Inference of sample-specific genetic interactions to increase accuracy of indication prioritization in oncology clinical trials and facilitate exploration of combined therapy opportunities.

Abstract

3123 Background: Precision oncology is growing rapidly in parallel with advances in high throughput sequencing. Development of new anti-cancer therapies is, however, still associated with low efficacy issues, leading to phase II and III clinical trial failures. Improved methodologies are required to identify clinical and molecular patient profiles associated with good drug response to inform decisions on indication prioritization. Methods: We used a sample-specific Genetic Interaction Graph Inference (ssGI2) algorithm, integrating bulk tumor transcriptomic data as well as data collected from 120 public databases and scientific literature in oncology, to infer genetic interactions (GI). More than 10,000 genes from 17,000 samples, covering 195 oncology ICD10 codes, were used to infer GIs for each individual sample. GIs involving a given drug target are selected from a compendium of 17,000 networks of 2M GIs each, and ranked based on their prevalence in the patient cohort and data-support. The mean Z-scored expression of genes from the top ranked GIs were subsequently used to predict drug response for each patient and to calculate the response rate for each indication. Detailed information on each drug target’s genetic interactors was used to characterize the drug’s mechanisms of action and explore opportunities for combined therapies. We investigated our method's ability to predict good responders using four FDA approved immune and targeted therapies (pembrolizumab, nivolumab, ipilimumab and sorafenib) across seven clinical studies. Importantly this methodology is suitable for drugs with no clinical studies available. Results: Our results show that the prediction of good responders can be achieved with Precision-Recall AUC on average 13% higher than predictions based on drug target expression level solely, in five out of seven studies. Also, for each drug target, between 30 to 140 genetic interactors with good performance (Precision=0.92; Recall=0.61) were identified, suggesting potential synergistic effects of drugs, some of which have already been confirmed by clinical studies on combined therapies. Conclusions: Our ssGI2-derived signatures are powerful predictors of good response to a drug even without available clinical data. Applying this methodology at a pre-clinical stage will significantly de-risk clinical trials, particularly for novel therapies, and could also support investigation of new combined therapies.