Challenges and opportunities for drug repurposing in cancers based on synthetic lethality induced by tumor suppressor gene mutations

9 Although two-thirds of cancers arise from loss-of-function mutations in tumor suppressor genes, there 10 are few approved targeted therapies linked to these alterations. Synthetic lethality offers a promising 11 strategy to treat such cancers by targeting vulnerabilities unique to cancer cells with these mutations. 12 To identify clinically relevant synthetic lethal interactions, we analyzed genome-wide CRISPR/Cas9 13 knock-out (KO) viability screens from the Cancer Dependency Map and evaluated their clinical 14 relevance in patient tumors through mutual exclusivity, a pattern indicative of synthetic lethality. Indeed, 15 we found significant enrichment of mutual exclusivity for interactions involving cancer driver genes 16 compared to non-driver mutations. To identify therapeutic opportunities, we integrated drug sensitivity 17 data to identify inhibitors that mimic the effects of CRISPR-mediated KO. This approach revealed 18 potential drug repurposing opportunities, including BRD2 inhibitors for bladder cancers with ARID1A 19 mutations and SIN3A-mutated cell lines showing sensitivity to nicotinamide phosphoribosyltransferase 20 (NAMPT) inhibitors. However, we discovered that pharmacological inhibitors often fail to phenocopy KO 21 of matched drug targets, with only a small fraction of drugs inducing similar effects. This discrepancy 22 reveals fundamental differences between pharmacological and genetic perturbations, emphasizing the 23 need for approaches that directly assess the interplay of loss-of-function mutations and drug activity in 24 cancer models. 25 26 27 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 2 Author Summary 28 Synthetic lethality is an emerging approach for targeting a biological dependency in cancer cells that 29 does not harm normal cells. This strategy is particularly valuable for targeting loss-of-function mutations 30 in tumor suppressor genes, which are more challenging to directly target. In an effort to accelerate 31 treatments for cancer patients, we aimed to map out these dependencies and overlap them with 32 responses to available drugs. We discovered different outcomes when a protein is targeted by a drug 33 versus when that same target is disrupted genetically. Thus, if a drug is to be effectively repurposed as 34 synthetic lethal agent, feasibility studies must capture drug biology, ideally by test the drug empirically 35 in relevant cancer models. A second notable discovery is that in vitro synthetic lethal interactions 36 involving cancer driver genes are significantly more likely to exhibit consistent patterns, such as mutual 37 exclusivity in human tumor samples. This is important since selection of relevant cell lines is often 38 critical in drug development to maximize potential for translation to clinical responses. 39 40 1 Introduction 41 Synthetic lethality (SL) provides a compelling framework for developing cancer therapies tailored to the 42 specific genetic profiles of individual patient tumors. SL relies on the genetic interaction between two 43 genes where disruption of either gene alone is tolerated, but their simultaneous disruption leads to cell 44 death. Applying this strategy in oncology usually involves pharmacologically targeting a protein whose 45 inhibition is lethal to cancer cells harboring a specific loss-of-function (LoF) mutation in a partnering 46 gene, but spare healthy cells without these mutations (S1A Fig). The appeal of exploiting SL lies in the 47 potential to increase treatment efficacy, minimize adverse effects and expand the range of actionable 48 therapeutic targets, particularly tumor suppressor genes (TSGs) that currently lack effective therapies 49 [1,2]. An exemplar of this approach is the use of PARP inhibitors, which harness SL to exploit DNA 50 repair deficiencies in homologous recombination deficient (HRD) cancers, commonly driven by LoF 51 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 3 alterations in BRCA1 and BRCA2 [3]. Generating over $3 billion annually, PARP inhibitors have 52 become standard of care treatment for several HRD cancers, and are increasingly used in combination 53 therapies that prove the value of exploiting the potential of SL-driven oncology. 54 Despite the success of PARP inhibitors, broader clinical adoption of additional SL drugs has not been 55 as effective as predicted and is limited to a handful of clinical-stage developments, none of which have 56 yet been approved. However, progress may have been limited by a lack of systematic SL mapping 57 abilities in human cells. Recent high-throughput screening efforts hold promise for expanding SL 58 treatments, allowing unbiased target exploration across the genome. Genome-wide CRISPR and RNA 59 interference-based genetic screens, including those from the Cancer Dependency Map Project 60 (DepMap), OnTarget, and Project DRIVE, have substantially expanded the repertoire of potential 61 therapeutic targets.[4] Since it’s public release in 2018, DepMap has become an essential tool for 62 hypothesis generation, facilitating the translation of numerous SL interactions into clinical trials. 63 Ongoing trials include targeting PRMT5 (NCT05732831; Tango Therapeutics), and MAT2A in MTAP-64 deleted cancers (NCT04794699; IDEAYA Biosciences), and WRN in MSI or mismatch repair deficient 65 tumors (NCT05838768; Novartis). As novel actionable SL targets and corresponding therapeutic 66 agents are identified, the potential of precision oncology advances toward increasingly meaningful 67 clinical impact. 68 Numerous studies have developed methods for identifying additional SL targets within genetically 69 defined contexts. Approaches range from univariate tests [5–8], linear models such as mixed-model 70 regressions [9], ANOVA [10], and machine learning techniques including random forests (RF) [11–16], 71 neural networks [17] and autoencoders [18,19]. Among these, RF models have gained prominence for 72 their capacity to identify molecular biomarkers linked to gene essentiality and drug sensitivity in cancer 73 cell lines. RF models capture complex, multivariate, and non-linear relationships that univariate tests 74 and linear models often fail to detect. RFs can account for interactions between variables, offer 75 interpretable feature importance rankings, and provide insights into complex biological systems. 76 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 4 Additionally, they are computationally efficient and adept at integrating diverse omics data alongside 77 confounding variables, making them especially well-suited for analyzing high-dimensional datasets, 78 such as those produced by the DepMap project [11]. 79 Despite methodological advances, statistical models often yield thousands of potential SL interactions, 80 necessitating the use of additional filters to enrich for the most clinically actionable interactions. To 81 achieve this, previous studies have leveraged biological pathway data [7,20], protein-protein 82 interactions [9,10], drug sensitivity profiles [5,20] and clinical information from The Cancer Genome 83 Atlas (TCGA) such as tumor gene expression [5], survival statistics [5] and mutual exclusivity [21,22]. 84 Mutual exclusivity is a powerful readout for filtering SL interactions using patient tumor mutations [23]. 85 The concept is based on the observation that certain genetic alterations rarely co-occur in tumors, 86 implying negative selective pressure on their coexistence (S1B Fig). Mutual exclusivity is thus a clinical 87 indicator for SL. Focusing on gene pairs or pathways that exhibit mutual exclusivity in tumors can 88 enrich potential in vitro SL interactions that are more likely to translate into therapeutic benefit. 89 Recognizing that drugs can elicit complex pharmacological responses on cell biology, several initiatives 90 have systematically screened drugs across extensive cancer cell line panels. Projects such as PRISM 91 (Profiling Relative Inhibition Simultaneously in Mixtures) [12], GDSC (Genomics of Drug Sensitivity in 92 Cancer) [24], and CTD2 (Cancer Target Discovery and Development) [5] have significantly advanced 93 our understanding of drug response across a diverse range of cancer types. These datasets enable the 94 modelling of relationships between molecular features and drug response [9,12,25], providing valuable 95 insight into mechanisms of action (MoA) as well as biomarkers of drug sensitivity and resistance 96 [26,27]. Integrating drug sensitivity profiles with SL interactions can highlight cases where 97 pharmacological inhibition and genetic perturbation of the same gene/protein induces similar synthetic 98 lethal effects, thus supporting the reliability of SL interactions. 99 In pursuit of robust pharmacogenomic interactions and potential drug repositioning opportunities, two 100 previous studies have conducted comprehensive comparisons of drug and CRISPR viability. Gonçalves 101 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 5 et al integrated drug sensitivity profiles from the GDSC with DepMap gene viability profiles, finding that 102 approximately 25% of 376 targeted oncology drugs had effects consistent with CRISPR deletion of the 103 same targets [9]. Similarly, using PRISM drug sensitivity profiles, Corsello et al. [12] showed that 15% 104 of active targeted oncology drugs and 0.8% of active non-oncology drugs displayed sensitivity profiles 105 resembling CRISPR KO of their intended targets. These findings suggest that, while pharmacological 106 inhibition and genetic disabling of a gene can produce similar outcomes, it is more common that the two 107 modes of targeting a gene product differ, which violates a fundamental assumption in SL studies. 108 Recognizing that there are typically differences between pharmacological inhibition and genetic KO of a 109 target is an important step in assessing the potential of drug repositioning, such as incorporating more 110 elements of drug biology. 111 In this study, we evaluate the principles underlying SL interactions identified from DepMap pooled 112 CRISPR KO screens to assess their potential as druggable opportunities, with a specific focus on 113 uncovering novel genetically targeted applications for existing clinical grade drugs. In contrast with 114 earlier studies, we found that only a small fraction of SL interactions identified in CRISPR KO screens 115 are likely to be replicated in tumors, with a significantly higher probability for interactions involving driver 116 genes. For the 6,550 drugs tested in cell viability screens, concordance with genetic KO was 117 unexpectedly low, highlighting the need for repurposing approaches that effectively capture drug 118 biology, such as empirical pharmacogenomic screens or additional mechanistic insights. 119 120 2 Results 121 Constructing a driver gene-centric network for repurposing and targeting cancer vulnerabilities 122 Genetic dependencies network 123 Our approach aimed to identify mutation-specific vulnerabilities in cancer cell lines by stratifying 124 frequently mutated genes based on mutation status and assessing differential fitness effects across 125 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 6 genome-wide gene KO’s. Systematically evaluating each mutated gene against all potential KO’s 126 enabled us to build a network of genetic dependencies central to our analyses. We employed an 127 adaptation of the Pan-Cancer Inferred Synthetic Lethalities (PARIS) framework [13] which applies the 128 Boruta algorithm – a Random Forest-based feature selection method designed to capture all significant 129 predictive features. Through our modified framework, we identified mutations associated with enhanced 130 sensitivity or resistance to gene inhibition, systematically mapping critical genetic dependencies across 131 diverse cancer cell line models without the need to exhaustively test each gene pair. This approach 132 enables a precise and scalable characterization of mutation-driven dependencies, enabling us to 133 explore potential therapeutic targets and drug repurposing opportunities. 134 Our analysis utilized input data from DepMap (release 23Q2) which included detailed mutation profiles 135 and gene essentiality scores obtained from CRISPR LoF (CRISPRi) screens across 1,087 cancer cell 136 lines. Gene essentiality scores – which reflect the impact of gene KO’s on cell proliferation – served as 137 response variables. The mutation data, including both hotspot and damaging mutations with detailed 138 zygosity information were used as predictor variables. We performed this analysis across all cell lines 139 with complete mutation and essentiality profiles, identifying interactions both in a pan-cancer context 140 and within 13 specific cancer types. After adjusting for confounders including microsatellite instability 141 and cancer type using linear regression (FDR < 0.1) and excluding interactions with insufficient 142 mutation data (< 4 mutations), our final pan-cancer network contained 1,428 significant genetic 143 interactions (details provided in Methods). To clarify terminology, within an SL interaction, we define 144 the mutated driver gene in the cell line as the ‘source’ gene and the gene targeted by CRISPR KO as 145 the ‘target’ gene. Comprehensive details of all interactions, both pan-cancer and cancer-type specific, 146 are documented in Table S1 and Table S2. 147 Identification of clinically relevant SL interactions 148 Since the source data was generated in cancer cell lines, which rarely recapitulate the complexity of 149 primary tumors [28–30], we were interested in exploring parameters that predicted which SL 150 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 7 interactions are more likely to translate into clinical benefit on the basis of tumor biology. We reasoned 151 that genetic interaction with driver genes [31–33], which are mutated at epidemiologically 152 overrepresented levels in patients with cancer, may be more likely to conserved since cell lines tend to 153 be dependent on their driver mutations [34]. As these genes are also more frequently mutated in patient 154 tumors, SL interactions could reveal treatment opportunities with significant potential for patient benefit. 155 To evaluate whether driver gene SL interactions are more likely to be conserved in tumors, we 156 analyzed mutual exclusivity (ME) using TCGA tumor mutation and copy number alteration data (Fig 1). 157 ME is based on the observation that certain genetic alterations co-occur less frequently than expected 158 in patient tumors, suggesting an intrinsic incompatibility and potential SL relationship (Fig S1B). As a 159 clinical indicator for SL, ME allows us to focus on gene pairs with greater translational potential. 160 Accordingly, for each cancer type we applied hypergeometric tests to assess ME among the statistically 161 powered gene interactions within our network (Methods). As hypothesized, in vitro SL interactions 162 involving driver mutations demonstrated significantly stronger ME signals in human tumors (P < 0.05) 163 compared to those lacking drivers (Fig 2C), even when controlling for power differences arising from 164 higher mutation rates of drivers versus non-drivers (Fig S2). This observation supports the central role 165 of driver mutations in dictating mutual exclusivity patterns and confirms that in vitro SL interactions 166 containing drivers are biologically relevant in clinical settings. 167 Network visualization of the genetic interactions revealed central hubs enriched by driver genes, such 168 as TP53, APC, RB1 and Ras/Raf/MAPK pathway oncogenes (Fig 2A). Given the clustering of central 169 hubs around common driver genes and the stronger ME signals observed in driver interactions relative 170 to non-driver interactions, we refined the network to focus exclusively on interactions involving ‘source’ 171 driver mutations - mutations pre-existing in the genomic profiles of the cell lines. These are interactions 172 involving a driver mutation inherent to a cancer cell line. This subset confined the network to 510 gene 173 pairs across 448 unique genes, including 46 driver genes (Fig 2B). We prioritized this subset to reduce 174 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 8 the search space and to concentrate on interactions more likely to have clinical relevance and 175 translational potential. 176 To improve interpretability, we introduced directionality to the genetic interactions using the correlation 177 coefficient; negative values indicated increased sensitivity in the mutated state, while positive values 178 indicated sensitivity in the wild-type. For drug repurposing purposes we were interested in SL 179 interactions, specifically cases where mutated cell lines showed heightened sensitivity to partner gene 180 KO. 181 Significant driver genetic interactions in the pan-cancer network were depicted in a heatmap (FDR < 182 0.1; Fig 3A), where numerous established dependencies were evident. These included self-paired 183 oncogenes such as BRAF (Box 1) and PIK3CA (Box 4), paralog dependencies such as SMARCA2-184 SMARCA4 (Fig 3C; Box 3), STAG1-STAG2 (Fig 3C; Box 2), RPL22-RP22L1, PDS5A-PDS5B, and 185 ARID1A-ARID1B, and other established pathway dependencies such as SHOC2-NRAS, PTPN11-186 BRAF, CREBBP-EP300 and RB1-CDK6 (Box 3). The enrichment of well-known, and experimentally 187 confirmed genetic dependencies gave us confidence that we were capturing biologically relevant signal 188 with power to identify clinically relevant cases downstream in our effort to repurpose cancer therapies 189 exploiting SL interactions. Our analysis revealed extensive SL and resistance interactions (Fig S3). 190 Overlap with PRISM 191 After establishing our driver-focused genetic interaction network, we integrated 6,550 drug sensitivity 192 profiles (6,337 unique drugs) from the PRISM 23Q4 dataset, encompassing two distinct screening 193 efforts: Repurposing-Primary (n = 5362) and recently released Repurposing-1M (n = 1278). Drug-gene 194 associations were mapped based on putative gene targets derived from Citeline and PRISM drug target 195 databases. Among the pan-cancer interactions we established, 20.5% (104/510) had at least one drug 196 with a matching target in the PRISM dataset. To explore the behavior of drug profiles targeting genes 197 implicated in our genetic interactions, we implemented two analytical strategies: (i) examining the linear 198 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 9 relationship between drug sensitivity and CRISPR KO effect scores, and (ii) assessing drug sensitivities 199 in cell lines stratified by driver mutation status (Fig 1). Both analyses – adjusted for microsatellite 200 instability (MSI), cell growth rates, and cancer types – aimed to identify interaction pairs where the drug 201 effect on cell proliferation mirrored that of gene KO. While the mutation-stratified analysis is limited by 202 smaller sample sizes, particularly in cancer-specific subsets and for less common driver mutations, it 203 provides direct insights into context-specific therapeutic potential with translational value. In contrast, 204 the linear correlation approach leverages the entirety of the dataset, providing better statistical power to 205 detect strong and generalizable dependencies. 206 We initially assessed correlations between drug sensitivity and gene essentiality for druggable targets 207 in pan-cancer genetic interaction pairs. Among these targets, 40.3% (42/104) displayed a significant 208 linear correlation between drug sensitivity and CRISPR KO (FDR < 0.05; S3 Table), indicating that 209 these genetic dependencies may be pharmacologically targeted to induce SL. Notably, many of these 210 'druggable SL' pairs involved well-characterized cancer genes, such as MDM2, TP53, MCL1, BRAF, 211 PTPN11, PIK3CA, EGFR, BCL2L1, CDK6, and ERBB2, for which targeted therapies are already 212 available, serving as positive controls. For example, dependencies on MDM2 and TP53 strongly 213 correlated with sensitivity to various MDM2 inhibitors, underscoring the parallel effects of 214 pharmacological inhibition and gene KO on MDM2-mediated p53 repression. Other key correlations 215 were observed with cell cycle-related genes, such as MCL1 dependencies with MCL1 inhibitors 216 (AZD5991, AMG-176), CDK6 with CDK4/6 inhibitors (Ribociclib, Palbociclib, Trilaciclib), and BCL2L1 217 with BCL inhibitors (Navitoclax, ABT-737; Fig S4). Furthermore, dependencies in epigenetic 218 regulators—such as EP300 with the p300/CBP inhibitor A-485 and BRD2 with BET inhibitors including 219 OTX015, TEN-010, and CPI-203 – highlight the potential for therapeutic intervention in these drug-gene 220 relationships (Fig S4). 221 However, across the entire PRISM database, correlations between drug sensitivity and CRISPR KO 222 essentiality of drug targets were rare. Of all PRISM drugs, only 154 (3%) displayed a significant 223 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 10 correlation with any target gene essentiality profile (FDR < 0.1; Table S3). Additionally, for each drug, 224 we compared the correlation coefficient between drug sensitivity and effect scores for all putative target 225 genes against a distribution of randomly permuted effect scores (Fig 4). In the majority of cases, 226 observed correlations did not significantly deviate from the permuted distribution, with the exception of 227 a skewed tail of highly correlated pairs, mostly from highly targeted oncology drugs. 228 These findings suggest that while CRISPRi screens are effective at identifying essential genes, the 229 genetic dependencies they reveal may not directly correlate with drug efficacy. Small-molecule 230 inhibitors typically target only the catalytic function of proteins, which can activate compensatory cellular 231 mechanisms, context-dependent dependencies, and off-target effects, complicating the translation of 232 genetic vulnerabilities into therapeutics. Furthermore, the promiscuity of many drugs, with multiple 233 putative targets, adds an additional layer of complexity. From a clinical translational perspective, 234 genetically-based predictions (e.g., SL interactions) are insufficient on their own to justify drug 235 development and rigorous validation through rigorous drug testing in model systems is essential to 236 better validate their therapeutic potential in humans. 237 We then investigated whether cell lines with specific driver mutations demonstrate consistent 238 responses to both genetic deletion and small-molecule inhibitors compared to their wild-type 239 counterparts. We addressed this by analyzing drug sensitivities in cell lines stratified by driver mutation 240 status, examining both hotspot oncogene mutations and deleterious mutations in TSGs across 59 241 driver genes (Table S1). Incorporating MSI as a confounder in the regression model was crucial for 242 accurately evaluating these associations, particularly for AKT inhibitors, which demonstrated high 243 efficacy in MSI-positive lines regardless of driver mutation status. Within our driver-focused SL network, 244 36 interaction pairs exhibited significant, directional agreement with drugs targeting the corresponding 245 protein products (FDR > 0.2; Table 1). As expected, several of the strongest associations involved 246 targeted oncology drugs in their anticipated genomic contexts—such as MDM2 inhibitors in TP53 wild-247 type lines, CHK inhibitors in TP53-mutated lines [35], BRAF inhibitors in BRAF- 248 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 11 Table 1. Driver-target interaction pairs showing significant directional agreement between genetic dependencies 249 and drug sensitivities (FDR > 0.2). Each row lists the driver gene, target, interaction type (synthetic lethal or 250 resistant), genetic effect, and corresponding drug effect size and targets. Color gradients indicate the direction 251 and strength of genetic and drug effects. 252 mutated lines, and CDK4/6 inhibitors in RB1 wild-type lines. Furthermore, BRAF-, NRAS-, and KRAS-253 mutated lines exhibited marked resistance to PTPN11/SHP2 deletion and SHP2 inhibitor RMC-4550, 254 but not other SHP2 inhibitors. PTEN-mutated lines showed heightened sensitivity to both PIK3CB 255 deletion and several pan-PI3K inhibitors. One potential novel interaction emerged, as SIN3A mutated 256 cell lines were sensitive to nicotinamide phosphoribosyltransferase (NAMPT) deletion and associated 257 NAMPT inhibitor GMX1778, although functional links are unclear. 258 Next, we broadened our analysis to investigate all mutation-specific drug dependencies across the 259 PRISM database, aiming to uncover how often these sensitivities arise beyond our predefined genetic 260 interaction pairs. Through this expanded analysis, we identified 44 significant interactions associated 261 with tumor suppressor gene mutations and 96 with oncogenic mutations (Table S4). Our findings 262 confirmed several known vulnerabilities, such as sensitivity to the Aurora kinase inhibitor LY3295668 in 263 SMARCA4-inactivated lines [36], and the SL of bafetinib – originally developed as a BCR-ABL and LYN 264 tyrosine kinase inhibitor – in BRAF-mutated cell lines. Notably, the latter association is supported by 265 recent target deconvolution studies identifying BRAF as an off-target of bafetinib [37]. This validation 266 underscores the utility of our approach in uncovering actionable vulnerabilities across diverse driver 267 mutations and drug combinations. Additionally, our analysis identified potentially novel drug 268 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 12 repurposing opportunities, including the sensitivity of KDM6A-mutated cancer cell lines to the enoyl-acyl 269 carrier protein reductase FABL inhibitor MUT056399 (Damaging; FDR = 0.00405) and the sensitivity of 270 FBXW7-mutated lines to PORCN inhibitor IWP-L6 (Hotspot; FDR = 0.025). FABL is a key enzyme in 271 the bacterial fatty acid biosynthesis (FAS II) pathway that has garnered interest in antibacterial drug 272 development. However, metabolic dependencies, such as those disrupted by FABL inhibitors, are 273 increasingly recognized in epigenetically mutated cancers, suggesting potential therapeutic relevance 274 [38]. 275 Cancer type-specific analysis 276 We transitioned our efforts towards cancer type-specific interactions to identify context-dependent 277 opportunities for drug repurposing, as these may better reflect the unique genetic and tumor 278 microenvironment landscapes of different cancer types. Previous experimental knockout studies have 279 consistently shown that SL gene pairs tend to exhibit cancer type specificity [39]. To explore these 280 interactions, we repeated the Boruta genetic dependency analysis across 13 cancer type subsets 281 (Table S2), yielding an average of 4,528 interactions per cancer type. To prioritize the highest-value 282 interactions, we applied stringent filtering criteria, including driver mutation status, mutation sample size 283 (n ≥ 4), and clustering based on importance scores using the heads/tails break algorithm, as previously 284 described. This algorithm, specifically designed to address heavily tailed distributions, allowed us to 285 isolate the most significant interactions. After filtering, 2,500 interactions remained across 13 cancer 286 types, with 379 interactions including a druggable target (Fig S5). 287 Apart from oncogene addiction-based SL interactions involving KRAS, and PIK3CA, and MDM2-TP53 288 interactions, most high-confidence SL interactions were rarely shared across multiple cancer types (Fig 289 S6). This is likely due to context-specific dependencies, but it also reflects the challenges posed by 290 insufficient sample sizes in some cancer types. Many cancer types lack the statistical power to identify 291 mutation stratified dependencies in cell lines screened for both KO and drug sensitivity assays. In many 292 cases, only a few highly mutated genes can be tested. Nevertheless, we identified some interactions 293 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 13 that merit experimental follow-up. PBRM1-mutated renal cancer lines show resistance to BRD2 KO and 294 BET inhibitors. Similar findings have been made in triple-negative breast cancer models using CRISPR 295 KO and small molecule inhibitor screens [40]. 296 Within our bladder cancer SL network, the interaction between ARID1A and BRD2 emerged as an 297 interesting candidate (Fig 6A). BRD2 has been previously highlighted as a therapeutic target in 298 ARID1A mutated gastric cancer [41], and ovarian clear cell carcinoma [42], suggesting it could also be 299 of clinical interest in bladder cancer as well. The ARID1A-BRD2 interaction was ranked on average 9.6 300 out of 47 significant SL interactions with matching drug targets in bladder cancer, combining ranks of 301 the three importance algorithms (Table S2). Additionally, BRD2 knockout phenocopied BET inhibitor 302 sensitivity in bladder cancer cell lines as BRD2 essentiality and BET inhibitor sensitivity are strongly 303 correlated in both (R = 0.60; FDR = 8.15e-05; ). In agreement with the CRISPRi interaction, ARID1A 304 mutated bladder cancer lines were significantly more sensitive in 11 of 19 BET inhibitors – an effect that 305 was not observed in other cancer types with recurrent ARID1A mutations (Wilcoxon; P < 0.05; Fig 6B). 306 A significant effect was also observed in the GDSC2 dataset, where bladder cancer cell lines showed 307 increased sensitivity to OTX015 (P = 0.0574; Cohen’s d = -1.33). When observing the interaction in 308 more detail, BRD2 essentiality was significantly enhanced in heterozygous mutated bladder cancer cell 309 lines (Fig 6C). In agreement, heterozygous ARID1A cell lines were also sensitive to several BET 310 inhibitors available in PRISM including ARV-825 and OTX015 (Fig 6D). Previous studies of mouse 311 gastric cancer in models have shown a similar dose dependent role of ARID1A where Arid1a-/+ tumors 312 facilitate global loss of enhancers resulting in p53 suppression and tumor progression, whereas Arid1a-313 /- tumors initiates p53 activation and confers a competitive disadvantage [43]. 314 3 Discussion 315 Despite extensive research efforts and the widespread adoption of high throughput KO and knockdown 316 screens, PARP inhibitors (approved in 2014) remain the only available therapy leveraging SL. This 317 leaves a significant gap between discovery and clinical application. Approximately 90% of LoF driver 318 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 14 mutations currently lack SL-targeted therapies [44]. In this study, we systematically evaluated SL 319 interactions involving cancer driver genes to assess their potential for drug repurposing opportunities 320 using DepMap. Although many clinically focused, CRISPR-based, genetic SL interactions were highly 321 statistically significant, only a subset translated effectively when mapped to drugs that target their gene 322 product, with most successful mappings corresponding to well-characterized interactions. These 323 findings reveal some inherent limitations of pan-cancer SL analyses and emphasize the need for drug-324 gene assays to improve the translatability and clinical impact of SL-based drug development. 325 Assuming that CRISPR-based KO of a gene fully replicates the effects of pharmacological inhibition of 326 its encoded protein has motivated the use of CRISPR LoF screens as high-throughput proxies for 327 identifying single-agent drugs with the potential to induce SL in cancer cells [45]. Although this 328 approach has achieved success in certain contexts [46], significant discrepancies between the 329 outcomes of genetic perturbation and pharmacological inhibition have been observed [47,48]. In our 330 analysis, only 3% of drugs in the PRISM database showed a statistically significant correlation with 331 survival effects observed in the corresponding genetic KOs, with these correlations typically confined to 332 highly targeted oncology agents. This discrepancy is not entirely unexpected. While small-molecule 333 inhibitors and CRISPR-mediated gene KOs can share functional similarities, they represent 334 fundamentally different modes of perturbation. Gene deletions result in the complete loss of gene 335 function and an abrupt cessation of expression. Although this may reveal acute cellular dependencies, 336 it also triggers compensatory responses such as the upregulation of paralogous genes [49], adaptive 337 rewiring of signaling pathways [50], and activation of the DNA damage response [51]. 338 In contrast, small molecules typically target specific protein domains, often impairing catalytic activity 339 while leaving other functional aspects of the protein, such as transcriptional and translational regulation 340 or structural roles, intact. These drugs frequently target conserved domains shared across families of 341 structurally related proteins. For example, BET inhibitors target bromodomains BD1 and BD2 present in 342 BRD2, BRD3, BRD4, and BRDT [52], and Imatinib is a competitive inhibitor of ATP binding to ABL 343 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 15 kinase and also targets the ATP-binding sites of other tyrosine kinases, including c-KIT and PDGFR 344 [53]. As a result, inhibitors often exhibit off-target effects that complicate the specificity of their action, 345 particularly in kinase families, where entire groups of related proteins may be inadvertently inhibited. 346 These off-target effects not only challenge therapeutic precision but can also confound efforts to identify 347 the true biological drivers of anti-tumor responses [54]. In some cases, off-target effects serve as the 348 primary mechanism of tumor suppression [55], highlighting the complexity of drug responses compared 349 to the complete LoF induced by gene KO’s. 350 We believe that unbiased, high throughput methods such as CRISPR KO screens remain 351 indispensable for advancing SL drug development. However, integrating small molecules directly into 352 genetic screens is likely crucial for capturing pharmacological effects and identifying clinically relevant 353 SL interactions, ultimately improving the translational relevance of these findings [34]. 354 Another major challenge in translating SL interactions identified from in vitro screens such as DepMap, 355 as demonstrated by our work and that of others [1,11], is the pronounced context specificity of 356 dependencies. SL dependencies can vary widely depending on factors such as cell of origin, cancer 357 type, cellular state, and experimental conditions, including media composition during screens [56]. This 358 variability presents significant obstacles to identifying SL interactions with broad applicability across 359 diverse cancer types, thereby limiting their translational potential. Controlling for these contextual 360 factors often results in diminishing returns in the discovery of novel SL targets in pan-cancer analyses, 361 while cancer-specific analyses often suffer from insufficient statistical power. These conclusions reflect 362 the inherent complexity of SL target discovery and highlights the need for: (i) accelerated molecular 363 characterization, genetic and drug sensitivity screening of cell lines across diverse cancer types and, (ii) 364 continued adoption of advanced screening strategies, such as integrating drug-gene 365 (pharmacogenomic), gene-gene (combinatorial CRISPR), and drug-drug (combinatorial drug) KO 366 screens, to identify robust, context-specific vulnerabilities and therapeutic opportunities. 367 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 16 The observed divergence between pharmacological inhibition and genetic knockout effects highlights 368 significant limitations of current large-scale genetic screens and the challenges of using genetic 369 perturbation to predict drug therapy effectiveness. These screens often fail to fully align drug sensitivity 370 profiles with SL targets, emphasizing the need for a vastly expanded dependency database, alongside 371 methodological advancements and integrative approaches to improve predictive accuracy. This aligns 372 with the ambitious goal of comprehensively perturbing all protein-coding genes and evaluating the 373 effects of 10,000+ drugs across 20,000+ cancer models, as envisioned by recent initiatives [57]. 374 However, in their current state, these screening efforts are still constrained by the limited representation 375 of many common driver genes in specific cancer contexts and the lack of experimental models for rare 376 cancers. This highlights the need to expand and diversify experimental systems to build a more 377 comprehensive dependency and drug sensitivity database to improve the predictive power of SL-based 378 drug discovery. 379 Our study highlights key limitations in current SL target discovery approaches, including the 380 pronounced context specificity of dependencies, the divergence between genetic KO’s and 381 pharmacological inhibitors, and the limited representation of diverse cancer models. Addressing these 382 challenges requires scaling up and diversifying experimental systems while leveraging advanced 383 strategies such as isogenic screens, combinatorial CRISPR KO’s, and pharmacogenomic screens. The 384 technology to achieve this exists, but the ambition and resources to expand these efforts at scale are 385 now critical to facilitate the translation of in vitro discoveries into clinical breakthroughs and accelerate 386 the development of impactful precision cancer therapies. A recent pharmacogenomics study 387 exemplifies these challenges by conducting genetic KO screens targeting LoF drivers and evaluating 388 their interactions with clinically approved targeted therapies and cancer drugs. Of the 32 driver-drug 389 interactions identified, only a KRAS hotspot mutations with RAF1 knockout was reproduced using 390 DepMap data (Pearson; P < 0.05; Fig. 7A). Interestingly, BRCA1 and STAG2 deletions sensitized cells 391 to PARP inhibitors in the pharmacogenomic screen; however, similar effects were absent in DepMap 392 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 17 when stratifying cell lines by BRCA1 and STAG2 mutation status. This discrepancy highlights the 393 complexity of PARP inhibitor mechanisms, which often rely on PARP trapping to stabilize PARP-DNA 394 complexes and induce cytotoxicity. Genetic KO’s, by contrast, do not replicate the formation of these 395 toxic drug-target intermediates or their downstream effects on genomic instability and cell death. These 396 findings emphasize the challenges of modeling pharmacogenetic mechanisms using genetic 397 perturbations alone and underscore the need to incorporate drug-specific effects into therapeutic 398 development. Expanding the scale and scope of experimental systems is essential to address these 399 gaps, improve the predictive power of SL-based drug discovery, and bridge the divide between in vitro 400 findings and clinical applications. 401 402 4 Methods 403 Synthetic lethal network construction 404 Input data 405 All cell line data, including CRISPR-Cas9 knockout fitness scores, and mutation data, were obtained 406 from the DepMap portal (DepMap Public 23Q2; https://depmap.org/portal/). Mutation matrices were 407 downloaded for both hotspot (OmicsSomaticMutationsMatrixHotspot.csv) and damaging mutations 408 (OmicsSomaticMutationsMatrixDamaging.csv). As described on the DepMap portal, DepMap 23Q2 409 mutations were called using Mutect2, with 0 indicating no mutation, 1 for heterozygous mutations, and 410 2 for homozygous mutations. Hotspot mutation sites were identified based on Hess et al. (2019). 411 CRISPR-Cas9 knockout effect scores were extracted from the CRISPRGeneEffect.csv file. For 412 DepMap 23Q2, gene-level knockout effect scores across all cell line models from both Achilles and 413 Project Score were computed using Chronos and harmonized using Harmonia. Microsatellite instability 414 scores were obtained from the DepMap Public 24Q2 release OmicsSignatures.csv file. 415 Selectively lethal genes 416 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 18 To refine feature selection, we excluded genes that did not significantly affect cell proliferation across 417 all examined cell lines. Specifically, genes were retained in our dataset if they exhibited effect scores ≤ 418 -1 or ≥ 0.4 in at least four distinct cell lines. This approach ensures that only genes influencing cell 419 viability are included. 420 Random forest feature selection. 421 To limit the search space and reduce the burden of multiple test correction we implemented feature 422 selection classification using the Boruta algorithm (v8.0.0), and a workflow largely based on the PARIS 423 pipeline (Benfatto et al., 2021). The Boruta algorithm works by first generating shadow features that are 424 random permutations of the original features. A random forest classifier is then employed to assess the 425 importance of each original feature relative to the maximum importance of the permuted shadow 426 features (shadowMax). If the importance of a feature – in this case gene mutation – significantly 427 exceeds shadowMax, it is retained; otherwise, it is rejected. In our analysis, Boruta was applied to each 428 gene using CRISPR-Cas9 knockout effect scores as the response variable, with concatenated hotspot 429 and damaging mutations as predictive features. Source genes with greater than 4 mutations were 430 included, as known SL interactions were observed as significant at this threshold (S7 Fig). The 431 algorithm was run across all cell lines (pan-cancer) and on cancer-specific subsets 432 (OncotreePrimaryDisease) that had ≥ 28 fully characterized cell lines in CCLE. We employed three 433 importance algorithms - getImpExtraGini, getImpRfZ, and getImpExtraRaw - to calculate feature 434 importance within the random forest models (ranger v0.16.0). Interactions were considered significant 435 only if all three algorithms concurred on feature importance. The significance threshold for feature 436 importance was set at a p-value of 0.01, and the maximum number of iterations for Boruta was capped 437 at 500. All code used in this study will be publicly available on GitHub: 438 https://github.com/michaelcvermeulen/cancer_driver_repurposing (in progress) 439 Defining mutual exclusivity 440 441 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 19 Tumor mutation and copy number data was downloaded from TCGA. ME was assessed for each 442 genetic interaction pair using Fisher’s Exact Test, a statistical method to test the non-random 443 association between the mutation or deep deletion status of two genes across a cohort of tumor 444 samples. For each pair, a 2x2 contingency table was constructed, with entries representing co-445 occurrence patterns: both genes mutated, only one gene mutated, or neither gene mutated. The 446 resulting p-values were used to determine significant ME, with the direction of association evaluated 447 using odds ratios: OR 1 suggested co-occurrence. 448 To account for cancer-type-specific effects, ME was tested separately in each TCGA cancer-type 449 subset, with the most significant interaction retained for each interaction pair. Given the larger sample 450 sizes associated with driver mutations, we performed random sampling of non-driver gene pairs with 451 equal or larger sample sizes (n) to compare ME signal distributions. FDR correction using the 452 Benjamini-Hochberg method was applied to adjust for multiple hypothesis testing, with gene pairs 453 considered significant if FDR < 0.05. 454 455 ANOVA model 456 All pan-cancer SL interaction pairs identified through Boruta analysis were further refined to assess 457 directionality, effect size, and potential confounding factors. Effect sizes were calculated using Cohen’s 458 d, while Pearson correlations were used to establish the directionality of interactions. To account for 459 confounding variables, including microsatellite instability (MSI), cancer type, and cell line growth 460 properties (adherent, suspension, or mixed), we applied a Type II ANOVA model adapted from Lord et 461 al. (2020). 462 463 The ANOVA model structure was defined as: 464 465 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 20 𝐶𝑅𝐼𝑆𝑃𝑅 𝐸𝑓𝑓𝑒𝑐𝑡 𝑆𝑐𝑜𝑟𝑒 ~ 𝑀𝑆𝐼 + 𝐺𝑟𝑜𝑤𝑡ℎ 𝑃𝑟𝑜𝑝𝑒𝑟𝑡𝑖𝑒𝑠 + 𝐶𝑎𝑛𝑐𝑒𝑟 𝑇𝑦𝑝𝑒 + 𝑀𝑢𝑡𝑎𝑡𝑖𝑜𝑛 𝑆𝑡𝑎𝑡𝑢𝑠 471 472 Damaging mutations were used for TSGs and hotspot mutations used for oncogenes. Mutation status 466 p-values were extracted for each interaction and adjusted for multiple tests using Benjamini-Hochberg. 467 To ensure the robustness of the final network, interactions were retained only if mutation status 468 exhibited an FDR < 0.05, excluding those where confounding factors contributed significantly to the 469 observed effects. 470 Intersection with drug sensitivity data 473 PRISM drug sensitivity data were obtained from the DepMap portal (DepMap Public 23Q4 release). 474 Gene targets for each drug in the PRISM dataset were compiled by merging information from Citeline 475 and PRISM repurposing target resources. For each pan-cancer synthetic lethal (SL) interaction, cell 476 lines were stratified based on the mutation status of the gene of interest. Drug sensitivity, specifically to 477 those targeting the corresponding gene, was evaluated using Cohen’s effect size, Pearson correlation 478 coefficients, and linear models, with adjustments for microsatellite instability (MSI) status, growth 479 properties, and cancer type. Only cell lines from cancer types with more than 10 available lines in the 480 CCLE database, as classified by the OncotreePrimaryDisease field, were included in the analysis. 481 To compare the effects of targeted drugs with the CRISPR-Cas9 knockout (KO) predictions, linear 482 models were fit between drug sensitivity profiles and CRISPR effect scores. These models were 483 adjusted for cancer type, MSI status, and growth properties to account for confounding factors when 484 testing pan-cancer interactions. 485 486 Overlap with pharmacogenomic data 487 Noteworthy pharmacogenomic interactions involving driver genes were sourced from Truesdell et al. 488 [34]. Each interaction involving targeted inhibitors was evaluated in both the DepMap and PRISM 489 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 21 datasets by assessing target gene/protein proliferative effects stratified by driver gene mutation, using 490 both (i) the pharmacological effect of the inhibitor and (ii) the impact of genetic knockout on the 491 corresponding target. 492 493 Network visualization 494 Pan-cancer genetic interaction networks were visualized using the visNetwork package (v2.1.2) in R, 495 offering an interactive and customizable platform for exploring genetic interactions. Node sizes were 496 scaled based on total connections, while edge widths were proportional to the getImpExtraGini 497 importance scores generated by Boruta. To enhance clarity, importance scores were capped at 30. 498 Interactions with druggable targets identified in PRISM were highlighted in blue. 499 500 Acknowledgments 501 We thank Christopher D Moyes for providing valuable guidance and feedback on the manuscript. We 502 thank Doris Coto Villa, Stephanie Young, Peter Truesdell and Ben Snider for valuable discussions and 503 contributions to early stages of this work. 504 Funding 505 This work was supported by a Canadian Institutes of Health Research Project Grant (PJT 178214) 506 awarded to Andrew Craig and Tomas Babak, and funds from the Ontario Graduate Scholarship 507 awarded to Michael Vermeulen. The funders had no role in study design, data collection and analysis, 508 decision to publish, or preparation of the manuscript. 509 510 Author Contributions 511 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 22 Conceptualization – TB, AC 512 Data curation – MV 513 Formal analysis – TB, MV 514 Funding acquisition – TB, AC, MV 515 Methodology – MV, TB 516 Project administration – TB, MV, AC 517 Supervision – TB, AC 518 Visualization - MV 519 Writing – original draft – MV, TB 520 Writing – review & editing – MV, TB, AC 521 522

523 1. O’Neil NJ, Bailey ML, Hieter P. Synthetic lethality and cancer. Nat Rev Genet. 2017;18: 613–623. 524 doi:10.1038/nrg.2017.47 525 2. Setton J, Zinda M, Riaz N, Durocher D, Zimmermann M, Koehler M, et al. Synthetic Lethality in Cancer 526 Therapeutics: The Next Generation. Cancer Discov. 2021;11: 1626–1635. doi:10.1158/2159-8290.CD-20-1503 527 3. Rose M, Burgess JT, O’Byrne K, Richard DJ, Bolderson E. PARP Inhibitors: Clinical Relevance, Mechanisms 528 of Action and Tumor Resistance. Frontiers in Cell and Developmental Biology. 2020;8. Available: 529 https://www.frontiersin.org/articles/10.3389/fcell.2020.564601 530 4. Arafeh R, Shibue T, Dempster JM, Hahn WC, Vazquez F. The present and future of the Cancer 531 Dependency Map. Nat Rev Cancer. 2024. doi:10.1038/s41568-024-00763-x 532 5. Srivatsa S, Montazeri H, Bianco G, Coto-Llerena M, Marinucci M, Ng CKY, et al. Discovery of synthetic 533 lethal interactions from large-scale pan-cancer perturbation screens. Nat Commun. 2022;13: 7748. 534 doi:10.1038/s41467-022-35378-z 535 6. Markowska M, Budzinska MA, Coenen-Stass A, Kang S, Kizling E, Kolmus K, et al. Synthetic lethality 536 prediction in DNA damage repair, chromatin remodeling and the cell cycle using multi-omics data from cell lines 537 and patients. Sci Rep. 2023;13: 7049. doi:10.1038/s41598-023-34161-4 538 7. Karimpour M, Totonchi M, Behmanesh M, Montazeri H. Pathway-driven analysis of synthetic lethal 539 interactions in cancer using perturbation screens. Life Sci Alliance. 2024;7: e202302268. 540 doi:10.26508/lsa.202302268 541 8. Rosenski J, Shifman S, Kaplan T. Predicting gene knockout effects from expression data. BMC Medical 542 Genomics. 2023;16: 26. doi:10.1186/s12920-023-01446-6 543 9. Gonçalves E, Segura‐Cabrera A, Pacini C, Picco G, Behan FM, Jaaks P, et al. Drug mechanism‐of‐action 544 discovery through the integration of pharmacological and CRISPR screens. Mol Syst Biol. 2020;16: e9405. 545 doi:10.15252/msb.20199405 546 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 23 10. Lord CJ, Quinn N, Ryan CJ. Integrative analysis of large-scale loss-of-function screens identifies robust 547 cancer-associated genetic interactions. Murphy ME, editor. eLife. 2020;9: e58925. doi:10.7554/eLife.58925 548 11. Tsherniak A, Vazquez F, Montgomery PG, Weir BA, Kryukov G, Cowley GS, et al. Defining a Cancer 549 Dependency Map. Cell. 2017;170: 564-576.e16. doi:10.1016/j.cell.2017.06.010 550 12. Corsello SM, Nagari RT, Spangler RD, Rossen J, Kocak M, Bryan JG, et al. Discovering the anticancer 551 potential of non-oncology drugs by systematic viability profiling. Nat Cancer. 2020;1: 235–248. 552 doi:10.1038/s43018-019-0018-6 553 13. Benfatto S, Serçin Ö, Dejure FR, Abdollahi A, Zenke FT, Mardin BR. Uncovering cancer vulnerabilities by 554 machine learning prediction of synthetic lethality. Molecular Cancer. 2021;20: 111. doi:10.1186/s12943-021-555 01405-8 556 14. Dempster JM, Krill-Burger JM, McFarland JM, Warren A, Boehm JS, Vazquez F, et al. Gene expression has 557 more power for predicting in vitro cancer cell vulnerabilities than genomics. Cancer Biology; 2020 Feb. 558 doi:10.1101/2020.02.21.959627 559 15. Dou Y, Ren Y, Zhao X, Jin J, Xiong S, Luo L, et al. CSSLdb: Discovery of cancer-specific synthetic lethal 560 interactions based on machine learning and statistic inference. Comput Biol Med. 2024;170: 108066. 561 doi:10.1016/j.compbiomed.2024.108066 562 16. Bazaga A, Leggate D, Weisser H. Genome-wide investigation of gene-cancer associations for the 563 prediction of novel therapeutic targets in oncology. Sci Rep. 2020;10: 10787. doi:10.1038/s41598-020-67846-1 564 17. Shi X, Verduzco D, Petiwala S, Jeffries C, Lu C, Anton T, et al. TCGADEPMAP – Mapping Translational 565 Dependencies and Synthetic Lethalities within The Cancer Genome Atlas. bioRxiv; 2022. p. 2022.03.24.485544. 566 doi:10.1101/2022.03.24.485544 567 18. He D, Liu Q, Wu Y, Xie L. A context-aware deconfounding autoencoder for robust prediction of 568 personalized clinical drug response from cell-line compound screening. Nat Mach Intell. 2022;4: 879–892. 569 doi:10.1038/s42256-022-00541-0 570 19. Chiu Y-C, Zheng S, Wang L-J, Iskra BS, Rao MK, Houghton PJ, et al. Predicting and characterizing a cancer 571 dependency map of tumors with deep learning. Science Advances. 2021;7: eabh1275. 572 doi:10.1126/sciadv.abh1275 573 20. Cohen Z, Petrenko E, Barisaac AS, Abu-Zhayia ER, Yanovich-Ben-Uriel C, Ayoub N, et al. SLAYER: 574 Synthetic Lethality Analysis for Enhanced Targeted Therapy Implicates AhR inhibitor as a Target in RB1-Mutant 575 Bladder Tumors. 2024. doi:10.1101/2024.05.01.592073 576 21. Fedrizzi T, Ciani Y, Lorenzin F, Cantore T, Gasperini P, Demichelis F. Fast mutual exclusivity algorithm 577 nominates potential synthetic lethal gene pairs through brute force matrix product computations. 578 Computational and Structural Biotechnology Journal. 2021;19: 4394–4403. doi:10.1016/j.csbj.2021.08.001 579 22. Liany H, Jayagopal A, Huang D, Lim JQ, Nbh NI, Jeyasekharan A, et al. ASTER: A Method to Predict 580 Clinically Relevant Synthetic Lethal Genetic Interactions. IEEE J Biomed Health Inform. 2024;28: 1785–1796. 581 doi:10.1109/JBHI.2024.3354776 582 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 24 23. Srihari S, Singla J, Wong L, Ragan MA. Inferring synthetic lethal interactions from mutual exclusivity of 583 genetic events in cancer. Biol Direct. 2015;10: 57. doi:10.1186/s13062-015-0086-1 584 24. Yang W, Soares J, Greninger P, Edelman EJ, Lightfoot H, Forbes S, et al. Genomics of Drug Sensitivity in 585 Cancer (GDSC): a resource for therapeutic biomarker discovery in cancer cells. Nucleic Acids Research. 2013;41: 586 D955–D961. doi:10.1093/nar/gks1111 587 25. Carli F, Chiaro PD, Morelli M, Arora C, Bisceglia L, Rosa NDO, et al. Learning and actioning general 588 principles of cancer cell drug sensitivity. bioRxiv; 2024. p. 2024.03.28.586783. doi:10.1101/2024.03.28.586783 589 26. Baek B, Jang E, Park S, Park S-H, Williams DR, Jung D-W, et al. Integrated drug response prediction 590 models pinpoint repurposed drugs with effectiveness against rhabdomyosarcoma. PLOS ONE. 2024;19: 591 e0295629. doi:10.1371/journal.pone.0295629 592 27. Pellecchia S, Viscido G, Franchini M, Gambardella G. Predicting drug response from single-cell expression 593 profiles of tumours. BMC Medicine. 2023;21: 476. doi:10.1186/s12916-023-03182-1 594 28. Goodspeed A, Heiser LM, Gray JW, Costello JC. Tumor-Derived Cell Lines as Molecular Models of Cancer 595 Pharmacogenomics. Molecular Cancer Research. 2016;14: 3–13. doi:10.1158/1541-7786.MCR-15-0189 596 29. Li A, Walling J, Kotliarov Y, Center A, Steed ME, Ahn SJ, et al. Genomic Changes and Gene Expression 597 Profiles Reveal That Established Glioma Cell Lines Are Poorly Representative of Primary Human Gliomas. 598 Molecular Cancer Research. 2008;6: 21–30. doi:10.1158/1541-7786.MCR-07-0280 599 30. Salvadores M, Fuster-Tormo F, Supek F. Matching cell lines with cancer type and subtype of origin via 600 mutational, epigenomic, and transcriptomic patterns. Science Advances. 2020;6: eaba1862. 601 doi:10.1126/sciadv.aba1862 602 31. Bailey MH, Tokheim C, Porta-Pardo E, Sengupta S, Bertrand D, Weerasinghe A, et al. Comprehensive 603 Characterization of Cancer Driver Genes and Mutations. Cell. 2018;173: 371-385.e18. 604 doi:10.1016/j.cell.2018.02.060 605 32. Campbell PJ, Getz G, Korbel JO, Stuart JM, Jennings JL, Stein LD, et al. Pan-cancer analysis of whole 606 genomes. Nature. 2020;578: 82–93. doi:10.1038/s41586-020-1969-6 607 33. Nulsen J, Misetic H, Yau C, Ciccarelli FD. Pan-cancer detection of driver genes at the single-patient 608 resolution. Genome Med. 2021;13: 12. doi:10.1186/s13073-021-00830-0 609 34. Truesdell P, Chang J, Coto Villa D, Dai M, Zhao Y, McIlwain R, et al. Pharmacogenomic discovery of 610 genetically targeted cancer therapies optimized against clinical outcomes. npj Precis Onc. 2024;8: 1–13. 611 doi:10.1038/s41698-024-00673-z 612 35. Zemanova J, Hylse O, Collakova J, Vesely P, Oltova A, Borsky M, et al. Chk1 inhibition significantly 613 potentiates activity of nucleoside analogs in TP53-mutated B-lymphoid cells. Oncotarget. 2016;7: 62091–62106. 614 doi:10.18632/oncotarget.11388 615 36. Tagal V, Wei S, Zhang W, Brekken RA, Posner BA, Peyton M, et al. SMARCA4-inactivating mutations 616 increase sensitivity to Aurora kinase A inhibitor VX-680 in non-small cell lung cancers. Nat Commun. 2017;8: 617 14098. doi:10.1038/ncomms14098 618 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 25 37. Ji H, Lu X, Zhao S, Wang Q, Liao B, Bauer LG, et al. Target deconvolution with matrix-augmented pooling 619 strategy reveals cell-specific drug-protein interactions. Cell Chem Biol. 2023;30: 1478-1487.e7. 620 doi:10.1016/j.chembiol.2023.08.002 621 38. Xu X, Peng Q, Jiang X, Tan S, Yang Y, Yang W, et al. Metabolic reprogramming and epigenetic 622 modifications in cancer: from the impacts and mechanisms to the treatment potential. Exp Mol Med. 2023;55: 623 1357–1370. doi:10.1038/s12276-023-01020-1 624 39. Wooller SK, Pearl LH, Pearl FMG. Identifying actionable synthetically lethal cancer gene pairs using 625 mutual exclusivity. FEBS Lett. 2024;598: 2028–2039. doi:10.1002/1873-3468.14950 626 40. Shu S, Wu H-J, Ge JY, Zeid R, Harris IS, Jovanović B, et al. Synthetic Lethal and Resistance Interactions 627 with BET Bromodomain Inhibitors in Triple-Negative Breast Cancer. Mol Cell. 2020;78: 1096-1113.e8. 628 doi:10.1016/j.molcel.2020.04.027 629 41. Geng H, Qian R, Zhong Y, Tang X, Zhang X, Zhang L, et al. Leveraging synthetic lethality to uncover 630 potential therapeutic target in gastric cancer. Cancer Gene Ther. 2024;31: 334–348. doi:10.1038/s41417-023-631 00706-y 632 42. Berns K, Caumanns JJ, Hijmans EM, Gennissen AMC, Severson TM, Evers B, et al. ARID1A mutation 633 sensitizes most ovarian clear cell carcinomas to BET inhibitors. Oncogene. 2018;37: 4611–4625. 634 doi:10.1038/s41388-018-0300-6 635 43. Loe AKH, Francis R, Seo J, Du L, Wang Y, Kim J-E, et al. Uncovering the dosage-dependent roles of Arid1a 636 in gastric tumorigenesis for combinatorial drug therapy. Journal of Experimental Medicine. 2021;218: 637 e20200219. doi:10.1084/jem.20200219 638 44. Hay M, Thomas DW, Craighead JL, Economides C, Rosenthal J. Clinical development success rates for 639 investigational drugs. Nat Biotechnol. 2014;32: 40–51. doi:10.1038/nbt.2786 640 45. Fellmann C, Gowen BG, Lin P-C, Doudna JA, Corn JE. Cornerstones of CRISPR–Cas in drug discovery and 641 therapy. Nat Rev Drug Discov. 2017;16: 89–100. doi:10.1038/nrd.2016.238 642 46. Xia Y, Ji X, Jang IS, Surka C, Hsu C, Wang K, et al. Genetic and pharmacological interrogation of cancer 643 vulnerability using a multiplexed cell line screening platform. Commun Biol. 2021;4: 834. doi:10.1038/s42003-644 021-02352-2 645 47. DeWeirdt PC, Sangree AK, Hanna RE, Sanson KR, Hegde M, Strand C, et al. Genetic screens in isogenic 646 mammalian cell lines without single cell cloning. Nat Commun. 2020;11: 752. doi:10.1038/s41467-020-14620-6 647 48. Lin A, Giuliano CJ, Sayles NM, Sheltzer JM. CRISPR/Cas9 mutagenesis invalidates a putative cancer 648 dependency targeted in on-going clinical trials. Settleman J, editor. eLife. 2017;6: e24179. 649 doi:10.7554/eLife.24179 650 49. Rossi A, Kontarakis Z, Gerri C, Nolte H, Hölper S, Krüger M, et al. Genetic compensation induced by 651 deleterious mutations but not gene knockdowns. Nature. 2015;524: 230–233. doi:10.1038/nature14580 652 50. El-Brolosy MA, Kontarakis Z, Rossi A, Kuenne C, Günther S, Fukuda N, et al. Genetic compensation 653 triggered by mutant mRNA degradation. Nature. 2019;568: 193–197. doi:10.1038/s41586-019-1064-z 654 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint 26 51. Haapaniemi E, Botla S, Persson J, Schmierer B, Taipale J. CRISPR–Cas9 genome editing induces a p53-655 mediated DNA damage response. Nat Med. 2018;24: 927–930. doi:10.1038/s41591-018-0049-z 656 52. Filippakopoulos P, Qi J, Picaud S, Shen Y, Smith WB, Fedorov O, et al. Selective inhibition of BET 657 bromodomains. Nature. 2010;468: 1067–1073. doi:10.1038/nature09504 658 53. Druker BJ, Talpaz M, Resta DJ, Peng B, Buchdunger E, Ford JM, et al. Efficacy and Safety of a Specific 659 Inhibitor of the BCR-ABL Tyrosine Kinase in Chronic Myeloid Leukemia. New England Journal of Medicine. 660 2001;344: 1031–1037. doi:10.1056/NEJM200104053441401 661 54. Green MR, Newton MD, Fancher KM. Off-Target Effects of BCR-ABL and JAK2 Inhibitors. Am J Clin Oncol. 662 2016;39: 76–84. doi:10.1097/COC.0000000000000023 663 55. Lin A, Giuliano CJ, Palladino A, John KM, Abramowicz C, Yuan ML, et al. Off-target toxicity is a common 664 mechanism of action of cancer drugs undergoing clinical trials. Sci Transl Med. 2019;11: eaaw8412. 665 doi:10.1126/scitranslmed.aaw8412 666 56. Hart T, Chandrashekhar M, Aregger M, Steinhart Z, Brown KR, MacLeod G, et al. High-Resolution CRISPR 667 Screens Reveal Fitness Genes and Genotype-Specific Cancer Liabilities. Cell. 2015;163: 1515–1526. 668 doi:10.1016/j.cell.2015.11.015 669 57. Boehm JS, Garnett MJ, Adams DJ, Francies HE, Golub TR, Hahn WC, et al. Cancer research needs a better 670 map. Nature. 2021;589: 514–516. doi:10.1038/d41586-021-00182-0 671 672 673 Supporting information captions 674 675 S1 Table. DepMap Pan-cancer SL network interactions and related PRISM drug correlations 676 S2 Table. DepMap Cancer-type specific SL network interactions and related PRISM drug correlations 677 678 S3 Table. Correlations between CRISPR effect score and PRISM drug sensitivity 679 S4 Table. Driver gene-stratified differential PRISM drug sensitivity for damaging and hotspot mutations 680 S1 Fig. Schematic illustrating genetic interaction concepts relevant to cancer genomics. 681 S2 Fig. Distribution of synthetic lethal (SL) interaction significance values stratified by mutation 682 frequency across cell lines. 683 S3 Fig. Overview of pan-cancer SL and resistance genetic interactions identified in the pan-cancer 684 analysis. 685 S4 Fig. Correlation of drug sensitivity with genetic dependency profiles for MCL1 inhibitors 686 S5 Fig. Overview of SL interactions identified across cancer types following filtering criteria. 687 S6 Fig. Top cancer-specific SL interactions identified across CCLE cell lines. 688 S7 Fig. Relationship between the statistical significance of SL genetic interactions and the number of 689 mutated cell lines. 690 .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint Figure 1. An overview of the approach to select for clinically actionable targeted drug repurposing opportunities. Using the Boruta algorithm, our pipeline integrates data from the DepMap consortium (release 23Q2) to construct a genetic interaction network. Fitness scores from CRISPR loss-of-function screens serve as dependent variables, while cell line mutation data (hotspot and damaging) are used as independent variables in a random forest (RF) feature selection process. For each perturbed gene, the algorithm assigns an importance score to mutation features, retaining those with any predictive value, and forming the foundation of our synthetic lethality (SL) network. To establish data features that are predictive of interactions holding up in the clinic (i.e. tumors), we utilized TCGA mutation data to assess mutual exclusivity (ME) across sufficiently powered genes.Non-overlapping mutations in human tumors suggest incompatibility and potential synthetic lethality. The resulting genetic interaction network was then integrated with PRISM drug screen data (release 23Q4) to identify drugs that mirror CRISPR effect scores and display differential sensitivity in driver-stratified subsets, generating a list of candidate drugs for targeted repurposing. .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint Figure 2. Cancer driver based genetic interaction network. A) Force-directed network of 1,428 high-confidence genetic interaction pairs identified from DepMap CRISPR viability screens, accounting for biases arising from cancer type and microsatellite instability. The network includes both synthetic lethal and synthetic viability interactions. B) Genetic interaction network filtered to show driver source mutations, with driver gene nodes colored red (oncogenes) and blue (tumor suppressor genes) and non-drivers in yellow. Blue edges indicate genetic interactions containing a drug target listed in the PRISM database, with edge width representing the strength of the interaction (Importance value). Node size reflects the total number of connections. C) Comparison of mutual exclusivity significance (y-axis) in TCGA data between synthetic lethal pairs containing a driver gene (blue) versus non-driver interactions (red) shows that interactions involving driver genes are more likely to be mutually exclusive. Grey lines represent the distribution of randomly permuted subsets of non-driver interactions. .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint Figure 3. A) Heatmap depicting effect sizes of pan-cancer SL interactions involving driver genes, derived from DepMap CRISPR effect scores (ANOVA; FDR < 0.1). Driver mutations are shown on the x-axis, and CRISPR target genes on the y-axis. Negative effect size values indicate that mutated drivers sensitize cells to target gene knockout, while positive values suggest that driver mutations confer resistance or that wild-type alleles sensitize cells. On the right, Pearson correlations between PRISM drug sensitivity and CRISPR knockout effects for sample interactions are provided. Higher Pearson correlation significance indicates alignment between drug sensitivity and CRISPR knockout profiles across all CCLE cell lines, implying comparable impacts on cell viability from drug inhibition and gene knockout. Drugs with negative correlations are labeled in blue. B) Selected examples are highlighted in white-outlined boxes: Box 1, BRAF; Box 4, PIK3CA; Box 5, CDK6/RB1. C) Known SL interactions, including paralog pairs, are provided as positive controls: Box 3, SMARCA2/4; Box 2, STAG1/2. .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint Figure 4: Drug perturbations typically do not replicate genetic perturbations of corresponding drug targets. Density plot showing the distribution of Spearman's rank correlation coefficients between gene effect scores and PRISM drug sensitivity for drugs with overlapping gene targets (blue). The red curve represents the Spearman's rank correlation coefficients between each PRISM drug and a set of randomly permuted gene effect scores, serving as a control. Bottom: Boxplot summarizing the same data, highlighting some of the strongest correlations between gene effects and drug sensitivity. .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint Figure 5: Pan-cancer synthetic lethal interactions recapitulated by PRISM drug sensitivity. Synthetic lethal interactions involving hotspot and damaging driver mutations were overlapped with PRISM drugs targeting the same genes. Tumor suppressor genes and oncogenes are indicated in blue and red, respectively. The inset provides a magnified view of the central plot region. For each SL interaction, cell lines were categorized as wild-type or mutant for the source gene, and differences between CRISPR effect scores (y-axis) and PRISM drug logFC values (x-axis) were assessed using type II ANOVA, controlling for cancer type and MSI status. Text labels follow the format "drug_source-mutation_target gene." P-values were -log10 transformed and adjusted by the sign of the effect size to reflect directionality. Points in the top right quadrant have a positive effect size for both drug sensitivity and CRISPR effect, indicating decreased viability in wild- type lines for the driver gene, while points in the bottom left quadrant show increased sensitivity in driver- mutant lines to both drug and CRISPR knockout. For instance, REFAMETINIB_BRAF_MAP2K1 in the bottom left quadrant demonstrates that BRAF-mutant cell lines are sensitive to both MAP2K1 CRISPR deletion and the MAP2K1 inhibitor Refametinib. Points in the top left and bottom right suggest discordance between drug and CRISPR effects on viability. When multiple PRISM drugs matched a target gene, the most significant interaction was plotted. Full data are available in S4 Table. .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint Figure 6. Overlap of DepMap synthetic lethal (SL) interactions and PRISM drug sensitivity in bladder urothelial carcinoma, highlighting ARID1A/BRD2 as a potential repurposing opportunity. A) SL interactions identified by CRISPR screens and corresponding drug sensitivities from the PRISM database in bladder urothelial carcinoma cell lines. The x-axis indicates the significance of genetic dependencies (CRISPR-Mut), and the y-axis shows the significance of drug sensitivity differences (Drug-Mut), represented as -log10(p) values. Circle size denotes sample size, while labeled red points highlight significant drug-gene interactions, notably the ARID1A-BRD2 interaction with BET inhibitor ARV-825. The top gene-drug interactions are shown when multiple drugs have the same predicted targets. B) Heatmap depicting differential sensitivity of bladder cancer cell lines carrying ARID1A mutations versus wild-type cell lines to various BET inhibitors from the PRISM dataset. ARID1A-mutated cell lines demonstrate consistent and heightened sensitivity to nearly all BET inhibitors in bladder cancer. C) Boxplot showing significantly lower BRD2 knockout (KO) fitness scores in bladder cancer cell lines harboring damaging ARID1A mutations compared to wild-type lines, suggesting increased dependency on BRD2 in the heterozygous mutant context. D) Boxplot illustrating heightened sensitivity (negative fold change in viability scores from PRISM primary screening) of ARID1A- mutated bladder cancer cell lines to the BET inhibitor ARV-825 compared to wild-type cell lines. .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint Figure 7. Limited replication of driver-drug interactions between pharmacogenomic and large-scale genomic screens. (A) Pearson correlation coefficients for the association between driver mutations and knockout of drug targets in DepMap CRISPR screens, stratified by interaction type (sensitizing vs. suppressive) that were detected in pharmacogenomic screens (Truesdell et al., 2024). (B) Corresponding correlations for drug sensitivity profiles from PRISM screens. Although KRAS–RAF1 was reproduced in DepMap (A), most interactions, including BRCA1–PARP1 and STAG2–RUCAPARIB, did not show consistent patterns across datasets. .CC-BY 4.0 International licenseavailable under a (which 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 preprintthis version posted March 22, 2025. ; https://doi.org/10.1101/2025.03.20.644275doi: bioRxiv preprint