop

men

tal D

ynam

ics

Additional Supporting Information may be found in the online version of this article.

School of Life Sciences and Center for Evolutionary Medicine and Informatics in the Biodesign Institute, Arizona State University, Tempe,ArizonaGrant sponsor: NIH; Grant number: 2R01HG002516-07.†Dr. Konikoff ’s present address is Department of Biology, University of Washington, Seattle, WA 98195*Correspondence to: Sudhir Kumar, School of Life Sciences and Center for Evolutionary Medicine and Informatics in theBiodesign Institute, Arizona State University, Tempe, AZ 85287. E-mail: [email protected]

DOI 10.1002/dvdy.22749Published online 29 September 2011 in Wiley Online Library (wileyonlinelibrary.com).

DEVELOPMENTAL DYNAMICS 241:150–160, 2012

VC 2011 Wiley Periodicals, Inc.

available today (Gurunathan et al.,2004; Peng et al., 2007; Walter et al.,2010). This problem has beenaddressed using textual descriptionsof gene expression images using con-trolled vocabularies (CVs; Janning,1997; Brody, 1999; FlyBase, 1999;Drysdale, 2001; Tomancak et al.,2002; Matthews et al., 2005; Grum-bling and Strelets, 2006).

Textual descriptions are alwayslimited by the size and nature of thevocabulary and the current state ofscientific knowledge. Furthermore, notextual annotations exist for a pleth-ora of spatial expression patterns dueto the tremendous resources required,as assigning such annotations mustbe done manually (Gurunathan et al.,2004). Furthermore, under manycircumstances it is not feasible to fullydescribe an expression pattern inwords alone. There is a great needfor computational approaches thatdirectly compare the primary infor-mation—images containing expres-sion patterns (Kumar et al., 2002;Khatri and Draghici, 2005; Walter etal., 2010). These tools have the uniquepotential to facilitate large-scalesynthesis of all available developmen-tal image data and to integrateinformation across experiments andlaboratories.

For instance, computational frame-works for sequence analysis have al-ready revolutionized the analysis ofDNA and proteins (Altschul et al.,1990). However, image analysis ofgene expression patterns poses uniquechallenges. Sequence data can be eas-ily expressed in four or twenty build-ing blocks (for DNA and proteins,respectively), but images containingspatial expression patterns show re-markable variation in orientation,size, lighting conditions, and color-spectrum depending on the laboratorytechniques used. Thus, biologicalimage analysis is more challengingand a new frontier of computationalbiology (Gurunathan et al., 2004;Peng et al., 2007; Walter et al., 2010).

A current issue is how to facilitatethe large-scale analysis of geneexpression patterns from D. mela-nogaster embryos. Over 100,000images capturing spatial aspects ofgene expression are now available(Tomancak et al., 2002; Lecuyer et al.,2007). These images are a key

resource for understanding the earlydevelopment of fruit flies and othermetazoans because a large number ofgenes are shared (Koonin et al., 2004;Bier, 2005). Thus, we developed theFlyExpress platform, which containsa unique digital library of standar-dized expression patterns and toolsthat facilitate comparative expressionanalysis (http://www.flyexpress.net).

Here, we show that image-basedsearching in FlyExpress often yieldsfewer but more high-quality resultscompared with text-based searchingwhen the specific goal is to find co-expressed genes. We then demon-strate how the FlyExpress resourcemay be used for scientific discovery byconducting an analysis of multigenefamily expression divergence fromspatial and temporal perspectives.Overall, we found more instances ofspatial than temporal divergenceamong paralogous genes with compa-rable data. Looking more closelyat this diversity, the data revealedmost of this expression divergenceoccurs after the maternal to zygotictransition.

RESULTS

By virtue of having a unique digitallibrary of expression patterns that areuniformly oriented, aligned andscaled, FlyExpress makes it possiblefor investigators to visually and com-putationally compare gene expressionpatterns within the Berkeley Dro-sophila Genome Project (BDGP; Tom-ancak et al., 2002) and Fly-FISH(Lecuyer et al., 2007) image collec-tions. These high-throughput in situhybridization datasets contain imagesdepicting the expression patterns ofvarious D. melanogaster genes acrossdevelopment (Fig. 1A). There are1,091 genes shared by both datasets,2,263 unique to BDGP and 1,342unique to Fly-FISH.

There are a total of 57,045 BDGPimages (of individual whole-mountembryos) in FlyExpress. This datasetcaptures expression throughoutembryogenesis. There are 43,065images of whole-mount embryos fromFly-FISH available for searching inFlyExpress. The majority of theseimages capture embryonic develop-ment through stage 9. Also note thatthe in situ hybridization protocols

used by each are different—BDGPvisualizes gene expression in whole-mount embryos using the single coloralkaline phosphatase reaction andlight microscopy while Fly-FISH usestwo color fluorescence in situ hybrid-ization and confocal microscopy.Thus, BDGP images are valuable forassessing expression presence or ab-sence in various parts of the embryoacross development, whereas Fly-FISH images are valuable for a high-resolution examination of geneexpression during the earliest stagesof development. Currently, searchingfor Fly-FISH images using a BDGPquery (and vice versa) cannot be con-ducted in FlyExpress, as the imagesfrom each dataset are sufficiently dif-ferent as to be incompatible. However,both BDGP and Fly-FISH curatorsassign images to stage ranges andtextually annotate them with CVterms. Thus, FlyExpress allows one tocompare image and text-based searchtechniques within either dataset.Image-based analyses are facili-

tated in FlyExpress by the BasicExpression Search Tool for Images(BESTi; Kumar et al., 2002). For eachgene, BESTi uses spatial profilesderived from images capturing geneexpression by means of a series ofcomputational filtering and adaptive-thresholding techniques (see the Ex-perimental Procedures section). Thisaccommodates images captured undervarying microscopy and experimentalconditions. Three binary spatial pro-files per standardized image areincluded to develop discrete represen-tations that accommodate expressionintensity differences across theembryo (black and white spatial pro-files). FlyExpress users may chooseone of these three spatial profiles (a,b, or c) to conduct a search for over-lapping expression. For each standar-dized image, the b spatial profile rep-resents expression most closelydepicted in the image itself, while thea and c profiles represent under orover-staining, respectively.Just as each standardized image is

assigned a unique identification num-ber in FlyExpress, their respectivespatial profiles are also assigned aunique identifier. Inclusion of spatialprofiles allows FlyExpress users tofurther customize searches. To dis-cover genes that show expression

Dev

elop

men

tal D

ynam

ics

IMAGE ANALYSIS OF EMBRYONIC GENE EXPRESSION 151

Fig. 2. Image searching method comparison for BDGP data. A,B: Images searches for overlapping expression patterns for the same queryimage were performed for twi using FlyExpress (A; FBim9035159 FlyExpress profile c) and BDGP’s image search tool that disregards embryo ori-entation (B). The top nine images of 86 are shown for FlyExpress while all six images are shown for the BDGP method. C:The number of CV termsa given FlyExpress identified gene shares with the query gene are graphed. Examples of four images with � 50% overlap but sharing 0 CV termswith the query gene are shown below the graph.

Fig. 1.

Fig. 1. twist gene expression patterns andspatial similarity. A: Standardized BDGP (left)and Fly-FISH images (right) viewable in FlyEx-press depicting expression of twist (twi) areshown at progressively older developmentalstages from top to bottom in both columns. InBDGP images gene expression is blue and inFly-FISH images gene expression is green/yellow. B: Spatial similarity is expressed as anS-score and measured based on the commonpresence and absence of expression at corre-sponding coordinates in the spatial profiles ofquery and database images. Using the twiimage presented in (A), we show a sna spatialprofile and corresponding standardized imagethat display considerable spatial overlap andFlyExpress calculates the pair has a highS-score (S ¼ 0.57 reflecting 57% overlap). Forimages with no overlap in comparison to twi,such as Doc3, FlyExpress calculates an S-scorefor the pair of zero (S ¼ 0.0).

Dev

elop

men

tal D

ynam

ics

152 KONIKOFF ET AL.

patterns similar to that of a querygene, FlyExpress compares the spa-tial profiles of that gene with all theblack and white spatial profiles in thedatabase. These comparisons are re-stricted to images from the same datasource, anatomical view, and stagerange as the query pattern.

Spatial similarity is measuredbased on the common presence andabsence of expression at correspond-ing coordinates in the spatial profilesof query and database images. Forpatterns with larger overlaps, thesimilarity value (S-score) will behigher (Fig. 1B). The statistical signif-icance of the observed match isassessed using the probability of ran-domly finding a match occurringwith the given similarity or better(P-value). The distribution used toderive the P-value for a given S-scoreis generated from pairwise compari-sons of all patterns in the query data-base in a stage and view specific man-ner (see the Experimental Proceduressection and Supp. Fig. S2, which isavailable online, for details). Thismethod is analogous to that used inthe BLAST search for homologoussequences (Altschul et al., 1990).

Image vs. Text-Based

Searching

We began our analysis of the relativevalue of data-driven (image-based) vs.text-based searching by using thewell-known twist (twi) expression pat-tern depicted in Figure 1A. twistexpression is activated by the dorsal–ventral axis organizing gene Dorsal;its expression defines the ventral do-main of the embryo and its functiondictates that these cells become meso-derm (reviewed in Stathopoulos andLevine, 2002). FlyExpress searching(FBim9035159, profile c) produces 86genes with a P-value < 0.02 andS-score � 0.5. The six of the nine mostsimilar expression patterns found bysearching with the query image shownin Figure 2A are biologically meaning-ful. They include known twi targetgenes such as sna (Ip et al., 1992), htl(Stathopoulos et al., 2004) and brk(Markstein et al., 2004) and additionalgenes with roles in dorsal–ventral axisor mesoderm specification such asMes2 (Zimmermann et al., 2006), tkv(Dorfman and Shilo, 2001) and zfh-1

(de Velasco et al., 2006). Importantly,genes that are not currently charac-terized, such as CG12177, CG4221,and CG8664, also exhibit significantoverlap. These genes are now candi-dates for testing to determine theirinvolvement, if any, in the twidevelopmental network or dorsal–ven-tral axis formation.

We next compared FlyExpressimage-based search results for twiwith those of BDGP’s more compleximage-based search tool (Frise et al.,2010). In contrast to FlyExpress data-driven searching that discovers simi-lar and overlapping expression pat-terns simply based on the presence orabsence of expression at a given point,BDGP data-driven searching uses amore sophisticated method that alsoconsiders expression intensity (Friseet al., 2010). Using the same queryimage, BDGP yields a set of six genes,all of which exhibit nearly congruentgene expression with the query pat-tern (Fig. 2B). This is in contrast toFlyExpress, which returns a longerlist of genes (including the six foundby BDGP) that exhibit a wider rangeof spatial overlap with the query andeach other.

We then conducted a text-basedsearch using CV terms associatedwith the twi query image (trunk meso-derm primordium P2, head mesodermprimordium P4 and anterior endo-derm anlage). A purely text-basedsearch will yield 116, 122, and 246genes sharing 1, 2, or 3 CV terms withthe query gene, respectively. The genecounts produced by text searching aresignificantly larger than thoseobtained from either type of image-based search and they have the widestrange of expression overlap with thequery (Fig. 2C). Of the 86 FlyExpressidentified genes having � 50% overlapwith the twi query 25, 4, 18, and 39genes shared 0, 1, 2, or 3 CV termsrespectively, with the query. The abil-ity of FlyExpress to identify 25 geneswith patterns that overlap � 50% withthe twi pattern but share no commonCV terms with twi (29% of the FlyEx-press list) is strong evidence of thevalue of image-based search techni-ques for biological discovery. Four ofthese 25 are shown in Figure 2C.

We then performed a FlyExpressimage-based search of a Fly-FISH twiimage (Fbim9497472, profile a) from a

similar developmental stage and ana-tomical view. Note that representedgenes and embryonic stage ranges forthe BDGP and Fly-FISH datasets aredifferent, and thus P-values will vary,even if similar query images are cho-sen to search within each image collec-tion. FlyExpress searching identifies132 genes with a P-value < 0.05 andS-score � 0.5 (Fig. 3A). Only one genefrom the top nine of the FlyExpresstwi search of BDGP (Fig. 2A; tkv) wasamong the top nine in the Fly-FISHsearch. To gain insight into this dis-crepancy, we first examined the Fly-FISH dataset for the presence of theother eight top genes identified in theFlyExpress BDGP twi search. All butone (htl) is absent from the Fly-FISHdataset. Upon closer inspection, wefound htl among our initial Fly-FISHresults, but the image displayedslightly < 50% overlap to the twiquery image and thus was not countedin the final tally. Note that manygenes showing ubiquitous or near-ubiquitous expression had slightly >50% overlap with the query spatialprofile, thus providing a possible ex-planation for this discrepancy. Notethat our 50% overlap cutoff is also rel-atively arbitrary. If the cutoff was low-ered to 45% then the htl image wouldbe included. As for BDGP above, aFlyExpress search of the Fly-FISHdataset returned genes with signifi-cant expression overlap to the queryand to each other. At this time there isno other image searching method foranalyzing Fly-FISH data.The Fly-FISH twi query image is

associated with four CV terms (blasto-derm nuclei, expression in mesoderm,subset blastoderm nuclei and zygotic).A purely text-based search will yield151, 98, 393, and 93 genes sharing 1,2, 3, and 4 CV terms with the queryimage. Text-based searching withinFlyExpress identified set of 132 genesrevealed that 103, 8, 1, 13, and 7share 0, 1, 2, 3, and 4 CV termsrespectively, with the query (Fig. 3B).The ability of FlyExpress to identify103 genes with patterns that overlap� 50% with the twi pattern but shareno common CV terms with twi (78% ofthe FyExpress list) is again strongevidence of the value of image-basedsearch techniques for biological dis-covery. Four of the 103 images belong-ing to genes associated with 0 shared

Dev

elop

men

tal D

ynam

ics


CV terms are shown in Figure 3B.Note that many genes displayed ubiq-uitous or near-ubiquitous expressionand are textually annotated as such.Thus, when looking for genes showingsignificant overlap with an expressionpattern covering a large area of theembryo an abundance of significantoverlap with ubiquitously expressedgenes is expected.Three more analyses of FlyExpress

identified images (derived from a va-riety of query images) with regard tothe number of CV terms shared bythe FlyExpress list and the querywere conducted to determine therobustness of this result beyond a twiquery image. Independently selectedexemplar images chosen from Interac-tive Fly (Brody, 1999), as well asrandomly selected from the BDGPand Fly-FISH datasets were exam-ined. These included embryos of vari-ous stages and different views (Supp.Table S1; Supp. Fig. S1). In each anal-ysis 2–82% of the FlyExpress identi-fied genes with > 50% overlap withthe query did not share a CV termwith the query and would have beenmissed in a text search. Examinationof the images missed by text-basedsearching revealed that if the queryembryo has an expression patterncovering a large area of the embryo,then ubiquitous expression patternswere often missed.

Fig. 3. Image vs. text searching method comparison for Fly-FISH data. A: A FlyExpress searchfor overlapping expression of twi using a similar image (Fbim9497472 FlyExpress profile a). Thetop nine images of 132 genes identified are shown. B: The number of CV terms a given FlyEx-press identified gene shares with the query gene are graphed. Examples of four images with �50% overlap but sharing 0 CV terms with the query gene are shown below the graph.

Fig. 4. Paralogous gene pair classification with analysis of spatially and temporally different gene pairs over developmental time. A: Example imagepairs classified as having the same (0) or different (1) spatial or temporal expression patterns. BDGP gene expression is in blue. Fly-FISH gene expres-sion is in green/yellow. N/A - no divergence was visible. B: For the BDGP dataset the frequencies of spatially different and temporally different genepairs within all gene pairs were graphed according to developmental stage and S-score. Images exhibiting ubiquitous or no expression are not included.Legend showing characters used on the graph for each developmental stage and it’s two expression pattern comparisons (similar or different).

Dev

elop

men

tal D

ynam

ics

154 KONIKOFF ET AL.

Paralogous Gene Expression

Analysis Using FlyExpress

Using our data-driven approach, wenext demonstrate how the FlyExpressresource may be used to analyze geneco-expression. Standardized in situimages from both datasets for 251multigene families ranging in sizefrom 2–20 members (801 genes total;Supp. Table S2) were manuallycompared, in a pairwise manner, todetermine the presence or absence ofspatial and/or temporal expressionpattern divergence. Standardizedimages enable this type of quantita-tive image-based analysis, as embryosand their expression patterns arecomparable if they are of the samedata source, stage and anatomicalview, due to uniform size and shape(Fig. 4A). Two family members (evolu-tionarily known as paralogs becausewe are examining genes in the samespecies) were judged to have the samespatial expression if their imagesfrom the same anatomical view,developmental stage range, and datasource exhibited expression in thesame embryonic regions. Two paral-ogs were judged to have the sametemporal expression if their imagesdepicted the presence or absence ofexpression in the same developmentalstage range. Corroborating microar-ray data was obtained to improve con-fidence (Arbeitman et al., 2002).

Image pairs were also comparedcomputationally (post hoc) using theBasic Expression Search Tool forImages (BESTi; (Kumar et al., 2002).As expected, image pairs showing thehighest levels of divergent expressionin the manual inspection assay alsohad the lowest S-scores (smallest over-lap) and image pairs with the lowestlevels of divergent expression had thehighest S-scores (highest overlap; Fig-ure 4B shows the graph for spatialexpression). For an S-score between0.6 and 0.7 the similar and differentexpression pairs are at roughly equalfrequency. However, for an S-scoreof 0.9 to 1.0 the clean correlationbetween S-score and manual assess-ment of expression overlap ceases.

To determine the source of the dis-crepancy between manual and compu-tational estimates of similarity wereexamined the datasets and soon real-ized that computers, unlike humans,

cannot recognize and discard noiseautomatically. In such cases, BESTimay underestimate the amount ofexpression similarity for a particulargene pair. For example, if one imagehas artifactual staining in addition tonormal gene expression while anotherdoes not, or if different artificial stain-ing is present in two images that oth-erwise depict the same expressionpattern true similarity will be under-estimated. A second example is whenone image is over-stained whileanother is under-stained for the samegene expression pattern. A thirdexample applies particularly to olderembryos during organogenesis where itresults from naturally occurring mor-phological movements. For instance,Malpighian tubules and peripheralneurons migrate in all directions, elon-gate, twist, extend and retract almostcontinuously during late stages of de-velopment. As a result, a comparison ofimages with expression of the samegene in two stage 16 embryos that are15 min apart in age may have a simi-larity score less than 1.0 due to mor-phological events that occurred dur-ing those 15 min. The skill necessaryto ascertain the age of an embryo,within 15 min, takes considerableeffort to develop.

Alternatively, BESTi may overesti-mate expression similarity betweentwo images. The most commoninstance of this is when expressionappears in the same two-dimensionalspace but is, in reality, present in dif-ferent regions of three-dimensionalspace (e.g., above and below the planeof focus). For example, visceral mus-culature and fat body expression maybe impossible to distinguish two-dimensionally in images of laterallyoriented stage 13 embryos. Compari-son of expression by a gene present inthe visceral musculature and anotherin fat body will lead to significantsimilarity overestimation. See Supp.Fig. S3 for examples illustrating howbinary similarity may be over or under-estimated. Thus, while the FlyExpressresource simplifies the identificationand quantification of similar expressionpatterns, the biological importance ofthese results must be evaluated by abiologist, as is true for all computa-tional systems.

Of 1,068 gene pairs in our analysis,807 (75.6%) exhibited spatial gene

expression divergence during at leastone stage of development. To deter-mine if spatial pattern divergence ismore frequent in genes with particu-lar functions, we tested the mostdivergent gene pairs for GO termenrichment. In the analysis, membersof gene pairs showing a predominanceof spatial divergence across develop-ment (the pair had divergent compari-sons over a majority of stages) werecompared with the larger group ofgene pairs that simply had enoughimage data to be assessed over themajority of developmental stages. Wefound three Biological Process termsat a statistically significant frequency(Fig. 5A; Supp. Table S3).To determine if spatial divergence

varied across developmental stages,the proportion of gene pairs exhibit-ing spatial differences was deter-mined for each developmental stagerange. These proportions for eachdataset, individually and combined,were then graphed according tomedian stage range developmentaltimes (Campos-Ortega and Harten-stein, 1985). The graph (Fig. 5B)reveals the fraction of spatially diver-gent gene pairs increases significantlyin all three cases (P-value < 0.05) asembryogenesis progresses. Just 2.9%of the pairs show spatial differencesin the youngest embryos (stages 1–3)that increased to 93.2% of pairs in theoldest (stages 13–16). Specifically, inthe earliest stages there is an excessof spatial similarity because ubiqui-tous expression is abundant. Theproportion of divergent pairs thenexpands swiftly with stages 4–6 show-ing roughly equal proportions of simi-lar and different pair expression. Bystages 10–12 (just halfway through de-velopment based on time), nearly 90%of the gene pairs show spatial diver-gence and spatial divergence reaches amaximum shortly thereafter.In contrast, there were just 168

(15.7%) gene pairs exhibiting tempo-ral divergence over at least one stagerange. Note that only 291 of the genepairs had comparable image data forat least one stage range in the Fly-FISH dataset, compared with 869 inBDGP (91 gene pairs had images inboth datasets). This is almost fivefoldless than the number of gene pairsexhibiting spatial divergence in atleast one stage range. This distinction

Dev

elop

men

tal D

ynam

ics


was truly unexpected as a priori therewas no reason to think that eitherspatial or temporal divergence wouldpredominate. To determine if tempo-ral divergence was associated withgene pairs of specific function, GOterm enrichment was again exam-ined. As above, members of gene pairsshowing a predominance of temporaldivergence across development (thepair had divergent comparisons overa majority of stages) were comparedwith the larger group of gene pairsthat simply had enough image data tobe assessed over the majority of devel-opmental stages. We found three Bio-logical Process terms at a statisticallysignificant frequency (Fig. 5C; Supp.Table S4). Two of these terms werealso found among spatially divergentpairs. The term ‘‘biological regulation’’is enriched only in spatially divergentgenes, while the term ‘‘developmentalprocess’’ is enriched only in tempo-rally divergent ones. Furthermore,temporally divergent pairs areenriched in the Molecular Functionterm ‘‘binding.’’

To determine if temporal diver-gence varied across developmentalstages, the proportion of gene pairsexhibiting temporal differences wasdetermined for each developmental

stage range. In the BDGP dataset,the fraction of gene pairs displayingtemporal expression differences isroughly stable during the first thirdof development at 20–24% (stages 1–9are not significantly different) butdecreases significantly (P-value <0.05) during the last two-thirds of de-velopment Figure 5D. The Fly-FISHdataset, with its emphasis on earlyembryos (the vast majority derivefrom stages 1–9), showed few tempo-rally different pairs, as well as rela-tively fewer comparable pairs overallcompared with BDGP, and thus wasunchanged over time. The decrease intemporal pairs for BDGP (which asthe larger dataset is responsible forthe combined results) is in sharp con-trast to the trajectory for spatialdivergence (continuously increasingin frequency). For BDGP, the fractionof temporally different pairs rangesfrom a high of 24% (stages 8–9) toa low of 0.6% in the oldest embryos(stages 13–16).

Taken together the spatial and tem-poral divergence data reveals thatgene pairs within a multigene familyare more likely to diverge in expres-sion temporally rather than spatiallybefore the maternal to zygotic expres-sion transition (stages 4–5; 20% tem-

porally vs.. 2.9% spatially for BDGP).In contrast, during late stages ofdevelopment gene pairs within amultigene family are more likely todiverge in expression spatially thantemporally (stages 13–16; 90.3% spa-tially vs. 0.6% temporally for BDGP).

DISCUSSION

Overall, our results for image vs. text-based searching illustrate that image-based searching for overlapping expres-sion patterns using the FlyExpressresource presents unique advantagescompared to text-based searching.Importantly, the comparison of imageand text-based searching suggeststhat co-expressed genes (even whenthey display substantial overlap) donot always share common text anno-tations. Therefore, if one’s aim is tofind genes with similar or overlappingpatterns of expression then directsearching of standardized expressionpatterns is more powerful than thetext-based searching. However, as ourstudy of spatial divergence suggests,users of image searching mustremember that, as with any computa-tional method, decisions about biologi-cal significance are theirs.In addition, textual descriptions

remain an important feature of dis-covery in developmental biologybecause they are crucial for communi-cating expression details across differ-ent embryonic stages, anatomicalviews and even species boundaries.Textual annotations also allow incor-poration of the three-dimensionalnature of the developing embryo, andkeen annotators may easily distin-guish between expression and back-ground or artifact staining. Therefore,FlyExpress incorporates BDGP andFly-FISH produced text informationalongside image data so users canmore easily discern the biological rele-vance of search results.Our analysis of expression diver-

gence among paralogous genes inmultigene families suggests spatialrather than temporal divergence ingene expression patterns is thepredominant contributor to the devel-opment of new morphological fea-tures, especially after the maternal tozygotic transition. One example of amultigene family with significant spa-tial and temporal divergence is the

Fig. 5. Paralogous gene pair expression analysis. A: Statistically significant GO term enrich-ment is present for three terms in members of gene pairs with spatial expression divergenceacross a majority of developmental stages. B: An average spatial divergence was calculated forall pairwise comparisons with spatial differences, and noted as percent divergence, for theBDGP, Fly-FISH, and combined datasets. These averages are graphed according to develop-mental stage after each stage was converted to time in minutes since fertilization based on theestablished timeline of embryonic development in Drosophila. Standard deviations are shownand statistical significance (P-value <0.05) was determined by the two-tailed P-value for a com-parison of the mean divergence values of the first and last developmental time points. The me-dian times of the seventeen Campos-Ortega and Hartenstein (1985) stages are also includedbelow the minute intervals. C: Statistically significant GO term enrichment is present for fourterms in members of gene pairs with temporal expression divergence across a majority of devel-opmental stages. D: The average divergence of gene pairs showing temporal differences wascalculated and graphed as above.

Dev

elop

men

tal D

ynam

ics

156 KONIKOFF ET AL.

innexin family. This family is highlyconserved and plays a variety of rolesin cell migration and in organogenesis(Bauer et al., 2002; Lechner et al.,2007). Genes associated primarilywith spatial divergence are alsoenriched in the GO term ‘‘biologicalregulation.’’ One example of a multi-gene family with significant spatialbut not temporal divergence is theMinichromosome maintenance family.This family is highly conserved acrossphyla and functions in genome main-tenance and DNA replication (Fors-burg, 2004). Genes primarily associ-ated with temporal divergence areenriched in two different GO terms:‘‘developmental process’’ and ‘‘binding.’’One example of a multigene familywith significant temporal but not spa-tial divergence is the Imaginal DiscGrowth Factor (IDGF) family. Thisfamily is also conserved (evolved fromchitinases) and influences cell prolif-eration (Kawamura et al., 1999; Zur-ovcova and Ayala, 2002).

Looking toward the future, increas-ingly sophisticated and intercon-nected computational tools for imageanalysis of gene expression patternswill speed the discovery of new rela-tionships between genes, genomic seg-ments, and ultimately species. To fosterthe discovery process, FlyExpress con-tains automated portals into FlyBaseand FlyMine. Looking more broadly,the search method used by FlyExpresscould potentially be applied to othertwo-dimensional expression datasets,provided that images are capable ofbeing effectively standardized. Ulti-mately, effective standardization willresult in significant overlap of morpho-logical features, and thus allow forexpression pattern comparisons.

Our usage of FlyExpress to identifyclues to molecular mechanisms under-lying the process whereby new genesacquire novel functions is just one of amyriad ways in which this discoveryplatform can contribute to our under-standing of developmental and evolu-tionary biology. This resource also hasmany other potential uses, limitedonly by the investigator’s imagination.One example is to examine expressionoverlap in gene networks, particularlythreshold activation of genes inresponse to important signaling gra-dients such as the one produced byDorsal (Stathopoulos and Levine,

2002; Papatsenko and Levine, 2005).Alternatively, FlyExpress could beused to predict interactions betweenembryonic genes with specific spatialexpression profiles, as has been donein oogenesis (Yakoby et al., 2008).

EXPERIMENTAL

PROCEDURES

FlyExpress Discovery

Platform: Gene Expression

Images and Image-Based

Searching

Digital images revealing patterns ofD. melanogaster gene expression, cap-tured by RNA in situ hybridization,were retrieved from BDGP Release 2(Tomancak et al., 2002) and Fly-FISHSeptember 2007 release (Lecuyeret al., 2007). All images were proc-essed using an in-house, semi-auto-mated pipeline to standardize andalign embryos, where multi-embryoimages were manually divided intoseparate images and partial embryoimages were discarded. Image proc-essing was carried out using Matlaband our own image processing rou-tines. In this pipeline, images wereextracted by means of the followingprocedure (Matlab functions areshown in italics): read original imagewith imread, convert to grayscalewith rgb2gray, apply windowed low-pass Gaussian filer (imfilter and fspe-cial) to blur the image a little so thatedge detection would detect only themain outside edges of the embryo andnot edges caused by expression pat-terns, shadows, etc., delineate embryoedges using edge Canny edge detection,expand points with imdilate, fill holesin the image with imfill, shrink dila-tions, blur again and sharpen theimage edges with imerode, strel, andmedfilt2. The bwlabel function is usedto identify individual embryo bounda-ries. Finally all pixels outside the selec-tion region (outside the embryo) are setto pure white. The resulting image issaved in RGB color as a bitmap file.

The next standardization step isembryo alignment, which is done byrotating embryos using imrotate, suchthat the major axis of the embryo isparallel to the horizontal, drawing abounding box to enclose the embryo,and cropping the embryo usingimcrop to automatically remove back-

ground external to the smallestembryo bounding box. For consistentorientation, we used anterior-on-the-left and dorsal-toward-the-top formatfor lateral images and anterior-on-the-left for all other views (e.g., dorsaland ventral; see also Kumar et al.,2002). During quality control, experi-enced biologists corrected orientationand alignment images using flipdim,as necessary.To size standardize and align all

images, we chose a cellular aspectratio of 2.5 (320 � 128 pixels) basedon natural aspect ratios (Markowet al., 2009), on the need to avoid pixelpadding in image representation andto make sure that each line of pixelsand all images end on byte, word andlong word boundaries. Using theimresize function, all embryo imageswere resized and a standardized col-lection created.Developmental stages for Fly-FISH

and BDGP embryos are available fromthe image source and were thusassigned to embryos. BDGP embryosare annotated with developmentalstage ranges (1–3, 4–6, 7–8, 9–10,11–12, or 13–16) using the Bownes sys-tem (16 stages; Roberts, 1986) whereasFly-FISH embryos are classified intostage ranges (1–3, 4–5, 6–7, 8–9, andlater) using the Campos-Ortega andHartenstein (1985; 17 stages) system.We found that a large fraction of Fly-FISH images contained multipleembryos from different developmentalstages. We carefully reviewed andassigned appropriate stage ranges toindividual embryos for these cases. Wealso assigned anatomical embryoviews (e.g., lateral, dorsal, and ventral)for all images, because computationalcomparison of spatial expression pat-terns is only biologically meaningful ifconducted within each view. Thesestage and view assignments wereadded during our quality control pro-cess, where all embryos were exam-ined by at least one developmental bi-ologist and image standardization andexpression extraction were carried outmanually, when needed. This produceda total of 99,148 standardized embryoimages: 42,065 Fly-FISH and 57,083BDGP. Within BDGP, there are 14,257dorsal, 37,825 lateral, and 5,964 ven-tral views, and within Fly-FISH, thereare 3,494 dorsal, 37,211 lateral, and1,360 ventral views.

Dev

elop

men

tal D

ynam

ics


Comparative expression analysisto identify images (and thus genes)with overlapping expression patternsrequires digital descriptions of spatialexpression patterns (Kumar et al.,2002). We used adaptive intensity andcolor thresh-holding to automaticallydelineate the expression profile fromthe embryo background. For Fly-FISH images, expression patterns arecaptured in green and yellow colors,and in BDGP images, blue color cap-tures the spatial profile. In the pro-cess of image extraction there arefour parameters: the color layer (red,green, or blue), whether to reverse in-tensity, an adjustment parameter forshift threshold and the amount of var-iance between the three expressionpatterns produced. The RGB image isloaded with imread. Based on thecolor channel (layer) parameter, a his-togram of intensities for just thatcolor layer is created with imhist. Thehistogram is ‘‘pre-cut’’ to only includeintensities between 5 and 250. Thesum of the intensities in the pre-cuthistogram is calculated with sum.Upper and lower limits of the histo-gram are computed for the rangebetween 10% and 90% of this sum. Anarea (‘‘integration’’) vector is calcu-lated for the color layer (specified bythe color parameter) of the originalimage. Using this area vector and theoriginal intensities, we calculate thefirst moment and the centroid of thearea under the under the curve aboutthe y-axis and the first moment andthe centroid of the area under thecurve about the x-axis. A thresholdvalue is calculated as the centroidabout the y-axis þ the lower limit cal-culated above. The minimum distancefrom the x–y center is calculated foreach point is the intensity histogram.The plus_da (the differential bestarea þ variance) and minus_da (thedifferential best area - variance) arecalculated. plus_da and minus_dathresholds are calculated from these.The image background is clearedusing imdilate (Image, SE).

The three binary extraction pat-terns are extracted from this imagewith cleared background based on theadjustment parameter and thethreshold (for the best extraction), theminus_da threshold (for the minusextraction) and the plus_da_threshold(for the plus extraction). The code

for this algorithm is available uponrequest. Three binary (black andwhite) patterns enable searches of thedatabase using patterns at differentlevels of expression intensity (a, b,and c patterns). Embryos with ubiqui-tous expression in the earliest stagesof development as well as those withno expression were noted and markedfor exclusion from comparative imageanalysis.

To identify the degree of spatialoverlap between patterns, we usedthe low-level bitmap Jaccard similar-ity index (Kumar et al., 2002), whichtraces its roots to the Tanimoto mea-sure (Tanimoto, 1958) and is a memberof a family of similarity measures thatinclude Taversky, Euclidian, Ham-ming, and Ochia measures (Bradshaw,2001). In this case, the similarity score(S) between two images (Q and D) isgiven by SQD ¼ |Q\D|/|Q|D|,where |Q\D| is the size of the inter-section of expression (count of blackpixels) between images Q and D and|Q|D| is the size of the unionbetween images Q and D. We havepreviously shown that this approachemphasizes spatial overlap, whichis biologically more meaningfulthan shape matching and invariantmoment based features (Kumar et al.,2002; Gurunathan et al., 2004). Wehave also found that it performs withan effectiveness similar to the compu-tationally more intensive GaussianMixture Model method (Peng et al.,2007), which in our hands is very sen-sitive to shifts in image properties,such as the color and contrast (Garge-sha et al., 2008, 2009a,b; Roy et al.,2009). Thus, we used multiple binaryfeature vectors to represent theexpression information in our analysisand provide the greatest flexibility.For each S-score, we also computedthe probability that any pair of imageswill show an equal or higher value bychance alone (P-value). EmpiricalS-score distributions derived fromimage pairs from the same data source(BDGP or Fly-FISH), developmentalstage and anatomical view are used todetermine P-values (Supp. Fig. S2).

For images from PubMed publica-tions, gene expression patterns weremanually extracted from the PDF fileby our pipeline team. Extractedimages were then standardized andexpression pattern representations

created using the same methodsdescribed above for BDGP and Fly-FISH. Images were manually anno-tated, for example for developmentalstage and anatomical view (lateral,dorsal, ventral), by our team of cura-tors at Harvard (FlyBase) and by biol-ogists on our pipeline team. Imagesfrom publications present specialchallenges due to their varying qual-ity. For images of low resolution orpoor quality, it is not possible to makespecific developmental stage determi-nation. In this case the best possiblestage range (from-stage/to-stage) isdetermined for the annotation. Publica-tion information (for example author,year, and title) for the papers wasobtained from the FlyBase publicallyavailable database.A detailed description of all compu-

tational aspects of FlyExpress and ofthe extracted gene expression patterndatabase can be found in Kumar et al.(2011) and the source code for all algo-rithms is available upon request.

Finding and Aligning

Multigene Families

All genes present in the BDGP (Tom-ancak et al., 2002) and Fly-FISH(Lecuyer et al., 2007) datasets wereanalyzed by tblastn, using the longestprotein isoform available from NCBIand the D. melanogaster genomesequence. Genes were consideredparalogous if reciprocal hits contain-ing �50% amino acid positives withE-value < 10�20 were returned. Addi-tional hits meeting these criteriawere analyzed by tblastn until nomore paralogous genes were retrievedcompleting that particular multigenefamily. The subset of multigene fami-lies with expression data for at leasttwo paralogs was retained for furtheranalysis. Multigene family sequenceswere aligned with default parametersin MAFFT (Katoh et al., 2005; Nuinet al., 2006) and sequences re-gappedwith GeneDoc (http://www.nrbsc.org/gfx/genedoc/index.html). Alignmentswere corrected by eye and modified(tails cut off) if necessary.

Expression Assessment

For each gene pair, the standardizedBDGP or Fly-FISH expression pat-terns in FlyExpress were manually

Dev

elop

men

tal D

ynam

ics

158 KONIKOFF ET AL.

compared to determine if the same ordifferent expression patterns werepresent, either spatially or tempo-rally. Manual comparison was neces-sary because we could not identifyan S-score cut-off that adequatelyrepresented spatial and temporaldivergence in comparisons involvingembryos with artifact staining, highbackgrounds, or embryo pairs withthree-dimensional expression consid-erations. Even when using manualinspection, posterior spiracles andother openings that are frequently asource of artifacts must show expres-sion in at least two images of thesame gene at the same stage and viewto be considered legitimate.

Image comparisons were performedwithout any additional information,but comparisons within a multigenefamily were verified to be free of logi-cal inconsistencies. Pattern similarityor difference was determined on thebasis of the presence or absence ofexpression within the same region ofthe embryo. Although only images ofthe same data source, developmentalstage and view were compared, genepairs were binned into BDGP stageranges and each pair given an overallassessment. The exception was stage1–3 BDGP embryos where the exactstage cannot be assigned due toinability to visualize the number ofnuclei. Gene pairs were assigned anoverall spatial and temporal assess-ment of Same (0), Different (1), orAmbiguous/No Data (null).

Spatially, only images of the samedata source, developmental stage, andanatomical view were compared. Agene pair was classified as having thesame expression in a stage if all imagepairs showed expression in the sameembryonic regions. A gene pair wasclassified as having divergent expres-sion if image pairs within a stageshowed expression in different embry-onic regions. Gene pairs with noimages meeting the above criteria(including those with no image data)were classified as Ambiguous/No Data.

Temporally, images of the samedata source and stage were examined.Two genes were classified as havingthe same expression in a stage if allimage pairs showed expression pres-ence or absence during that stage,accompanied by corroborating micro-array data (Arbeitman et al., 2002).

Two genes were classified as havingdifferent expression in a stage if atleast one image pair showing expres-sion presence for one gene andabsence for the other was present,again accompanied by corroboratingmicroarray data. Gene pairs not fall-ing into one of the above categorieswere classified as Ambiguous/NoData. Gene pairs exhibiting temporalexpression differences in a stage werenot assessed spatially for that stage.

GO Term Analysis

FlyBase (release FB2009_09) GO termannotations (Tweedie et al., 2009)were retrieved for paralogous BDGPand Fly-FISH genes (http://flybase.bio.indiana.edu/static_pages/termlink/termlink.html). Overall enrichmentwas determined in the context of totalFlyBase term annotations by meansof hypergeometric distribution. Sig-nificantly enriched terms were thosewith P-value < 0.0001. For spatialanalysis, there were 211 genes from158 gene pairs where spatial patternswere assessed over at least 2/3 ofBDGP stages. Of these 167 genesfrom 119 gene pairs showed a spatialexpression divergence over a majorityof BDGP stages. For temporal analy-sis, there were 392 genes from 285pairs assessed over at least 2/3 ofBDGP stage ranges. Of these 92genes (from 57 gene pairs) exhibitedtemporal expression divergence overa majority of BDGP stages.

Developmental Time Course

Analysis

Individual pairwise comparisons werebinned according to developmentalstage range. The fraction of genesexhibiting spatial or temporal diver-gence was calculated for each bin(1/0þ1) and plotted against the medianBDGP stage range times (Campos-Ortega and Hartenstein, 1985). Statis-tical significance (P-value < 0.05) wasdetermined by the two-tailed P-value.

ACKNOWLEDGMENTSWe thank Toni Marco and RachelSipes for assistance with paralogousgene analysis. S.K. was funded by aNIH grant.

REFERENCES

Altschul S, Gish W, Miller W, Myers E,Lipman D. 1990. Basic local alignmentsearch tool. J Mol Biol 215:403–410.

Arbeitman M, Furlong E, Imam F, John-son E, Null B, Baker B, Krasnow M,Scott M, Davis R, White K. 2002. Geneexpression during the life cycle of Dro-sophila melanogaster. Science 297:2270–2275.

Bauer R, Lehmann C, Fuss B, Eckardt F,Hoch M. 2002. The Drosophila gapjunction channel gene innexin 2 con-trols foregut development in responseto Wingless signalling. J Cell Sci 115:1859–1867.

Bier E. 2005. Drosophila, the golden bug,emerges as a tool for human genetics.Nat Rev Genet 6:9–23.

Bradshaw J. 2001. YAMS - Yet anothermeasure of similarity. In: Euromug01,13th–14th September 2001, Cambridge(UK).

Brody T. 1999. The interactive fly: genenetworks, development and the Inter-net. Trends Genet 15:333–334.

Campos-Ortega JA, Hartenstein V. 1985.The embryonic development of Drosoph-ila melanogaster. New York: Springer-Verlag. xi, 227 p.

de Velasco B, Mandal L, Mkrtchyan M,Hartenstein V. 2006. Subdivision anddevelopmental fate of the head meso-derm in Drosophila melanogaster. DevGenes Evol 216:39–51.

Dorfman R, Shilo BZ. 2001. Biphasic acti-vation of the BMP pathway patternsthe Drosophila embryonic dorsal region.Development 128:965–972.

Drysdale R. 2001. Phenotypic data inFlyBase. Briefings in Bioinformatics 2:68–80.

FlyBase. 1999. The FlyBase database ofthe Drosophila genome projects andcommunity literature. Nucleic AcidsRes 27:85–88.

Forsburg SL. 2004. Eukaryotic MCM pro-teins: beyond replication initiation.Microbiol Mol Biol Rev 68:109–131.

Frise E, Hammonds AS, Celniker SE. 2010.Systematic image-driven analysis of thespatial Drosophila embryonic expressionlandscape. Mol Syst Biol 6:345.

Gargesha M, Jenkins M, Rollins A, Wil-son D. 2008. Denoising and 4D visual-ization of OCT images. Opt Express 16:12313–12333.

Gargesha M, Jenkins M, Wilson D, Roll-ins A. 2009a. High temporal resolutionOCT using image-based retrospectivegating. Opt Express 17:10786–10799.

Gargesha M, Qutaish M, Roy D, SteyerG, Bartsch H, Wilson D. 2009b.Enhanced volume rendering techniquesfor high-resolution color cryo-imagingdata. Proc Soc Photo Opt Instrum Eng7262:72655V.

Grumbling G, Strelets V. 2006. FlyBase:anatomical data, images and queries.Nucleic Acids Res 34:D484–D488.

Gurunathan R, Van Emden B, Panchana-than S, Kumar S. 2004. Identifying spa-tially similar gene expression patterns

Dev

elop

men

tal D

ynam

ics


in early stage fruit fly embryo images:binary feature vs. invariant momentdigital representations. BMC Bioinfor-matics 5:202.

Ip YT, Park RE, Kosman D, YazdanbakhshK, Levine M. 1992. dorsal-twist interac-tions establish snail expression in thepresumptive mesoderm of the Drosophilaembryo. Genes Dev 6:1518–1530.

Janning W. 1997. FlyView, a Drosophilaimage database, and other Drosophiladatabases. SeminCell DevBiol 8:469–475.

Katoh K, Kuma K, Toh H, Miyata T.2005. MAFFT version 5: improvementin accuracy of multiple sequence align-ment. Nucleic Acids Res 33:511–518.

Kawamura K, Shibata T, Saget O, Peel D,Bryant P. 1999. A new family of growthfactors produced by the fat body andactive on Drosophila imaginal disc cells.Development 126:211–219.

Khatri P, Draghici S. 2005. Ontologicalanalysis of gene expression data: cur-rent tools, limitations, and open prob-lems. Bioinformatics 21:3587–3595.

Koonin EV, Fedorova ND, Jackson JD,Jacobs AR, Krylov DM, Makarova KS,Mazumder R, Mekhedov SL, NikolskayaAN, Rao BS, Rogozin IB, Smirnov S, Sor-okin AV, Sverdlov AV, Vasudevan S, WolfYI, Yin JJ, Natale DA. 2004. A compre-hensive evolutionary classification of pro-teins encoded in complete eukaryoticgenomes. GenomeBiol 5:R7.

Kumar S, Jayaraman K, PanchanathanS, Gurunathan R, Marti-Subirana A,Newfeld S. 2002. BEST: a novel compu-tational approach for comparing geneexpression patterns from early stages ofDrosophila melanogaster development.Genetics 162:2037–2047.

Kumar S, Konikoff C, Van Emden B,Busick C, Davis K, Ji S, Wu L, Ramos H,Brody T, Panchanathan S, Ye J, Karr T,Gerold K, McCutchan M, Newfeld SJ.2011. FlyExpress: visual mining of spa-tiotemporal patterns for genes and publi-cations in Drosophila embryogenesis.Bioinformatics (in press; Aug. 15)

Lechner H, Josten F, Fuss B, Bauer R,Hoch M. 2007. Cross regulation ofintercellular gap junction communica-tion and paracrine signaling pathwaysduring organogenesis in Drosophila.Dev Biol 310:23–34.

Lecuyer E, Yoshida H, Parthasarathy N,Alm C, Babak T, Cerovina T, Hughes T,Tomancak P, Krause H. 2007. Globalanalysis of mRNA localization reveals aprominent role in organizing cellulararchitecture and function. Cell 131:174–187.

Markstein M, Zinzen R, Markstein P, YeeKP, Erives A, Stathopoulos A, LevineM. 2004. A regulatory code for neuro-genic gene expression in the Drosophilaembryo. Development 131:2387–2394.

Markow T, Beall S, Matzkin L. 2009. Eggsize, embryonic development time andovoviviparity in Drosophila species.J Evol Biol 22:430–434.

Matthews KA, Kaufman TC, GelbartWM. 2005. Research resources for Dro-sophila: the expanding universe. NatRev Genet 6:179–193.

Nuin P, Wang Z, Tillier E. 2006. The accu-racy of several multiple sequence align-ment programs for proteins. BMCBioinformatics 7:471.

Papatsenko D, Levine M. 2005. Quantita-tive analysis of binding motifs mediatingdiverse spatial readouts of the Dorsalgradient in the Drosophila embryo. ProcNatl Acad Sci U S A 102:4966–4971.

Peng H, Long F, Zhou J, Leung G, EisenM, Myers E. 2007. Automatic imageanalysis for gene expression patterns offly embryos. BMC Cell Biol 8(suppl 1):S7.

Roberts DB. 1986. Drosophila: a practicalapproach. Washington DC: IRL Press.xix, 295 p.

Roy D, Steyer G, Gargesha M, Stone M,Wilson D. 2009. 3D cryo-imaging: avery high-resolution view of the wholemouse. Anat Rec 292:342–351.

Stathopoulos A, Levine M. 2002. Linearsignaling in the Toll-Dorsal pathway ofDrosophila: activated Pelle kinase

specifies all threshold outputs of geneexpression while the bHLH proteinTwist specifies a subset. Development129:3411–3419.

Stathopoulos A, Tam B, Ronshaugen M,Frasch M, Levine M. 2004. pyramusand thisbe: FGF genes that pattern themesoderm of Drosophila embryos.Genes Dev 18:687–699.

Tanimoto T. 1958.An elementary mathe-matical theory of classifications andprediction In: IBM Corp. InternalReport 17th November.

Tomancak P, Beaton A, Weiszmann R,Kwan E, Shu S, Lewis S, Richards S,Ashburner M, Hartenstein V, CelnikerS, Rubin G. 2002. Systematic determi-nation of patterns of gene expressionduring Drosophila embryogenesis. Ge-nome Biol 3:RESEARCH0088.

Tweedie S, Ashburner M, Falls K, LeylandP, McQuilton P, Marygold S, Millburn G,Osumi-Sutherland D, Schroeder A, SealR, Zhang H. 2009. FlyBase: enhancingDrosophila Gene Ontology annotations.Nucleic Acids Res 37:D555–D559.

Walter T, Shattuck DW, Baldock R, BastinME, Carpenter AE, Duce S, EllenbergJ, Fraser A, Hamilton N, Pieper S,Ragan MA, Schneider JE, Tomancak P,Heriche JK. 2010. Visualization ofimage data from cells to organisms. NatMethods 7:S26–S41.

Yakoby N, Bristow CA, Gong D, SchaferX, Lembong J, Zartman JJ, Halfon MS,Schupbach T, Shvartsman SY. 2008. Acombinatorial code for pattern forma-tion in Drosophila oogenesis. Dev Cell15:725–737.

Zimmermann G, Furlong EE, Suyama K,Scott MP. 2006. Mes2, a MADF-contain-ing transcription factor essential forDrosophila development. Dev Dyn 235:3387–3395.

Zurovcova M, Ayala F. 2002. Polymorphismpatterns in two tightly linked develop-mental genes, Idgf1 and Idgf3, ofDrosoph-ilamelanogaster. Genetics 162:177–188.

Dev

elop

men

tal D

ynam

ics

160 KONIKOFF ET AL.

Get PDF (1011K) Get it!@ASU

Find more content: like this article

Find more content written by: Charlotte E. Konikoff Timothy L. Karr Michael McCutchan Stuart J. Newfeld Sudhir Kumar All Authors

Additional Supporting Information may be found in the online version of this article.

Filename FormatSize Description

DVDY_22749_sm_SuppFig1A.tif 675K Supp. Fig. S1. Image-based versus text-based search results. A–C: Interactive Fly (A),randomly selected BDGP images (B), and randomly selected Fly-FISH images (C) as queryimages. Left: Query image. Center - a square represents a single image identified byFlyExpress searching of the BDGP dataset with the indicated query. The S-score (similarityindex stated as percent of overlapping expression assessed with BESTi) on the Y-axis isshown for identified images with S-score of ≥ 0.5. X-axis is the number of shared CV termsbetween the identified FlyExpress image and the query. Images identified by FlyExpressthat would be missed by text searching are shown above the 0 on the X-axis. Right:Examples of FlyExpress identified images with ≥ 50% overlap to the query image thatwould be missed by text searching.

DVDY_22749_sm_SuppFig1B.tif 517K Supp. Fig. S1. Image-based versus text-based search results. A–C: Interactive Fly (A),randomly selected BDGP images (B), and randomly selected Fly-FISH images (C) as queryimages. Left: Query image. Center - a square represents a single image identified byFlyExpress searching of the BDGP dataset with the indicated query. The S-score (similarityindex stated as percent of overlapping expression assessed with BESTi) on the Y-axis isshown for identified images with S-score of ≥ 0.5. X-axis is the number of shared CV termsbetween the identified FlyExpress image and the query. Images identified by FlyExpressthat would be missed by text searching are shown above the 0 on the X-axis. Right:Examples of FlyExpress identified images with ≥ 50% overlap to the query image thatwould be missed by text searching.

DVDY_22749_sm_SuppFig1C.tif 763K Supp. Fig. S1. Image-based versus text-based search results. A–C: Interactive Fly (A),randomly selected BDGP images (B), and randomly selected Fly-FISH images (C) as queryimages. Left: Query image. Center - a square represents a single image identified byFlyExpress searching of the BDGP dataset with the indicated query. The S-score (similarityindex stated as percent of overlapping expression assessed with BESTi) on the Y-axis isshown for identified images with S-score of ≥ 0.5. X-axis is the number of shared CV termsbetween the identified FlyExpress image and the query. Images identified by FlyExpressthat would be missed by text searching are shown above the 0 on the X-axis. Right:Examples of FlyExpress identified images with ≥ 50% overlap to the query image thatwould be missed by text searching.

DVDY_22749_sm_SuppFig2.tif 833K Supp. Fig. S2. The P-value distributions of pairwise S-scores for BDGP and Fly-FISHexpression data. For each stage, view, and image-source the probability of randomlymatching a query expression pattern with greater than or equal to a given S-score is plottedfor all S > 0.25. This approach is analogous to that taken by BLAST (Altschul, et al., 1990).

DVDY_22749_sm_SuppFig3.tif 1562K Supp. Fig. S3. FlyExpress images illustrating circumstances when expression overlap canbe under or overestimated. Image pairs representing ways in which binary similarity can beover or underestimated are shown below. A,B: Underestimation of similarity can be due toartifact staining (e.g., expression in the esophagus and dorsal trunk branch on the left isdue to a technical flaw and not associated with faf). C,D: Underestimation of similarity canoccur when expression occurs in tissues that display rapid morphologically movements(e.g., Malpighian tubules for CG17209 on the left and CG10086 on the right). E,F: Over- orunderestimation of similarity can be due to differences in expression intensity and/or thelevel of background staining can cause (e.g., overstaining for Doc3 on the left). G,H:Similarity can be overestimated if different three-dimensional (3D) expression domainsoverlap in 2D space (e.g., somatic muscle expression near the embryo surface forCG14122 on the left versus visceral muscle and fat body expression near the embryocenter for ade5).

DVDY_22749_sm_SuppTab.pdf 583K Supporting Information Tables

Please note: Wiley-Blackwell are not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the correspondingauthor for the article.

Jump to…Supporting Information

More content like this

Comparison of embryonic expression within multigene families using the flyexpress discovery plat... http://onlinelibrary.wiley.com/doi/10.1002/dvdy.22749/full

7 of 7 10/3/11 9:07 AM

For Peer Review

Supplemental Figure S1A

114x95mm (300 x 300 DPI)

Page 38 of 42

John Wiley & Sons, Inc.

Developmental Dynamics

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

Supplemental Figure S1B

114x95mm (300 x 300 DPI)

Page 39 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

Supplemental Figure S1C

114x95mm (300 x 300 DPI)

Page 40 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

Supplemental Figure S2

165x132mm (300 x 300 DPI)

Page 41 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

Supplemental Figure S3

76x70mm (300 x 300 DPI)

Page 42 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

1

Table S1. Image-based vs. text-based searching using exemplar images.

Query

Genea

Image ID Query image

FlyExpress

spatial

profileb

FlyExpress

Results

(≥50%overlap)

Text-based

Resultsc

Genes with

≥50% overlap

in both

searches

tsh FBim9036362 C 311 199 65 (20.9%)

sna* FBim9036229 C 37 273 31 (83.8%)

NetA* FBim9038181 B 44 234 27 (61.4%)

tkv* FBim9037528 C 31 685 30 (96.8%)

GATAe* FBim9039229 B 102 952 97 (95.1%)

hb FBim9039002 C 294 595 52 (17.7%)

Dr FBim9044026 B 88 1069 40 (45.5%)

Acf1 FBim9044260 B 59 660 54 (91.5%)

a. A random subset of Interactive Fly exemplar images from the BDGP dataset were subjected to

image-based (FlyExpress) and text-based (CV terms) searching of the BDGP dataset. The subset

of that group shown here consists of images returning at least 10 different genes with at least one

image of ≥50% expression overlap to the query image in FlyExpress.

b. Chosen from among the three spatial profiles available in FlyExpress for each gene expression

pattern image.

c. For every gene except tsh text searching returned more images than FlyExpress searching but

text searching did not identify roughly 35.5% of genes found by FlyExpress with the same query

image (ranging from 1.6% to 82.3% for individual genes). A non-matching Fly-Express and text

result is an image identified by FlyExpress with ≥50% expression similarity that has no shared

CV terms with the query image. This is shown in Fig. S1A for genes marked with an asterisk.

Page 32 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

2

Table S2. List of the 801 genes belonging to multigene families analyzed in this study

Gene names are given in FlyExpress Release 4 and listed alphabetically.

26-29-p 4EHP abd-A Acer Acox57D-d Act57B Act5C Act87E

Actn Adgf-A Adgf-B Ahcy13 Ahcy89E al Aldh alphaTub67C

Ama Ance Ance-4 Ank Ank2 AnnX Antp Anxb11

ap ApepP aret Arf51F Arp87C arr Art1 Art3

Art4 Art8 As ash1 Atet aux Bap55 Bc

betaInt-nu betaTub56D betaTub60D betaTub85D bgm bib boi brat

bru-3 Bsg25A btd bw Bx CaBP1 Caf1 CalpA

CalpB Cam CanB CanB2 cbt Ccp84Aa Ccp84Ab Ccp84Ag

Cct5 Cctgamma Cdc42 Cdc6 cenB1A cenG1A CG10000 CG10026

CG10211 CG10300 CG10359 CG10621 CG10623 CG10638 CG10657 CG10804








CG1444 CG14521 CG14528 CG14605 CG14715 ssp5 CG14801 CG14935





CG18493 CG18609 Cht7 CG1941 CG1942 CG2065 CG2191 CG2246

CG2254 Cg25C CG2663 CG2767 CG2781 CG2811 CG2818 CG2852






























CG9993 Chd1 Chd64 cher Chit chrb Cht2 Cht3

coro Cpr62Bb Cpr62Bc Cpr65Eb Cpr65Ec Cpr67Fa1 Cpr97Ea Cpr97Eb

cry csw cue CycA CycB CycB3 CycE CycK

CycT Cyp1 Cyp304a1 Cyp305a1 cyp33 Cys D d4

Page 33 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

3

DAT dbo Dcp-1 Ddc decay desat1 Dif dl

dnc Doc1 Doc2 Doc3 dom Dox-A3 dpa dpld

dpn Dys EcR EfTuM eIF-4E eIF4E-7 eIF4G Eip75B

Elf emp en enok ERp60 ERR Esp Ets21C

Ets65A eyg Fatp fd96Cb fdl Fkbp13 FKBP59 fkh

flw Fmo-1 Fmo-2 Frq1 Frq2 ftz-f1 fus fz

fz3 fz4 fzy GalNAc-T1 GalNAc-T2 Galpha49B gammaTub23C garz

Gasp GATAe gb Gbeta13F Gbeta5 Gbeta76C Gdh G-ialpha65A

GlcAT-I GlcAT-P glo glu GluRIIA GluRIIC Got1 Got2

Gp93 Gs1 Gs2 G-salpha60A gsb gsb-n HBS1 Hel89B

Hexo1 Hexo2 HLHm5 HLHmbeta Hn HP1b HP1c Hr39

Hr46 Hr78 Hsp26 Hsp27 Hsp83 ia2 Ice Idgf1

Idgf2 Idgf3 Idgf4 iHog ine insv inv inx2

inx3 inx7 Iswi Jafrac1 Jafrac2 Jheh1 Jheh2 JhI-21

Kap-alpha3 kar Keap1 kis kst l(1)G0007 l(1)G0156 l(1)G0232

l(2)09851 l(2)37Cb l(2)44DEa l(2)efl l(2)k01209 Lac lace Lam

LamC lid Lim3 Liprin-alpha Lis-1 Lmpt LpR1 LpR2

LvpD LvpH Mad mask Mbs Mcm2 Mcm3 Mcm5

Mcm7 Mdh mdy mei-P26 Men Mes-4 mib2 MICAL-like

miple miple2 mnd Moca-cyp Mp20 Msp-300 Mtl mtm

mts mys nAcRalpha-96Ab nAcRbeta-21C Nak Nckx30C Nedd4 Nep1

Nep2 Nep4 NetA NetB ninaA ninaD Nmda1 Nmdar1

nrv1 nrv2 nrv3 nub Obp44a Obp99a obst-A obst-B

ofs ogre okr olf413 Optix Orc1 org-1 Osbp

Osi7 Osi9 P5cr path Pde11 Pde8 Pdi pdm2

pdm3 pea PebIII pen pgant2 Pgant35A pgant5 pgant6

PH4alphaEFB PH4alphaPV PH4alphaSG2 Phk-3 ple pod1 pont Pp4-19C

PpD3 PpD5 Pph13 prd Pros25 Pros28.1 Pros35 ProsMA5

Prx2540-2 Prx6005 Ptp4E Ptp61F Ptp69D Ptpmeg put Pxn

pxt Rac1 Rac2 rap rdgC Rel rept RfC40

rhea Rho1 RhoBTB RhoL RpS15Aa RpS15Ab Rx santa-maria

sar1 sax sca scf scramb1 scramb2 scyl Sema-2a

Sema-5c Sep1 Sep5 SerT Set2 Sh Shab Shal

shot Side sim siz sle slmb slp1 slp2

SMC2 Smox so Sox102F SoxN Spn27A Spn4 Spn43Aa

Spn6 Spt-I srp SRPK Sry-alpha st step Su(dx)

Su(var)205 sut1 syt Syt7 SytIV Taf5 Tb T-cp1

Tdc1 TepII TepIV term tgy Tina-1 Tip60 tkv

tld toe tok toy TpnC25D TpnC73F Tps1 Tre1

trh Tsp42Ea Tsp42Eb Tsp42Ec Tsp42Ed Tsp42Ee Tsp42Eh Tsp42Ei

tth tup TwdlB TwdlD TwdlG Uba1 Uba2 Ubx

usp Vha100-1 Vha68-1 Vha68-2 vkg vvl wds wg

wkd Wnt4 wun XNP yin yip2 Yp1 Yp3

zpg

Page 34 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

4

Table S3. FlyBase annotated GO term analysis for BDGP and Fly-FISH gene pairs

showing spatial divergence over a majority of developmental stages.

Category FlyBase Term ID % Genesa P Value

b

Biological anatomical structure formation 89.29% 1.99E-02

Process biological adhesion 100.00% 1.10E-01

biological regulation 85.88% 7.49E-05

cellular process 81.82% 2.20E-08

death 100.00% 2.09E-01

developmental process 83.15% 9.38E-04

establishment of localization 77.27% 1.94E-01

growth 75.00% 4.26E-01

immune system process 90.91% 1.28E-01

localization 87.50% 4.00E-03

locomotion 94.74% 1.50E-02

metabolic process 69.57% 8.87E-02

multicellular organismal process 84.54% 4.34E-05

multi-organism process 100.00% 2.88E-01

reproductive process 90.32% 9.23E-03

reproduction 90.32% 9.23E-03

response to stimulus 81.82% 8.84E-02

rhythmic process 100.00% 3.94E-01

Molecular antioxidant activity 100.00% 3.94E-01

Function binding 77.34% 1.05E-02

catalytic activity 67.11% 2.98E-02

enzyme regulator activity 81.82% 2.46E-01

molecular transducer activity 76.47% 2.26E-01

structural molecule activity 100.00% 3.94E-01

transcription regulator activity 83.87% 6.51E-02

translation regulator activity 100.00% 7.35E-01

transporter activity 90.00% 1.61E-01

a. The percentage represents the subset of gene pairs described by the indicated GO

term of all the genes pairs with predominantly spatial divergence.

b. Hypergeometric P values <1.00E-04 were considered statistically significant.

Page 35 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

For Peer Review

5

Table S4. FlyBase annotated GO term analysis for BDGP and Fly-FISH gene pairs

showing temporal divergence over a majority of developmental stages.

Category FlyBase Term ID % Genesa P Value

b

Biological anatomical structure formation 16.22% 8.46E-02

Process biological adhesion 45.45% 7.00E-02

biological regulation 27.68% 5.91E-02

cellular process 30.96% 1.07E-05

death 11.11% 2.38E-01

developmental process 43.41% 1.39E-12

establishment of localization 23.08% 1.59E-01

immune system process 21.43% 2.51E-01

localization 24.64% 1.27E-01

locomotion 21.74% 1.98E-01

metabolic process 26.87% 6.55E-02

multicellular organismal process 38.06% 1.11E-07

multi-organism process 33.33% 2.28E-01

pigmentation 100.00% 5.70E-02

reproductive process 31.11% 7.16E-02

reproduction 31.11% 7.16E-02

response to stimulus 28.00% 1.09E-01

rhythmic process 33.33% 4.19E-01

Molecular antioxidant activity 33.33% 4.19E-01

Function binding 33.68% 9.85E-10

catalytic activity 22.00% 7.51E-02

electron carrier activity 50.00% 2.01E-01

molecular transducer activity 26.09% 1.90E-01

structural molecule activity 33.33% 2.92E-01

transcription regulator activity 27.03% 1.43E-01

transporter activity 40.74% 2.19E-02

a. The percentage represents the subset of gene pairs described by the indicated GO

term of all the genes pairs with predominantly temporal divergence.

b. as described for Table S3.

Page 36 of 42



123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

comparison of embryonic expression within multigene families...

Documents