Abstract 395: Prediction of genomic based drug sensitivity signatures to enable optimal drug positioning

Abstract

Abstract Cancer by nature, is a heterogeneous disease, which can lead to highly variable patient responses for targeted therapies - even within the same clinically defined cancer type or subtype. Identifying drug sensitivity biomarkers, patient specific traits that are highly correlated with positive response, has proven to be an effective strategy to identify positioning opportunities and selecting patients. Here, we describe our biomarker identification workflow which identifies drug specific genomic sensitivity signatures, a collection of genomic aberrations that are found to be predictive of patient response, which can be used for patient selection as well as understanding a compound's mechanism of action. Using available drug efficacy data (in vitro or in vivo results) with paired baseline genomics data, we build diverse models to predict sample sensitivity to a drug. For every model, we train and measure the performance and how each genomic aberration contributes to model performance, those that do not sufficiently increase predictability are removed. This process is repeated, with each new model training on a smaller subset of genomic aberrations, until we have the minimum number of genomic aberrations needed to reach peak predictive performance. The resulting set of genomic aberrations make up a genomic sensitivity signature. Instead of using one type of algorithm to build these models, we train three distinct model types: linear, decision tree and custom kernel based. The multiple model architecture allows us to model the unique underlying biological mechanism of a compound's efficacy, such as synthetic lethal relationships or resistant pathway deactivation, ultimately giving us a collection of robust biomarker signatures. In addition to mechanistic understanding, clinical feasibility was a top priority. Therefore, our workflow architecture has been tested and validated to incorporate mutation, copy number alteration and/or gene expression based biomarkers either in combination or alone. This flexibility lowers the data required to use our workflow, making it more widely available to investigational drugs. Our identification of genomic sensitivity signatures can then be used for clinical patient selection or positioning a compound for a clinically defined cancer type, based on the prevalence of the predicted signature. Overall, this workflow is a powerful approach identifying diverse, mechanism based drug sensitivity signatures that can enable identifying the right patient for a therapy of interest. Citation Format: Coryandar Gilvary, Olivier Elemento, Neel Madhukar. Prediction of genomic based drug sensitivity signatures to enable optimal drug positioning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 395.