Use of a systems-biology informed machine learning model to predict drug response using clinically available NGS data.

Abstract

3136 Background: Tremendous progress towards understanding the molecular underpinnings of cancer has ushered in an era of precision medicine, where actionable biomarkers inform rational design of novel drugs, repurposing of existing medications or new combination therapies, agnostic of tumor type. However less than 20% of cancer patients match to therapies using existing predictive biomarkers. There is therefore an acute need for new methods to expand the population of patients who could benefit from existing drugs and improve patient selection and design of clinical trials evaluating new or repurposed drugs. Methods: We developed systems biology-informed AI methods trained on vast multi-omics and multi-modal datasets from DepMap, TCGA and other publicly available sources, to uncover latent vulnerabilities existing in tumors. These vulnerabilities map to pharmacological sensitivities and can be assigned to individual patients using clinically available molecular data. To evaluate our model, we focused on predicting sensitivities to a third-generation tyrosine kinase inhibitor (TKI), which we assessed in non-small cell lung cancer (NSCLC) cell lines, patient derived xenografts (PDXs) and a clinicogenomics cohort from AACR Project GENIE (n=2004 patients, n=273 patients treated with TKI of interest). Where applicable, whole exome data used as input for validation model predictions were filtered to ~300 genes included on commercial NGS panels to reflect data available in RWD cohorts. Results: Our model accurately predicts response to the TKI across diverse cancer cell lines with high accuracy (r2= 0.768; rho= 0.83; mean absolute error = 0.065). Using NSCLC cancer cell lines and PDX model systems heterogeneous for this TKI’s current FDA molecular label (but within the scope of the drug’s current cancer indication), we were able to further validate our model’s predictions, show they were not dependent on obvious molecular confounders, and demonstrate a potential label expansion opportunity for this drug beyond the current narrow molecular label. While all NSCLC patients from AACR Project GENIE treated with this TKI were “on label”, NSCLC patients our model predicted to be sensitive to the TKI had a 4-month increased median PFS compared to patients predicted to be less sensitive (Log-rank test p-value< 0.05; HR= 1.64; 1.23-2.22, 95% CI). The PFS partition was not due to obvious clinical confounders such as stage (p= 0.91), age (p=0.86), sex (p=0.63), smoking status (p=0.57) or on-target mutations (p=0.70). Biological networks associated with each prediction provide rich context and highlight a putative rational drug combination strategy for this drug. Conclusions: We have developed an ML model that uses clinically obtainable NGS data to accurately predict increased sensitivity to a third-generation TKI validated in NSCLC cell lines, PDX models and a RWD patient cohort.