Privacy preserving validation for multiomic prediction models.

Reproducibility of results obtained using ribonucleic acid (RNA) data across labs remains a major hurdle in cancer research. Often, molecular predictors trained on one dataset cannot be applied to another due to differences in RNA library preparation and quantification, which inhibits the validation of predictors across labs. While current RNA correction algorithms reduce these differences, they require simultaneous access to patient-level data from all datasets, which necessitates the sharing of training data for predictors when sharing predictors. Here, we describe SpinAdapt, an unsupervised RNA correction algorithm that enables the transfer of molecular models without requiring access to patient-level data. It computes data corrections only via aggregate statistics of each dataset, thereby maintaining patient data privacy. Despite an inherent trade-off between privacy and performance, SpinAdapt outperforms current correction methods, like Seurat and ComBat, on publicly available cancer studies, including TCGA and ICGC. Furthermore, SpinAdapt can correct new samples, thereby enabling unbiased evaluation on validation cohorts. We expect this novel correction paradigm to enhance research reproducibility and to preserve patient privacy.

Landscape of homologous recombination deficiencies in solid tumours: analyses of two independent genomic datasets.

BACKGROUND: DNA repair deficiencies are characteristic of cancer and homologous recombination deficiency (HRD) is the most common. HRD sensitizes tumour cells to PARP inhibitors so it is important to understand the landscape of HRD across different solid tumour types. METHODS: Germline and somatic BRCA mutations in breast and ovarian cancers were evaluated using sequencing data from The Cancer Genome Atlas (TCGA) database. Secondly, a larger independent genomic dataset was analysed to validate the TCGA results and determine the frequency of germline and somatic mutations across 15 different candidate homologous recombination repair (HRR) genes, and their relationship with the genetic events of bi-allelic loss, loss of heterozygosity (LOH) and tumour mutation burden (TMB). RESULTS: Approximately one-third of breast and ovarian cancer BRCA mutations were somatic. These showed a similar degree of bi-allelic loss and clinical outcomes to germline mutations, identifying potentially 50% more patients that may benefit from precision treatments. HRR mutations were present in sizable proportions in all tumour types analysed and were associated with high TMB and LOH scores. We also identified numerous BRCA reversion mutations across all tumour types. CONCLUSIONS: Our results will facilitate future research into the efficacy of precision oncology treatments, including PARP and immune checkpoint inhibitors.

Validation of genomic and transcriptomic models of homologous recombination deficiency in a real-world pan-cancer cohort.

BACKGROUND: With the introduction of DNA-damaging therapies into standard of care cancer treatment, there is a growing need for predictive diagnostics assessing homologous recombination deficiency (HRD) status across tumor types. Following the strong clinical evidence for the utility of DNA-sequencing-based HRD testing in ovarian cancer, and growing evidence in breast cancer, we present analytical validation of the Tempus HRD-DNA test. We further developed, validated, and explored the Tempus HRD-RNA model, which uses gene expression data from 16,750 RNA-seq samples to predict HRD status from formalin-fixed paraffin-embedded tumor samples across numerous cancer types. METHODS: Genomic and transcriptomic profiling was performed using next-generation sequencing from Tempus xT, Tempus xO, Tempus xE, Tempus RS, and Tempus RS.v2 assays on 48,843 samples. Samples were labeled based on their BRCA1, BRCA2 and selected Homologous Recombination Repair pathway gene (CDK12, PALB2, RAD51B, RAD51C, RAD51D) mutational status to train and validate HRD-DNA, a genome-wide loss-of-heterozygosity biomarker, and HRD-RNA, a logistic regression model trained on gene expression. RESULTS: In a sample of 2058 breast and 1216 ovarian tumors, BRCA status was predicted by HRD-DNA with F1-scores of 0.98 and 0.96, respectively. Across an independent set of 1363 samples across solid tumor types, the HRD-RNA model was predictive of BRCA status in prostate, pancreatic, and non-small cell lung cancer, with F1-scores of 0.88, 0.69, and 0.62, respectively. CONCLUSIONS: We predict HRD-positive patients across many cancer types and believe both HRD models may generalize to other mechanisms of HRD outside of BRCA loss. HRD-RNA complements DNA-based HRD detection methods, especially for indications with low prevalence of BRCA alterations.

VarDict: a novel and versatile variant caller for next-generation sequencing in cancer research.

Accurate variant calling in next generation sequencing (NGS) is critical to understand cancer genomes better. Here we present VarDict, a novel and versatile variant caller for both DNA- and RNA-sequencing data. VarDict simultaneously calls SNV, MNV, InDels, complex and structural variants, expanding the detected genetic driver landscape of tumors. It performs local realignments on the fly for more accurate allele frequency estimation. VarDict performance scales linearly to sequencing depth, enabling ultra-deep sequencing used to explore tumor evolution or detect tumor DNA circulating in blood. In addition, VarDict performs amplicon aware variant calling for polymerase chain reaction (PCR)-based targeted sequencing often used in diagnostic settings, and is able to detect PCR artifacts. Finally, VarDict also detects differences in somatic and loss of heterozygosity variants between paired samples. VarDict reprocessing of The Cancer Genome Atlas (TCGA) Lung Adenocarcinoma dataset called known driver mutations in KRAS, EGFR, BRAF, PIK3CA and MET in 16% more patients than previously published variant calls. We believe VarDict will greatly facilitate application of NGS in clinical cancer research.

BubbleTree: an intuitive visualization to elucidate tumoral aneuploidy and clonality using next generation sequencing data.

Tumors are characterized by properties of genetic instability, heterogeneity, and significant oligoclonality. Elucidating this intratumoral heterogeneity is challenging but important. In this study, we propose a framework, BubbleTree, to characterize the tumor clonality using next generation sequencing (NGS) data. BubbleTree simultaneously elucidates the complexity of a tumor biopsy, estimating cancerous cell purity, tumor ploidy, allele-specific copy number, and clonality and represents this in an intuitive graph. We further developed a three-step heuristic method to automate the interpretation of the BubbleTree graph, using a divide-and-conquer strategy. In this study, we demonstrated the performance of BubbleTree with comparisons to similar commonly used tools such as THetA2, ABSOLUTE, AbsCN-seq and ASCAT, using both simulated and patient-derived data. BubbleTree outperformed these tools, particularly in identifying tumor subclonal populations and polyploidy. We further demonstrated BubbleTree's utility in tracking clonality changes from patients' primary to metastatic tumor and dating somatic single nucleotide and copy number variants along the tumor clonal evolution. Overall, the BubbleTree graph and corresponding model is a powerful approach to provide a comprehensive spectrum of the heterogeneous tumor karyotype in human tumors. BubbleTree is R-based and freely available to the research community (https://www.bioconductor.org/packages/release/bioc/html/BubbleTree.html).

Multi-omic measurement of mutually exclusive loss-of-function enriches for candidate synthetic lethal gene pairs.

BACKGROUND: Identification of synthetic lethal interactions in cancer cells could offer promising new therapeutic targets. Large-scale functional genomic screening presents an opportunity to test large numbers of cancer synthetic lethal hypotheses. Methods enriching for candidate synthetic lethal targets in molecularly defined cancer cell lines can steer effective design of screening efforts. Loss of one partner of a synthetic lethal gene pair creates a dependency on the other, thus synthetic lethal gene pairs should never show simultaneous loss-of-function. We have developed a computational approach to mine large multi-omic cancer data sets and identify gene pairs with mutually exclusive loss-of-function. Since loss-of-function may not always be genetic, we look for deleterious mutations, gene deletion and/or loss of mRNA expression by bimodality defined with a novel algorithm BiSEp. RESULTS: Applying this toolkit to both tumour cell line and patient data, we achieve statistically significant enrichment for experimentally validated tumour suppressor genes and synthetic lethal gene pairings. Notably non-reliance on genetic loss reveals a number of known synthetic lethal relationships otherwise missed, resulting in marked improvement over genetic-only predictions. We go on to establish biological rationale surrounding a number of novel candidate synthetic lethal gene pairs with demonstrated dependencies in published cancer cell line shRNA screens. CONCLUSIONS: This work introduces a multi-omic approach to define gene loss-of-function, and enrich for candidate synthetic lethal gene pairs in cell lines testable through functional screens. In doing so, we offer an additional resource to generate new cancer drug target and combination hypotheses. Algorithms discussed are freely available in the BiSEp CRAN package at http://cran.r-project.org/web/packages/BiSEp/index.html .

Network-driven cancer cell avatars for combination discovery and biomarker identification for DNA damage response inhibitors.

Combination therapy is well established as a key intervention strategy for cancer treatment, with the potential to overcome monotherapy resistance and deliver a more durable efficacy. However, given the scale of unexplored potential target space and the resulting combinatorial explosion, identifying efficacious drug combinations is a critical unmet need that is still evolving. In this paper, we demonstrate a network biology-driven, simulation-based solution, the Simulated Cell™. Integration of omics data with a curated signaling network enables the accurate and interpretable prediction of 66,348 combination-cell line pairs obtained from a large-scale combinatorial drug sensitivity screen of 684 combinations across 97 cancer cell lines (BAC = 0.62, AUC = 0.7). We highlight drug combination pairs that interact with DNA Damage Response pathways and are predicted to be synergistic, and deep network insight to identify biomarkers driving combination synergy. We demonstrate that the cancer cell 'avatars' capture the biological complexity of their in vitro counterparts, enabling the identification of pathway-level mechanisms of combination benefit to guide clinical translatability.

Mixed responses to targeted therapy driven by chromosomal instability through p53 dysfunction and genome doubling.

The phenomenon of mixed/heterogenous treatment responses to cancer therapies within an individual patient presents a challenging clinical scenario. Furthermore, the molecular basis of mixed intra-patient tumor responses remains unclear. Here, we show that patients with metastatic lung adenocarcinoma harbouring co-mutations of EGFR and TP53, are more likely to have mixed intra-patient tumor responses to EGFR tyrosine kinase inhibition (TKI), compared to those with an EGFR mutation alone. The combined presence of whole genome doubling (WGD) and TP53 co-mutations leads to increased genome instability and genomic copy number aberrations in genes implicated in EGFR TKI resistance. Using mouse models and an in vitro isogenic p53-mutant model system, we provide evidence that WGD provides diverse routes to drug resistance by increasing the probability of acquiring copy-number gains or losses relative to non-WGD cells. These data provide a molecular basis for mixed tumor responses to targeted therapy, within an individual patient, with implications for therapeutic strategies.

Contrived Materials and a Data Set for the Evaluation of Liquid Biopsy Tests: A Blood Profiling Atlas in Cancer (BLOODPAC) Community Study.

The Blood Profiling Atlas in Cancer (BLOODPAC) Consortium is a collaborative effort involving stakeholders from the public, industry, academia, and regulatory agencies focused on developing shared best practices on liquid biopsy. This report describes the results from the JFDI (Just Freaking Do It) study, a BLOODPAC initiative to develop standards on the use of contrived materials mimicking cell-free circulating tumor DNA, to comparatively evaluate clinical laboratory testing procedures. Nine independent laboratories tested the concordance, sensitivity, and specificity of commercially available contrived materials with known variant-allele frequencies (VAFs) ranging from 0.1% to 5.0%. Each participating laboratory utilized its own proprietary evaluation procedures. The results demonstrated high levels of concordance and sensitivity at VAFs of >0.1%, but reduced concordance and sensitivity at a VAF of 0.1%; these findings were similar to those from previous studies, suggesting that commercially available contrived materials can support the evaluation of testing procedures across multiple technologies. Such materials may enable more objective comparisons of results on materials formulated in-house at each center in multicenter trials. A unique goal of the collaborative effort was to develop a data resource, the BLOODPAC Data Commons, now available to the liquid-biopsy community for further study. This resource can be used to support independent evaluations of results, data extension through data integration and new studies, and retrospective evaluation of data collection.

Knowledge graph-based recommendation framework identifies drivers of resistance in EGFR mutant non-small cell lung cancer.

Resistance to EGFR inhibitors (EGFRi) presents a major obstacle in treating non-small cell lung cancer (NSCLC). One of the most exciting new ways to find potential resistance markers involves running functional genetic screens, such as CRISPR, followed by manual triage of significantly enriched genes. This triage process to identify 'high value' hits resulting from the CRISPR screen involves manual curation that requires specialized knowledge and can take even experts several months to comprehensively complete. To find key drivers of resistance faster we build a recommendation system on top of a heterogeneous biomedical knowledge graph integrating pre-clinical, clinical, and literature evidence. The recommender system ranks genes based on trade-offs between diverse types of evidence linking them to potential mechanisms of EGFRi resistance. This unbiased approach identifies 57 resistance markers from >3,000 genes, reducing hit identification time from months to minutes. In addition to reproducing known resistance markers, our method identifies previously unexplored resistance mechanisms that we prospectively validate.

The landscape of therapeutic vulnerabilities in EGFR inhibitor osimertinib drug tolerant persister cells.

Third-generation EGFR tyrosine kinase inhibitors (EGFR-TKIs), including osimertinib, an irreversible EGFR-TKI, are important treatments for non-small cell lung cancer with EGFR-TKI sensitizing or EGFR T790M resistance mutations. While patients treated with osimertinib show clinical benefit, disease progression and drug resistance are common. Emergence of de novo acquired resistance from a drug tolerant persister (DTP) cell population is one mechanism proposed to explain progression on osimertinib and other targeted cancer therapies. Here we profiled osimertinib DTPs using RNA-seq and ATAC-seq to characterize the features of these cells and performed drug screens to identify therapeutic vulnerabilities. We identified several vulnerabilities in osimertinib DTPs that were common across models, including sensitivity to MEK, AURKB, BRD4, and TEAD inhibition. We linked several of these vulnerabilities to gene regulatory changes, for example, TEAD vulnerability was consistent with evidence of Hippo pathway turning off in osimertinib DTPs. Last, we used genetic approaches using siRNA knockdown or CRISPR knockout to validate AURKB, BRD4, and TEAD as the direct targets responsible for the vulnerabilities observed in the drug screen.

DeePaN: deep patient graph convolutional network integrating clinico-genomic evidence to stratify lung cancers for immunotherapy.

Immuno-oncology (IO) therapies have transformed the therapeutic landscape of non-small cell lung cancer (NSCLC). However, patient responses to IO are variable and influenced by a heterogeneous combination of health, immune, and tumor factors. There is a pressing need to discover the distinct NSCLC subgroups that influence response. We have developed a deep patient graph convolutional network, we call "DeePaN", to discover NSCLC complexity across data modalities impacting IO benefit. DeePaN employs high-dimensional data derived from both real-world evidence (RWE)-based electronic health records (EHRs) and genomics across 1937 IO-treated NSCLC patients. DeePaN demonstrated effectiveness to stratify patients into subgroups with significantly different (P-value of 2.2 × 10-11) overall median survival of 20.35 months and 9.42 months post-IO therapy. Significant differences in IO outcome were not seen from multiple non-graph-based unsupervised methods. Furthermore, we demonstrate that patient stratification from DeePaN has the potential to augment the emerging IO biomarker of tumor mutation burden (TMB). Characterization of the subgroups discovered by DeePaN indicates potential to inform IO therapeutic insight, including the enrichment of mutated KRAS and high blood monocyte count in the IO beneficial and IO non-beneficial subgroups, respectively. Our work has proven the concept that graph-based AI is feasible and can effectively integrate high-dimensional genomic and EHR data to meaningfully stratify cancer patients on distinct clinical outcomes, with potential to inform precision oncology.

BloodPAC Data Commons for Liquid Biopsy Data.

PURPOSE: The Blood Profiling Atlas in Cancer (BloodPAC) Data Commons (BPDC) is being developed and is operated by the public-private BloodPAC Consortium to support the liquid biopsy community. It is an interoperable data commons with the ultimate aim of serving as a recognized source of valid scientific evidence for liquid biopsy assays for industry, academia, and standards and regulatory stakeholders. METHODS: The BPDC is implemented using the open source Gen3 data commons platform (https://gen3.org). In particular, the BPDC Data Exploration Portal, BPDC Data Submission Portal, the BPDC Workspace Hub, and the BloodPAC application programming interface (API) were all automatically generated from the BloodPAC Data Model using the Gen3 data commons platform. BPDC uses Gen3's implementation of the data commons framework services so that it can interoperate through secure, compliant APIs with other data commons using data commons framework service, such as National Cancer Institute's Cancer Research Data Commons. RESULTS: The BPDC contains 57 studies and projects spanning more than 4,100 cases. This amounts to 5,700 aliquots (blood plasma, serum, or a contrived sample) that have been subjected to a liquid biopsy assay, quantified, and then contributed by members of the BloodPAC Consortium. In all, there are more than 31,000 files in the commons as of December 2020. We describe the BPDC, the data it manages, the process that the BloodPAC Consortium used to develop it, and some of the applications that have been developed using its API. CONCLUSION: The BPDC has been the data platform used by BloodPAC during the past 4 years to manage the data for the consortium and to provide workspaces for its working groups.

A pan-cancer organoid platform for precision medicine.

Patient-derived tumor organoids (TOs) are emerging as high-fidelity models to study cancer biology and develop novel precision medicine therapeutics. However, utilizing TOs for systems-biology-based approaches has been limited by a lack of scalable and reproducible methods to develop and profile these models. We describe a robust pan-cancer TO platform with chemically defined media optimized on cultures acquired from over 1,000 patients. Crucially, we demonstrate tumor genetic and transcriptomic concordance utilizing this approach and further optimize defined minimal media for organoid initiation and propagation. Additionally, we demonstrate a neural-network-based high-throughput approach for label-free, light-microscopy-based drug assays capable of predicting patient-specific heterogeneity in drug responses with applicability across solid cancers. The pan-cancer platform, molecular data, and neural-network-based drug assay serve as resources to accelerate the broad implementation of organoid models in precision medicine research and personalized therapeutic profiling programs.

Identification of Intrinsic Drug Resistance and Its Biomarkers in High-Throughput Pharmacogenomic and CRISPR Screens.

High-throughput drug screens in cancer cell lines test compounds at low concentrations, thereby enabling the identification of drug-sensitivity biomarkers, while resistance biomarkers remain underexplored. Dissecting meaningful drug responses at high concentrations is challenging due to cytotoxicity, i.e., off-target effects, thus limiting resistance biomarker discovery to frequently mutated cancer genes. To address this, we interrogate subpopulations carrying sensitivity biomarkers and consecutively investigate unexpectedly resistant (UNRES) cell lines for unique genetic alterations that may drive resistance. By analyzing the GDSC and CTRP datasets, we find 53 and 35 UNRES cases, respectively. For 24 and 28 of them, we highlight putative resistance biomarkers. We find clinically relevant cases such as EGFRT790M mutation in NCI-H1975 or PTEN loss in NCI-H1650 cells, in lung adenocarcinoma treated with EGFR inhibitors. Interrogating the underpinnings of drug resistance with publicly available CRISPR phenotypic assays assists in prioritizing resistance drivers, offering hypotheses for drug combinations.

A statistical framework for assessing pharmacological responses and biomarkers using uncertainty estimates.

High-throughput testing of drugs across molecular-characterised cell lines can identify candidate treatments and discover biomarkers. However, the cells' response to a drug is typically quantified by a summary statistic from a best-fit dose-response curve, whilst neglecting the uncertainty of the curve fit and the potential variability in the raw readouts. Here, we model the experimental variance using Gaussian Processes, and subsequently, leverage uncertainty estimates to identify associated biomarkers with a new Bayesian framework. Applied to in vitro screening data on 265 compounds across 1074 cancer cell lines, our models identified 24 clinically established drug-response biomarkers, and provided evidence for six novel biomarkers by accounting for association with low uncertainty. We validated our uncertainty estimates with an additional drug screen of 26 drugs, 10 cell lines with 8 to 9 replicates. Our method is applicable to any dose-response data without replicates, and improves biomarker discovery for precision medicine.

Adenosine Signaling Is Prognostic for Cancer Outcome and Has Predictive Utility for Immunotherapeutic Response.

PURPOSE: There are several agents in early clinical trials targeting components of the adenosine pathway including A2AR and CD73. The identification of cancers with a significant adenosine drive is critical to understand the potential for these molecules. However, it is challenging to measure tumor adenosine levels at scale, thus novel, clinically tractable biomarkers are needed. EXPERIMENTAL DESIGN: We generated a gene expression signature for the adenosine signaling using regulatory networks derived from the literature and validated this in patients. We applied the signature to large cohorts of disease from The Cancer Genome Atlas (TCGA) and cohorts of immune checkpoint inhibitor-treated patients. RESULTS: The signature captures baseline adenosine levels in vivo (r 2 = 0.92, P = 0.018), is reduced after small-molecule inhibition of A2AR in mice (r 2 = -0.62, P = 0.001) and humans (reduction in 5 of 7 patients, 70%), and is abrogated after A2AR knockout. Analysis of TCGA confirms a negative association between adenosine and overall survival (OS, HR = 0.6, P < 2.2e-16) as well as progression-free survival (PFS, HR = 0.77, P = 0.0000006). Further, adenosine signaling is associated with reduced OS (HR = 0.47, P < 2.2e-16) and PFS (HR = 0.65, P = 0.0000002) in CD8+ T-cell-infiltrated tumors. Mutation of TGFß superfamily members is associated with enhanced adenosine signaling and worse OS (HR = 0.43, P < 2.2e-16). Finally, adenosine signaling is associated with reduced efficacy of anti-PD1 therapy in published cohorts (HR = 0.29, P = 0.00012). CONCLUSIONS: These data support the adenosine pathway as a mediator of a successful antitumor immune response, demonstrate the prognostic potential of the signature for immunotherapy, and inform patient selection strategies for adenosine pathway modulators currently in development.

Targeting melanoma's MCL1 bias unleashes the apoptotic potential of BRAF and ERK1/2 pathway inhibitors.

BRAF and MEK1/2 inhibitors are effective in melanoma but resistance inevitably develops. Despite increasing the abundance of pro-apoptotic BIM and BMF, ERK1/2 pathway inhibition is predominantly cytostatic, reflecting residual pro-survival BCL2 family activity. Here, we show that uniquely low BCL-XL expression in melanoma biases the pro-survival pool towards MCL1. Consequently, BRAF or MEK1/2 inhibitors are synthetic lethal with the MCL1 inhibitor AZD5991, driving profound tumour cell death that requires BAK/BAX, BIM and BMF, and inhibiting tumour growth in vivo. Combination of ERK1/2 pathway inhibitors with BCL2/BCL-w/BCL-XL inhibitors is stronger in CRC, correlating with a low MCL1:BCL-XL ratio; indeed the MCL1:BCL-XL ratio is predictive of ERK1/2 pathway inhibitor synergy with MCL1 or BCL2/BCL-w/BCL-XL inhibitors. Finally, AZD5991 delays acquired BRAFi/MEKi resistance and enhances the efficacy of an ERK1/2 inhibitor in a model of acquired BRAFi + MEKi resistance. Thus combining ERK1/2 pathway inhibitors with MCL1 antagonists in melanoma could improve therapeutic index and patient outcomes.

MEK1/2 inhibitor withdrawal reverses acquired resistance driven by BRAFV600E amplification whereas KRASG13D amplification promotes EMT-chemoresistance.

Acquired resistance to MEK1/2 inhibitors (MEKi) arises through amplification of BRAFV600E or KRASG13D to reinstate ERK1/2 signalling. Here we show that BRAFV600E amplification and MEKi resistance are reversible following drug withdrawal. Cells with BRAFV600E amplification are addicted to MEKi to maintain a precise level of ERK1/2 signalling that is optimal for cell proliferation and survival, and tumour growth in vivo. Robust ERK1/2 activation following MEKi withdrawal drives a p57KIP2-dependent G1 cell cycle arrest and senescence or expression of NOXA and cell death, selecting against those cells with amplified BRAFV600E. p57KIP2 expression is required for loss of BRAFV600E amplification and reversal of MEKi resistance. Thus, BRAFV600E amplification confers a selective disadvantage during drug withdrawal, validating intermittent dosing to forestall resistance. In contrast, resistance driven by KRASG13D amplification is not reversible; rather ERK1/2 hyperactivation drives ZEB1-dependent epithelial-to-mesenchymal transition and chemoresistance, arguing strongly against the use of drug holidays in cases of KRASG13D amplification.

Community assessment to advance computational prediction of cancer drug combinations in a pharmacogenomic screen.

The effectiveness of most cancer targeted therapies is short-lived. Tumors often develop resistance that might be overcome with drug combinations. However, the number of possible combinations is vast, necessitating data-driven approaches to find optimal patient-specific treatments. Here we report AstraZeneca's large drug combination dataset, consisting of 11,576 experiments from 910 combinations across 85 molecularly characterized cancer cell lines, and results of a DREAM Challenge to evaluate computational strategies for predicting synergistic drug pairs and biomarkers. 160 teams participated to provide a comprehensive methodological development and benchmarking. Winning methods incorporate prior knowledge of drug-target interactions. Synergy is predicted with an accuracy matching biological replicates for >60% of combinations. However, 20% of drug combinations are poorly predicted by all methods. Genomic rationale for synergy predictions are identified, including ADAM17 inhibitor antagonism when combined with PIK3CB/D inhibition contrasting to synergy when combined with other PI3K-pathway inhibitors in PIK3CA mutant cells.