Pooled CRISPR pharmacogenomic screening to identify druggable driver-mutation dependenciesTargeted therapies are designed to treat tumors with driver mutations in either oncogenes (GoF) or tumor suppressor genes (LoF). Unlike oncogenes, where mutations result in a targetable protein, LoF mutations in tumor suppressor genes typically yield reduced functionality of the protein that cannot be easily reversed pharmacologically (Fig. 1a). Consequently, the majority of approved genetically targeted therapies are for cancers driven by oncogene mutations (Supplementary Fig. 1); whereas there remains a major unmet need for treatment of tumors driven by mutation of tumor suppressor genes.Fig. 1: CRISPR-pooled pharmacogenomic screening to identify druggable genetic dependencies of LoF TSG mutations.a Schematic showing the drug-targetability of GoF versus LoF mutations. GoF mutations result in a protein product that can be pharmacologically targeted, while LoF mutations result in the absence of a protein product. b Schematic showing the synergistic effects of SL mutations on cell viability. Single mutations in either white or blue genes do not impact cell growth, but when both are mutated in the same cell the combination is lethal. c Schematic of CODA-PGX platform for rapid, high throughput, pharmacogenomic screening. Cancer cell lines are transduced with a pool of driver gene sgRNAs to create knockouts via CRISPR and grown in the presence or absence of drug. The viability of each knockout is detected by NGS sequencing of the corresponding guide RNAs used to create it. Data is visualized by comparing sgRNA distribution between drug and control samples.It is widely accepted that identifying vulnerabilities of tumors through SL interactions is key to unlocking the potential of TSG targeting (Fig. 1b) and driver genes have recently become recognized as important to clinical translation11. While large consortia have made great advances to identify genes which when knocked out together kill cancer cells, challenges remain. Despite the availability of whole-genome pooled CRISPR screening data from more than 1000 cell lines, we are likely severely underpowered to complete a comprehensive mapping of SL networks without the use of isogenic screens15. Furthermore, translating SL discoveries in cell lines is challenging for both scientific11 and practical2 reasons.CODA-PGX platform designTo address these challenges, we developed a method that efficiently identifies sensitizing driver mutations for small molecule drugs (Fig. 1c). In brief, the approach is to: (i) perform driver-focused CRISPR screens with and without drug in a pool of isogenically controlled cell lines, (ii) use NGS to identify genes which, when mutated, cause cells to ‘drop out’ of the pool. We refer to the genes containing these mutations as ‘hits’ and expect that their presence in tumors may identify patients whose clinical outcomes could be improved by treatment with the corresponding drug. We curated driver genes represented on our platform (Supplementary Table 1) from published lists16,17,18 and subset to genes mutated in at least 5% of one cancer type in The Cancer Genome Atlas (TCGA; cancer.gov/tcga). 95.1% of all patients in TCGA have at least one mutation in a gene targeted on our platform. 3–4 guide sequences targeting these genes were selected from the Brunello library19. Importantly, we also added 96 guides that target non-conserved intergenic regions to serve as negative controls (Supplementary Table 3).To visualize the data, we plot, for each driver gene, the mean ratio of guide sequences present in drug-treated cells versus guide sequences present in control (vehicle-only) on a log scale (Fig. 1c, bottom). Guide sequences with negative log values represent genes which when mutated reduce the viability of the cells in the presence of the drug. We then gauge significance (y-axis) using a ranksum test of the 3–4 guides/gene ratios against the distribution of ratios of the 96 negative controls. Negative controls establish a conservative null distribution of platform technical variation, thereby yielding a robust drug-driver synergy signal.Motivated by the recent discovery that driver biology is important for translating discoveries made in cell lines to clinical outcomes11, we optimized several parameters to maximize representation of native driver biology. We found that oncogene GoF and TSG LoF effects on viability are reproduced in public genome-wide pooled CRISPR screens such as the Dependency Map (20,21; Supplementary Fig. 3). For example, cell lines with an oncogenic BRAFV600E activating mutation were more susceptible to BRAF-knockout (KO) than BRAF-wildtype (WT) lines, and KO of the TSG PTEN in PTEN-WT lines increased the relative fitness of cells. We therefore identified, for each driver gene, the set of cell lines where these expected effects on proliferation were maintained. We also used Celligner22 to identify cell lines whose expression profile matches its tumor of origin. Furthermore, to best capture the founder-driver biology of our genetic perturbations, we reasoned that TSGs present as 2 WT copies, with evidence of expression23 best represent the native initial tumor state. A full summary of the genetic data (mutation, copy number, gene expression, and genetic perturbation fitness scores) used to select the 4 cell lines (MDA-MB-231, KP-4, A549 and LS180 cells) for this study is shown in Supplementary Table 4.In summary, a “valid” TSG perturbation in our platform: 1) results in a fitness increase when knocked out in DepMap, 2) is expressed from two WT copies, and 3) is in a cellular context that matches its tumor of origin. Since pooled CRISPR screens are not yet empowered to generate specific edits, we rely on the GoF mutation to be present in the cell line background. For a GoF oncogene to be a “valid” signal in our CODA-PGX platform, it similarly needs to yield a fitness effect in DepMap; this time negative demonstrating the dependency on that oncogene. It similarly needs to be expressed23, present in 2 or more copies, and the cell line needs to reflect the tumor of origin. We then optimized cell line selection to maximize representation of “valid” perturbations (see Methods for details) and quantified screening results on “valid” perturbations whenever possible.
CODA-PGX platform testing and validation
To test if our approach could accurately identify known SL interactions, we performed proof-of-concept experiments with four currently marketed PARP (poly-ADP ribose polymerase) inhibitor drugs – the only approved class of drug currently available to treat LoF TSG drivers. PARPs are nuclear proteins that detect and initiate a cellular response to repair single strand breaks in DNA, and PARP inhibitors cause cells to accumulate DNA damage and result in cell death. Accordingly, PARP inhibitors have established SL interactions with genes involved in homologous recombination, such as BRCA1, BRCA2, ATM, ATR, RAD21, and STAG224,25.In order to validate the platform, establish parameters for selecting an appropriate drug screening concentration, and to explore the effect of cell population doubling (PD), we screened all four currently approved PARP inhibitors in several cell line backgrounds (see Methods) while varying the above parameters. We examined viability inhibitory concentration (IC) values from IC5-IC50 and population doublings from 5–15. We determined that IC15 and 7–10 PDs are optimal to maximize the signal of known clinically validated SL interactions (Supplementary Fig. 4; Fig. 2). From the various screens, we consistently identified well-established SL genes with PARPi’s, including BRCA1, STAG2, ATM, and RAD21 (Fig. 2), as well as other genes, including FBXW7, PPP6C and PTEN (Supplementary Fig. 4), that are also reported to be involved in DNA repair processes26,27,28. Not surprisingly, at lower concentrations there are no significant signals, and at higher concentrations general drug toxicity/resistance effects start to dominate. For example, for hits above IC30, metabolic genes (e.g. mTOR) and genes involved in more general biology (e.g. DIS3, an exoribonuclease), reproducibly dominate the most significant sensitizers (Supplementary Fig. 4).Fig. 2: CODA-PGX platform proof-of-concept experiment with PARP inhibitors in MDA-MB-231 cells.Volcano plots (see Methods) depict SL interactions between driver mutations and four currently marketed PARP inhibitor pharmaceuticals. Known SL interactions that were identified by CODA-PGX are highlighted in red. MDA-MB-231 cells were treated with 3.44 uM Olaparib, 0.002 uM Talazoparib, or 0.368 uM Rucaparib and grown for ~15 PDs, whereas cells treated with 0.200 uM Niraparib were grown for 10 PDs. The identity of all the hits can be found in Supplementary Fig. 4.Mutation-induced drug sensitivity landscape of common chemotherapeuticsTo discover novel patient selection possibilities for existing treatments, we screened a selection of marketed and characterized drugs. Specifically, we analyzed SL interactions of our pooled driver mutations in 85 cell line-drug combinations and visualized the results in a heatmap (Fig. 3a, Supplementary Fig. 5, Supplementary Table 4) where cell line-drug combinations are plotted on the y-axis and driver mutations on the x-axis. Yellow squares indicate sensitization in the presence of drug (i.e., slower growth compared to control), while blue squares indicate that a drug conferred a fitness benefit (i.e., faster growth compared to control). Data were clustered and diagonalized to highlight specific drug-driver combinations along the diagonal29. Vertical stretches of yellow squares indicate that a driver mutation was significantly sensitizing in multiple screens. In general, these patterns correspond to drugs within a single class (e.g., PARP inhibitors: Olaparib, Rucaparib and Talazoparib; Fig. 3a, Box 1).Fig. 3: Mutation-induced drug sensitivity landscape of common chemotherapeutics.a Heatmap of driver mutations (x-axis) and cell line/drug combinations (y axis) where yellow squares indicate reduced fitness in the presence of drug and blue squares indicate increased fitness. The sensitization/suppressive scale is log2(“fold-change”) or ratio of drug treated to vehicle/DMSO treated cells which have the same gene knocked out. The scale is from -2 to +2 (i.e. showing effect size up to ±4-fold up/down). Values with ranksum p > 0.1 were set to zero. b PCA of interactions identified by CODA-PGX. c Dose–response curves of isogenic KP-4 cells with and without sensitizing mutations identified by CODA-PGX. WT and KO cells were treated with a 5-fold dilution series of Talazoparib, Olaparib, Carboplatin or Taladagib. All pairwise comparisons of isogenic treated sets of were significant (p < 0.01; see Methods).Interestingly, we note that both Cisplatin and Carboplatin also occur within this same vertical pattern, suggesting that they have similar SL interactions to PARP inhibitors. Platinum-based chemotherapies are a class of chemotherapy drugs whose mechanism of action involves delivery of platinum ions to cells that cause DNA crosslinks which inhibit DNA replication and synthesis in all dividing cells. Although not currently prescribed as targeted therapy, platinum-based chemotherapies have been associated with improved outcomes in BRCA1- and BRCA2-deficient tumors30,31,32.A second group of SL interactions is detected between the TP53BP1 gene and Etoposide and Bleomycin (Fig. 3a, Box 2). TP53BP1(tumor-suppressor p53 binding protein 1) has a known role in promoting the non-homologous end joining pathway to repair DNA double strand breaks33,34. Etoposide and Bleomycin are both chemotherapy drugs, often used in combination, known to impact DNA damage and repair; Etoposide prevents the repair of DNA damage while Bleomycin causes DNA degradation. While neither drug is currently used in targeted therapy approaches, the observed SL interaction between these drugs and TP53BP1 is consistent with reports in the literature35,36,37 and with the mechanism of action of the drugs and biological role of the TP53BP1 gene.We also noted a SL interaction between TBL1XR1 and Taladegib (Fig. 3a, Box 3). Taladegib is a small molecule antagonist to the smoothened receptor designed to inhibit Hedgehog (Hh) signaling, a pathway that promotes cell division, differentiation, and migration, and is often upregulated in cancer cells38. TBL1XR1 has been identified as a potential therapeutic target39 as mutations are widely documented in cancer40. While no direct connection between TBL1XR1 and Hh signaling is known, the gene is tightly linked to Wnt signaling41 and there is well-established crosstalk between these two signaling pathways which often regulate similar biological functions (cell division, differentiation, and migration)42,43.These data also allow us to make some general observations about the value and potential impact of CODA-PGX. We note that many of the driver mutations identified as SL hits are consistent with the known mechanism of action of the drugs tested. Specifically, both platinum-based chemotherapy and PARP inhibitors are SL with chromatin-associated proteins STAG2 and EZH2, and BRCA1 (known to function in DNA damage repair). We observed smaller but significant interactions between Taladegib and FBXW7 and SMARCB1 genes, both of which are known to regulate Hh signaling44,45. Additional noteworthy interactions are summarized in Supplementary Table 7, including a variety of suppressive interactions. For example, MDA-MB-231 cells, which have an activating KRAS mutation, were less sensitive to Gefitinib (Fig. 3a, Box 4), a previously reported effect observed in the clinic46. Taken together, these observations suggest that in addition to identifying established SL interactions between drugs and driver genes, CODA-PGX uncovers new and biologically relevant connections.To explore whether drug sensitizing mutation signals arising from CODA-PGX screening are confounded by any technical variables, we performed a PCA analysis (Fig. 3b, see Methods). Briefly, the proximity of data points (each of which represents one screen – or cell line/drug combination) in the plot reflects how similar they are. We observe that, in general, drug mechanism of action is the dominant variable that drives the signal, not the cell line in which the screen was conducted, or the drug concentration. For example, consistent with the heatmap (Fig. 3a), we observed that PARP inhibitors and platinum-based chemotherapy cluster tightly together, as well as clustering of drugs with other unique mechanisms (Fig. 3b). This suggests that the druggable dependencies arising from cancer driver biology, when filtered appropriately (e.g. on driver context; see Methods), tend to persist across different cell of origin contexts.To orthogonally validate some novel hits from our CODA-PGX screens, we generated isogenic pairs of cell lines differing only by the CRISPR/Cas9-mediated KO of the indicated driver gene (EZH2, TBL1XR1, FBXW7) in KP-4 cells. We then compared WT and KO isogenic lines for their sensitivity to drugs tested in our CODA-PGX platform. In all cases, we observed greater inhibition of growth in the KO lines compared to WT lines at the same drug concentration (Fig. 3c). This confirms that sensitizing driver mutations identified by our CODA-PGX platform indeed reduce cellular fitness in the presence of drug.In vivo validation of STAG2 – a novel Carboplatin-sensitizing mutationWe next investigated the ability of CODA-PGX to identify driver mutations that improve the efficacy of specific drugs. We looked specifically at Carboplatin, a chemotherapy drug commonly used to treat ovarian and other cancers, but not designated for targeted therapy. Our preliminary screening and heatmap analysis (Fig. 3) identified that Carboplatin shares several SL interactions with PARP inhibitor drugs. To explore whether there are additional driver mutations that improve the efficacy of Carboplatin, we screened for SL interactions in three different cell lines (MDA-MB-231, A549, KP-4) using our CODA-PGX approach (Fig. 4a and Supplementary Fig. 6). In addition to BRCA1, mutations in STAG2, ATM, DDX3X, EZH2, and APC showed decreased viability in the presence of Carboplatin compared to the vehicle control (Fig. 4a). Of these, ATM, STAG2 and EZH2 have known involvement in homologous recombination and DNA repair24,25,47. Since STAG2 was identified as a strong hit and is not currently used as an HRD patient-selection marker, we pursued more extensive validation of its potential as a sensitizing mutation. We also noted from our data that the KO of TP53 causes resistance to Carboplatin (Fig. 4a) and aligns with documented resistance mechanisms in the clinic48.Fig. 4: In vivo validation of STAG2 – a novel Carboplatin-sensitizing mutation.a Volcano plot identifying Carboplatin-sensitizing mutations detected in A549 cells. b Dose–response curve showing inhibitory effects of Carboplatin on WT (pink) and STAG2 KO (blue) OvCar-3 cells. The shift toward increased sensitivity in STAG2 KO cells was significant (p < 0.001; see Methods). c–f Tumor growth curves of cell-derived xenografts (CDXs) from WT and STAG2 KO OvCar-3 cells with and without Carboplatin treatment. n = 10 mice per group, data presented as the mean ± SEM. A two-way ANOVA test was used. WT OvCar-3 Vehicle v.s. STAG2 KO OvCar-3 Vehicle, p = 0.16. WT OvCar-3 Vehicle v.s. Carboplatin, p = 0.17; STAG2 KO Vehicle v.s. Carboplatin, p = 0.006; *P < 0.05 was considered statistically significant. g Individual tumors at Day 10 from all groups of mice were collected and the weights were measured. One-tail T-test was used for the statistical analysis. P < 0.05 was considered statistically significant. h Normalized tumor growth inhibition (TGI) values for WT and STAG2 KO CDXs shown in (c). i Plot of tumor volume over time for a gastric STAG2 LoF model (CrownBio model GA0151) patient-derived xenograft (PDX). n = 8 mice per group. j Plot of tumor volume over time for additional Crown Bio PDX models of gastric, lung, and ovarian cancer cells. Turquoise line corresponds to model shown in (i).We first screened STAG2 LoF in a controlled viability assay. We compared the growth of cells with a CRISPR-generated KO of STAG2 to WT cells in the OvCar-3 ovarian cancer cell line (see Methods) in presence of increasing concentrations of Carboplatin and plotted a dose–response curve (Fig. 4b). The results confirm that the absence of STAG2 causes greater growth inhibition at lower concentrations of Carboplatin, indicating a SL response between STAG2 and Carboplatin.Cancer cell-derived xenografts (CDX) provide a valuable system in which to validate the genetic sensitization of LoF mutations in vivo. We established CDXs in mice by injecting NSG mice with either WT or STAG2 KO OvCar-3 cells. We observed that the CRISPR-generated STAG2 KO cells grew faster than WT cells (Fig. 4c, d, KO Vehicle vs WT Vehicle), but the difference was modest and didn’t reach statistical significance (p = 0.16).While the overall inhibitory trend of Carboplatin treatment in the WT group didn’t reach statistical significance (p = 0.17), STAG2 KO tumors were more sensitive to Carboplatin treatment and displayed significantly slower tumor growth in response to the drug (p = 0.006) (Fig. 4e, f). In addition, the tumor weight measurements on Day 10 showed a consistent differential effect of Carboplatin between WT and STAG2 KO groups (Fig. 4g). We calculated the normalized effect of Carboplatin on STAG2 KO vs WT as tumor growth inhibition and plotted these data (Fig. 4h). We observed that the normalized effect of Carboplatin on STAG2 mutant growth is greater than the normalized effect of Carboplatin on WT cells. The study was ended on Day 10 because several tumors derived from the vehicle groups of WT and STAG2 KO-injected mice reached the approved endpoint tumor size of 1500 mm3. Overall, the inhibitory effect within the same genetic background indicates that Carboplatin has a greater effect on the growth of STAG2 KO tumors.We next used a patient-derived xenograft (PDX) system to further validate these observations in a pre-clinical system that closely models clinical response. PDX models generally capture more of the heterogeneity of tumors compared to CDX models and have been shown to have reasonable predictive capacity for genetically-guided precision oncology49. To this end, we partnered with Crown Bio to interrogate their collection of Carboplatin standard-of-care in ovarian, gastric, and lung cancer PDX models. We plotted tumor volumes from model GA0151, a gastric tumor PDX with STAG2 LoF mutant model treated with vehicle or 40 mg/kg Carboplatin Q4Dx4 (Fig. 4i) and observed that the presence of the LoF mutation slowed tumor growth compared to WT when treated with Carboplatin. We expanded this analysis to examine tumors in the ten available models for these three cancer types and plotted the difference in tumor volume between WT and mutant for BRCA1, ATM, STAG2, and APC genes (Fig. 4j). In all cases, the presence of the driver mutation in the PDX showed reduced tumor growth with Carboplatin compared to WT PDXs of the same cancer type. These data indicate that the mutations identified by CODA-PGX were sensitive to Carboplatin and, by extension, that CODA-PGX can accurately identify sensitizing mutations.
Retrospective clinical validation of hits
The ultimate test for a new therapy is clinical efficacy, and the gold standard measure of this is a prospective randomized control clinical trial. In the last 10 years we have seen a dramatic increase in cancer patient data collection, including NGS/molecular readouts that capture the mutation spectrum of the patient’s tumor, treatment, and outcomes. The Cancer Genome Atlas contains more than ten thousand patients with tumors profiled by whole-exome sequencing and their treatment histories and outcome meticulously documented. More recently, the AACR’s Project Genie is a cancer registry of patient genomic and demographic data for more than 150,000 patients across 19 oncology centers50. While molecular profiling in Project Genie is largely limited to targeted sequencing of cancer associated genes, and drug treatment/outcome data is not yet broadly available, an industry-academia partnered effort known as the Biopharma Collaborative (BPC)51 is gradually releasing these curated RWE datasets. We downloaded all mutation, treatment history, and clinical outcome data from TCGA for 33 cancer types and BPC data from lung and colon cancers that were available at the time of publication.To evaluate the ability of CODA-PGX to predict drug-driver interactions likely to yield meaningful improvements in clinical outcomes, we considered the input data from our exploration of the mutation-induced drug sensitivity landscape of common chemotherapeutics (Fig. 3). For each cell line/drug combination (i.e., row), we identified genes that sensitized the cells to drug treatment (negative on volcano plot) at p < 0.05 and a minimum fold-change of 2. In order to avoid potentially confounding variables arising from cross-cancer analysis, we selected the cancer type with the highest number of patients where the drug in question was used as a frontline therapy. We then subset these patients into two groups based on their tumor mutation status: with or without one or more of the sensitizing mutations identified in the screen (see Methods for details). We then compared overall survival (OS) of these populations based on driver mutation status (Fig. 5a, b). With our CODA-PGX hit linking SMARCB1 LoF to Oxaliplatin efficacy (Fig. 3a, Box 5), we compared overall survival of colorectal cancer patients who had mutations to the SMARCB1 gene and observed that patients with mutations treated with Oxaliplatin (Fig. 3a, Box 5) survived longer than patients without SMARCB1 mutations (Fig. 5a, p = 0.033, logrank test). A parallel analysis of glioma and glioblastoma patients treated with Temozolomide revealed significantly improved survival in patients with CODA-PGX-informed driver mutation-positive tumors compared to patients without the indicated driver mutations (Fig. 5b, p < 0.001, logrank test). We also repeated the above analysis for mutations that suppressed drug sensitivity (positive in volcano plots, blue squares in Fig. 3; e.g. Fig. 5c). We observed significantly decreased probability of survival in non-small cell lung cancer patients with these mutations compared to patients without the mutations who received the same treatment (p = 0.05, logrank test). To confirm that the mutations themselves are not responsible for the difference in OS, whenever possible (i.e., data exists), we repeated this analysis for the next best represented drug and observed no difference in overall survival (Supplementary Fig. 7A–C). We attempted to substantiate our findings related to increased survival in STAG2 mutant ovarian cancer patients treated with Carboplatin but were limited by the sample size. We did, however, capture a sufficient number of Carboplatin treated ovarian patients with either STAG2, EZH2, BRCA1, APC, or DDX3X mutations (recurring sensitizing mutations for Carboplatin identified by CODA-PGX) and observed an increase in OS with a p-value of 0.089 (Supplementary Fig. 8).Fig. 5: Retrospective clinical validation of CODA-PGX hits.a Plot of survival vs. time for colorectal cancer patients with (Mut/Del) mutations to SMARCB1 gene treated with Oxaliplatin compared to WT patients. b Plot of survival versus time for brain cancer patients with (Mut/Del) or without (WT) CODA-PGX-identified sensitizing mutations, treated with Temozolomide. c Plot of survival versus time for lung cancer patients with (Mut/Del) or without (WT) drug-suppressive mutations, treated with Cisplatin. d CODA-PGX predictions are generally associated with expected OS outcomes: the majority of sensitizing mutation/drug combinations are associated with improved OS, and the majority of suppressive mutation/drug combinations with worse OS.These data suggest that the CODA-PGX platform can identify common cancer-causing mutations with potential to guide treatment decisions in the clinic. To more broadly assess whether CODA-PGX is predictive of clinical outcomes, we repeated the analysis above for all drugs and asked what proportion of the time CODA-PGX identified hit/driver mutations predict the expected trend in OS compared to patients without these mutations in their tumors. Despite availability of RWE and molecular data from thousands of patients, subsetting on cancer type, the drug in question being used as frontline therapy, and the presence of the mutation(s), yielded 41 drug-mutation combinations (i.e. “analyses”) where we were able to assess an association with overall survival (OS; Fig. 5d). Both sensitizing and suppressive mutations yielded an enrichment for OS in the expected directions, with sensitizing mutations associated with improved OS, and vice versa. These effects were not caused by the mutations alone (Supplementary Fig. 7D). Agreement was higher at more significant OS differences (i.e., to the left of Fig. 5d), suggesting that statistical power may be limiting the overall analysis and that agreement overall would be higher if more patient data were available. Given that statistical power is severely limited for most of the CODA-PGX predictions due to patient data limitations, we reasoned that less significant results (e.g. p-value of 0.25) should nonetheless have predictive capacity for the direction of OS association. Sensitizing mutations should lead to longer OS, and suppressive mutations lower OS. If the platform hits were non-informative, we should observe ~50% agreement with OS since it has two possible states across any analysis regardless of the KM p-value threshold. Instead, we observed a high rate of OS concordance when KM p-values are significant (likely due to sufficient sample numbers) and the rate of concordance gets closer to 0.5 as analyses with lower KM thresholds are included (Fig. 5d). Repeating the same analysis, but randomly substituting in a different drug that has at least as many patients represented (i.e. power is greater) for each analysis, effectively yields no analyses that have significant OS differences (Supplementary Fig. 7D).Overall, when requiring that at least 10 patients with the mutation exist in the analysis, 12/16 (75%) are associated with the predicted OS direction and the majority are at KM logrank p < 0.5. Taken together, these data support that CODA-PGX predicted driver mutation–drug relationships are enriched for patient selection criteria with clinical relevance.
Pharmacogenomic discovery of genetically targeted cancer therapies optimized against clinical outcomes
