DGRPool: A web tool leveraging harmonized Drosophila Genetic Reference Panel phenotyping data for the study of complex traits

Genome-wide association studies have advanced our understanding of complex traits, but studying how a GWAS variant can affect a specific trait in the human population remains challenging due to environmental variability. Drosophila melanogaster is in this regard an excellent model organism for studying the relationship between genetic and phenotypic variation due to its simple handling, standardized growth conditions, low cost, and short lifespan. The Drosophila Genetic Reference Panel (DGRP) in particular has been a valuable tool for studying complex traits, but proper harmonization and indexing of DGRP phenotyping data is necessary to fully capitalize on this resource. To address this, we created a web tool called DGRPool (dgrpool.epfl.ch), which aggregates phenotyping data of 1034 phenotypes across 135 DGRP studies in a common environment. DGRPool enables users to download data and run various tools such as genome-wide (GWAS) and phenome-wide (PheWAS) association studies. As a proof-of-concept, DGRPool was used to study the longevity phenotype and uncovered both established and unexpected correlations with other phenotypes such as locomotor activity, starvation resistance, desiccation survival, and oxidative stress resistance. DGRPool has the potential to facilitate new genetic and molecular insights of complex traits in Drosophila and serve as a valuable, interactive tool for the scientific community.


Introduction
3 to index all existing literature about DGRP phenotyping data -where possible-in order for 68 users to quickly search through the website using simple keywords. We manually associated 69 each study with broad and tailored categories such as "ageing", "metabolism", or "olfactory". 70 We specifically spent important time curating the datasets to avoid any errors or 71 misrepresentations of datasets. To avoid the "maintenance issue" that is common to online 72 tools, and keep the data up to date, we implemented specific curators tools, to help maintain 73 the web application in the future. These tools allow any user to submit a novel dataset, which 74 is then attributed to a curator, in order to manually format and validate all phenotyping data 75 and metadata associated with the study. Importantly, any user can become a curator, as 76 advertised on the main page of the resource, since we strongly believe that such a 77 community-run resource architecture is most optimal to keep a web tool state-of-the-art and 78 allow crowd-based curation work 4 . 79 In addition, we set out to build important tools for the DGRP community such that DGRPool 80 would not only be a static repository for downloading phenotyping data but could also be 81 used as an interactive data analysis tool. For example, users can correlate phenotypes 82 together, from the same study or across studies. We also implemented an automated GWAS 83 analysis (using PLINK2, and known covariates) which we pre-calculated on all the 84 phenotyping data that are currently available. Using this data, users can simply browse 85 through their genes or variants of interest and directly find related phenotypes. A PheWAS 86 page also allows exploration of each variant's impact across multiple phenotypes. Moreover, 87 these tools are applicable to user-submitted phenotypes, so that anyone can upload their 88 own phenotypes to search the DGRPool database for correlated phenotypes or to run 89 GWAS analyses. 90 Our goal is to ensure that DGRP phenotyping data is findable, accessible, interoperable, and 91 reusable (FAIR) 5 to fully leverage the opportunities that stem from this unique genotyping-92 phenotyping resource. To this end, we made user access our priority, both for removing the 93 bottleneck of data harmonization, and also to allow for better, more reproducible research. 94 To showcase the potential of our tool in facilitating new biological discoveries, we conducted 95 a proof-of-concept study focusing on the longevity phenotype, a well-studied trait in 96 Drosophila research with clear relevance to human longevity 6 . By leveraging the data 97 harmonization and curation efforts in DGRPool, we identified multiple phenotypes that are 98 significantly associated with longevity across 18 different studies, such as oxidative stress 99 resistance 7 , sleep duration 8,9 , desiccation survival 10,11 , and starvation resistance 10,12,13 . 100 Interestingly, we also observed correlations between shorter lifespan and certain 101 phenotypes, such as locomotor activity 14 and food intake 15,16 . These results validate prior 102 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint 4 knowledge and illustrate how our tool can provide novel biological insights with just a few 103 clicks. Therefore, we firmly believe that tools such as DGRPool -which ultimately could 104 become entirely community-driven-are essential not only for catalyzing novel research, but 105 also for leveraging the diversity and richfulness of existing datasets. 106

107
A thousand phenotypes across 125 studies 108 To start our data collection, we searched for DGRP studies that reference any phenotyping 109 data and in parallel implemented diverse tools to automatically aggregate these data and 110 their associated metadata from the journals hosting the datasets. However, we quickly 111 realized that it was difficult to automate the entire process. Specifically, the import of 112 phenotyping data proved challenging since i) datasets tended to be hosted in very different 113 formats such as Excel files or PDF, ii) data was stored within the journal's supplementary 114 section, or in external repositories such as Figshare; and iii) the format of the phenotyping 115 data differed from one publication to another. Because of these challenges, we implemented 116 a curation page to manually review, edit, and correct datasets that were automatically 117 aggregated, aiming to prevent errors in the imported datasets. In addition, this allows the 118 curator to add relevant remarks or comments on the study under review, thus providing 119 enhanced context for future analyses of these datasets. 120 In line with the community-resourcing philosophy of DGRPool, we created a specific 121 "curator" role that any logged-in user can claim, again with the underlying rationale of 122 assuring long-term sustainability of our web application. With this role, the user has access 123 to additional functionalities on the DGRPool website, including the modification of any 124 metadata attached to a study (title, authors, description, categories), and the submission or 125 modification of attached phenotypes (see Supp. Figure S1). Although this may entail a 126 considerable amount of time, we assert that this approach is the most effective means of 127 furnishing high-quality data. Consistent with this philosophy, we have incorporated a 128 functionality on the homepage which empowers any user to submit a DOI as a 129 recommendation for a study that could be absent from the DGRPool repository. If the DOI is 130 not in the database, it triggers the same automated scripts that were originally used to 131 incorporate the 125 studies. The corresponding study is then created on DGRPool, and its 132 metadata (authors, links, …) are automatically imported. Once a study has been created, 133 one of three possible labels can be assigned to describe the state of curation of a study: 1) 134 Submitted (default), when no curator is yet assigned to the study, 2) Under curation, when 135 a curator is assigned, and 3) Curated when all phenotyping data and metadata have been 136 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint algorithms. Usually, the data are presented in the form of a matrix, with DGRP lines in rows 189 and phenotypes in columns. But sometimes, they can be in a more "exotic" format 25 , 190 requiring a hands-on approach to format it appropriately. Also, the provided phenotyping 191 data are often not sufficiently self-informative and thus require in-depth reading of the 192 original manuscript to grasp abbreviations or identify the correct measurement units. These 193 are important, in particular, to assure reproducibility, but especially when aggregating 194 multiple studies together such that the scale of the values is similar. In DGRPool, we 195 therefore created a common matrix format to represent all studies, and we implemented a 196 "Unit" metadata for each phenotype. Then, for each study, we mapped all phenotypes to 197 their appropriate format and units (Supplement Figure S3). This part is fully accessible to 198 the curator, who can update or add any phenotype that would be missing, with their 199 corresponding units and meta-data description. 200 Another issue that we faced is that phenotypes are often averaged across multiple individual 201 flies and that the authors only provide these "Summary datasets". This can be problematic in 202 terms of reproducibility, since some figures may show boxplots or distributions of values for 203 each DGRP line, but these plots are not reproducible when only summary data is available 204 (i.e. means or medians). Fortunately, some studies do provide "raw datasets" which contain 205 multiple phenotypic values per DGRP line, often corresponding to replicate flies of the same 206 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint 7 genotype. These values tend to be of much greater interest since they enable statistical 207 analyses and/or the computation of further summary statistics (not only mean or median, but 208 also the standard error of means or other often non-provided summary values). 209 Finally, for some studies, phenotyping data were not or no longer available from the journal's 210 website 26-28 , which is often the journal's responsibility. However, in all cases, we were able 211 to contact the authors directly to recover the missing datasets. 212 To avoid such issues in the future, we have formulated a couple of good practice guidelines 213 for authors to facilitate and improve upon our and future datasets with the aim of enabling 214 harmonized and reproducible research. These guidelines are detailed in the Discussion 215 section of this manuscript. All curated datasets in DGRPool are formatted following these 216 guidelines (where possible), and phenotypes can now be easily downloaded in a standard 217 TSV format from a particular study, or from a phenotype page. 218

How to leverage these datasets by correlating phenotypes 219
Our formatting and harmonizing of all datasets now enables interesting cross-phenotype 220 analyses to generate new biological insights. One strategy to perform such analyses is to 221 download a summary table that contains all the phenotypes in a common format and that is 222 available from DGRPool's front page. However, we deemed this still insufficient as a 223 catalyzing resource, which is why we implemented tools to correlate existing and user-224 submitted phenotypes with all the other phenotypes in DGRPool (Supp. Figure S4). 225 To better understand the structure of these phenotypes, and how they relate together, we 226 also computed a global visualization of the phenotype correlations across all curated studies 227 (Figure 2A, Supp. Figure S5). This revealed a clear trend, with phenotypes belonging to the 228 same study (within-study) correlating in general stronger than those from different studies 229 ( Figure 2B, Supp. Figure S6). This is expected since a given study will typically contain 230 phenotypes that have been acquired for a given research topic, thus they will share 231 similarities. Another potential factor that could explain this similarity is the well-known "batch 232 effect". Indeed, phenotypes acquired in the same environment (same lab, technician, 233 reagents etc.) may sometimes show greater similarity than those acquired across different 234 labs and conditions 29 . The longevity phenotype however, assessed in at least six of the 235 studies in DGRPool 27,30-34 across different laboratories, illustrates that phenotype and its 236 measurements not only exhibits strong correlation across sexes ( Figure 2C), but are also 237 sufficiently robust between laboratories ( Figure 2D). This example illustrates both the high 238 robustness of results acquired in the context of DGRP studies (stable genotype, stable 239 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint 8 environment) and the robustness of the phenotype itself, which highlights its potential high 240 heritability. 241 These examples were all generated using DGRPool phenotype correlation tools, supporting 256 our notion that it can leverage cross-study comparisons of multiple phenotypes to unveil 257 potentially new interesting phenotype interaction/associations. As a further proof of concept 258 and given society's strong interest in defining "healthy aging" determinants 38 , we continued 259 investigating the "mean longevity" phenotype from (Arya et al, 2010) 30 and we selected 50 260 phenotypes that were significantly correlated with it at 25% FDR threshold ( Figure 3C). The 261 hierarchical clustering clearly separated the phenotypes into three clusters: longevity-like 262 phenotypes (strongly correlated together), other longevity-associated phenotypes (correlated 263 with longevity), and phenotypes that seem antagonistic to longevity (anti-correlated 264 phenotypes). Among the phenotypes that positively correlated with longevity, some may be 265 expected such as starvation resistance 10,12,13 and oxidative stress resistance 7 but some are 266 less intuitive such as desiccation survival 10,11 , certain cuticular components of the 267 epicuticle 39 , and sleep duration 8,9 , whose relationship to longevity is complex and still not 268 fully understood 40 . Although we cannot exclude spurious correlations, some of these more 269 surprising correlations appear biologically highly interesting, illustrating the capacity of 270 DGRPool to unveil new research avenues that seem worth exploring in greater molecular 271 detail. Also of interest is the group of often unexpected phenotypes that significantly anti-272 correlates with longevity. These include locomotor activity 14 , some other cuticular 273 components of the epicuticle 41 , and food intake 15,16 , suggesting that higher locomotor activity 274 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

Cross-study correlations highlight phenotype relationships 242
The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint 9 or food intake is linked to reduced longevity. Whether these are direct or indirect links 275 remains unanswered, but appears worthy for a more in-depth scrutiny that is beyond the 276 scope of this paper. 277 Inversely, our analyses also revealed that some expected phenotype correlations could not 278 be detected. For example, in the context of metabolic energy expenditure 42 , it might seem 279 intuitive that higher activity 43 would lead to greater food intake 44 . However, we did not 280 observe such a correlation. Similarly, higher activity levels may reflect increased mating 281 behaviour 37 , but this was also not observed. These are just a few examples of several cases 282 where expected correlations did not materialize, collectively signifying that the genetic 283 architecture underlying such traits appears inherently complex. 284 These proof-of-concept examples demonstrate in our opinion the utility of the DGRP lines 285 and by extension DGRPool to serve as powerful tools that will facilitate the identification of 286 non-intuitive phenotype correlations and their underlying molecular basis as well as the 287 discovery of putative genotype to phenotype relationships, as detailed below. 288

From phenotypes to associated genotypes 289
The goal of most DGRP phenotyping studies is to eventually be able to link the phenotypes 290 to potentially causal variants or sets of variants 45 . In response, tools like DGRP2 GWAS 291 (http://dgrp2.gnets.ncsu.edu/) 1,2 have been put in place to accommodate geno-phenotype 292 relationship analyses. 293 With the goal of providing an integrative analytical environment, we therefore also 294 implemented GWAS tools within DGRPool, aiming to assist researchers with performing 295 GWAS analyses and interpreting the respective output. Specifically, we precalculated GWAS 296 analyses using PLINK2 on every existing phenotype in DGRPool (see Methods), thereby 297 considering all ~4M available DGRP variants while correcting for six main covariates 298 (Wolbachia status, and five major insertions) 2 . Consequently, users can browse the GWAS 299 results from any phenotype page on DGRPool (Supp. Figure S7). These comprise a 300 QQplot, for assessing the validity of the results, or potentially over-estimated p-values, and a 301 Manhattan plot, for visualizing the significant loci across the D. melanogaster genome. It also 302 displays a table with the top 1000 associated variants and allows the user to download the 303 table of all significant hits, at a p-value<0.01 threshold. The tool further runs an ANOVA 304 between the phenotype and the six main covariates to uncover potential confounder effects 305 (prior correction), which is displayed as a "warning" table to inform the user about potential 306 associations of the phenotype and any of the covariates. The interface also allows plotting 307 an independent boxplot for each variant to visualize the effect of each allele on the 308 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint phenotype. Importantly, for each variant, we also implemented a PheWAS button to visualize 309 the effect of a particular variant over all existing phenotypes in DGRPool. We also annotated 310 all the variants for impact (non-synonymous effects, stop-codon gain, etc.) and for potential 311 regulatory effect (transcription factor binding motif disruption), which should assist 312 researchers with prioritizing the variants in terms of potential consequences. For all of these 313 variants, we also provide links to their description in Flybase 4 . 314 As mentioned, these GWAS results are available for each existing phenotype in DGRPool, 315 directly from the phenotype's page. But users can also submit their own phenotype files 316 (through the 'Tool' menu in the header), and visualize the same information for their own 317 phenotypes. The GWAS analysis runs in the backend and takes about 1-2 minutes before 318 displaying the results. This is implemented using a queuing system which prevents 319 overloading the server in case of a peak of users or requests. 320 After having run GWAS on all phenotypes in DGRPool, we observed the distribution of the 321 number of significant variants per phenotype at p -5 threshold, which is an often used 322 arbitrary threshold for GWAS analyses in DGRP studies ( Figure 4A). This threshold yields 323 on average 382 significant hits per tested phenotype, which is skewed due to some 324 phenotypes leveraging lots of results (median = 38). Conveniently, this threshold seems 325 sufficient for avoiding an over-abundant number of false positives, as is clearly visible from 326 other, less stringent, thresholds (Supp. Figure S8). Another very often used threshold, is the 327 Bonferroni one, which is much more stringent and varies from p -8 ) yielded 73 330 significant hits on average (median = 0, Supp. Figure S8) which could be limiting for many 331 studies as it may mask potentially interesting variants that, while minimally contributing on an 332 individual basis, may collectively point to implicated pathways or biological processes 46 . 333 Thus, while choosing an optimal threshold is in general challenging, our results indicate that 334 any threshold below 1 x 10 -5 is reasonable given that at this threshold, the p-values appear 335 not over-estimated, as observed on the respective QQplots. We also verified if any variant is 336 over-selected across all phenotypes to uncover a possible bias in our studies (Figure 4B), 337 but we did not find such variants, even at different thresholding values (data not shown). 338 As a proof-of-concept and a validation of our approach, we compared our results with a 339 randomly selected study that identified several variants associated with survival to azinphos-340 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint methyl at different doses (0.25, 0.5, 1, and 2 µg/ml) 26 . Of note, this study is available in 341 DGRPool under https://dgrpool.epfl.ch/studies/3. In particular, this study showed that 342 survival to azinphos-methyl is highly variable among DGRP lines, even at a "low" 0.25 µg/ml 343 dose. Importantly, the results of this study are reproduced in DGRPool as can be observed 344 on the respective phenotype's page (https://dgrpool.epfl.ch/phenotypes/20, Figure 4C) LTR insertion is associated with increased resistance to organophosphates, suggesting that 355 derived alleles of Cyp6g1 confer organophosphate resistance in the DGRP (Figure 4E). 356 These results show that DGRPool is able to accurately reproduce results from existing 357 studies, and that new biological findings can be leveraged from its interactive results and 358 plots. Revisiting the same organophosphate study 26 , the PheWAS page present in the 359 GWAS results shows that this top variant is not only significant at other doses, but that it is 360 also significant in the context of other studies, in particular one study on cuticular 361 hydrocarbon composition 23 , and another study investigating Drosophila microbiota 22 . This 362 could help with fine-tuning putative causal variants, but also with uncovering potential 363 associations between certain phenotypes that in turn could enable studies aimed at 364 providing underlying genetic and molecular mechanisms. 365

Extreme phenotypes 366
After having collected and harmonized thousands of DGRP phenotypes, we investigated if 367 we could identify outliers amongst DGRP lines that would potentially bias phenotypic 368 associations. Indeed, if a particular DGRP line is repeatedly ranked in the extreme of all 369 phenotypes, it could be that there are unknown cofactors that make the line "weaker" in 370 general, or inversely. Although it is difficult to judge what phenotype is particularly 371 advantageous or disadvantageous due to the presence of potential trade-offs 50,51 , we can 372 determine how often a DGRP line is in the top or bottom 15% of a given phenotype. By 373 focusing on phenotypes that are likely impacting overall viability, we ranked DGRP lines for 374 each associated phenotype. Upon ranking the DGRP lines, we calculated whether the rank 375 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint 12 falls within the top or bottom 15% performers of the phenotype. We then assessed for each 376 DGRP line how often they are 'extreme' and divided this by the total number of phenotypes 377 in which the DGRP line has been included to obtain a "fraction of extremeness" (FoE). 378 Finally, we filtered for lines which had at least 50 phenotypic measures available to ensure 379 that our values were not driven by a low number of observations ( Figure 5A) however, their counterpart may not be as extreme or moderate. We therefore also looked for 394 DGRP lines which can be considered extreme in both females and males, and are 395 potentially more extreme on a population-wide basis. Figure 5D describes such populations 396 where the overall fraction of extremeness between males and females differed on average at 397 most 0.05. In these cases, DGRP_852 and DGRP_042 are more likely to be extreme across 398 sexes, which may be attributed to at least two factors. First, this may indicate that the 399 population is generally not healthy if they consistently display a low lifespan, or second, and 400 conversely, well-documented trade-offs of life history traits such as lifespan vs fecundity may 401 be strongly at play here. The former does not however seem to be the case, as shown in 402 There are many studies across organisms where collated phenotyping data has led to novel 413 insights 52,53 . Even though the Drosophila Genetic Reference Panel was formally released 414 more than ten years ago, the resulting phenotype data of over 100 studies has so far not 415 been combined into a single accessible resource. We anticipate that providing wider access 416 to this data, as driven by FAIR principles 5 , will therefore facilitate our general understanding 417 of the relationship between genotypes and phenotypes. 418 We have previously shown that using a subset of this resource effectively enabled us to 419 establish a relationship between mitochondrial haplotypes and feeding behavior which we 420 and sleep 43 . Therefore, we believe that DGRPool will either aid with validating the findings of 430 a given study (i.e. higher bacterial resistance linked to overall resistance phenotypes) or by 431 placing a study's phenotype data into a wider context (for example, linking brain size to 432 behavioral phenotypes). 433 Moreover, having access to multiple studies studying similar phenotypes can also be of help 434 for meta-analyses and increased statistical power. In the case of longevity for example, there 435 are multiple studies that aggregated this phenotype, across similar or complementary DGRP 436 lines. Therefore, one could conduct a meta-GWAS analysis 61 by leveraging the replicates or 437 combining the different lines into a single dataset. This tends to be a challenging process 438 given the need for data harmonization and curation, which is exactly what we aimed to 439 address by establishing with DGRPool. Of course, since similar DGRP lines across 440 laboratories still have the same genotype, they should not be treated as biological replicates, 441 but phenotypes could be averaged across similar lines, which would reduce hidden 442 covariates such as laboratory adaptation or batch effects. Moreover, complementary lines 443 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made DGRPool. This is because such lines have often been included as controls in DGRP 460 studies 34 , and for most of these, genomic information is also available. 461 Finally, in order to sustain the value of the DGRP as a resource and to promote more 462 findings, we provide the following guidelines for future DGRP phenotyping studies: 463 was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint modes respectively, or Jquery autocomplete for phenotype search combined with a SOLR 510 search engine running on the server side (used for the phenotype comparison tool). 511

Semi-automated referencing of studies and/or phenotypes 512
To submit a new study, any user can submit a DOI from the front page. Then, all metadata 513 associated with this study (authors, journal, date, …) are automatically imported from the 514 Crossref 65 API. When the study is created, it acquires the "Submitted" state, and 515 administrators are notified. Then, a curator is assigned to the study and needs to manually 516 verify all information. A specific curator page allows him/her to 1) edit the metadata, 2) edit 517 the categories associated with the study, or 3) add/remove/modify the phenotyping data and 518 edit their names/types/units. 519 Identifiers from GEO 66 , ArrayExpress 67 , or the Sequence Read Archive (SRA) 68 can be 520 associated manually with any study, for example for referencing additional gene expression 521 data that would be published along with the phenotyping data. 522

Phenotypes correlated with longevity 523
We computed the correlation of the "mean longevity" phenotype from ( 12'591 MNPs) with options "--glm --geno 0.2 --maf 0.05". We corrected the model for six 533 main covariates (Wolbachia status, and 5 major insertions) that were described in 2 and also 534 used on the DGRP2 website. Of note, these covariates are phenotypes, and thus are also 535 available as a separate, browsable study on DGRPool (https://dgrpool.epfl.ch/studies/17). 536 Extremeness 537 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint Fraction of extremeness was calculated for each phenotypic spectrum separately by ranking 538 the values with ties being assigned the minimum rank. We then calculated a cut-off to assign 539 ranks in the bottom or upper 15% of a phenotypic range. This rank cut-off was further 540 rounded up to be more inclusive on either end (i.e. if the cut-off was 1.2 or 1.8, the cut-off 541 would become 2). Phenotypes equal or lower than the cut-off were assigned -1, whereas 542 phenotypes equal to the max rank minus the cutoff or higher were assigned 1. was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint Author contributions 566 RB and BD initiated the project. VG, RB and BD wrote the article. RR implemented the 567 automatic pipeline to retrieve phenotypic data from articles. VG and ER curated the studies. 568 FPAD designed and implemented the web application and its database. FPAD designed and 569 set up the unified format to represent phenotype data. FPAD and VG implemented the 570 different tools (GWAS, PheWAS, Correlation). VG tested the web application extensively. 571 VG and RB performed supporting analyses (e.g. GWAS, extremeness analysis). 572

Competing interests 573
The authors declare that they have no conflict of interest.
. CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint D rosophila 718 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. plots were generated using the "phenotype correlation" tool in DGRPool. 753 and B plots were generated using the "phenotype correlation" tool in DGRPool. 761   was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint

Supplement Figures 802
Supplemental Figure S1. Screenshot from the curator's view for a given study -Metadata 803 section. This screenshot shows the metadata section of the editing page for a study, where the 804 curator can edit any of the fields. We expect the curator to set a description (short abstract) for the 805 study, and associate some categories. The curator can also deactivate a phenotype if he/she 806 considers that it is not a proper phenotype (like the number of replicates). Once the curation is done, 807 the "Status" field can be changed to "Validated", which signifies that the curation process is finished, 808 allowing the study to be widely visible to the users.

844
. CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

855
. CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023.  was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted June 5, 2023. ; https://doi.org/10.1101/2023.06.01.543194 doi: bioRxiv preprint