The Development And Application Of Machine Learning For Drug Discovery And Drug Response Prediction For Personalized Cancer Treatment

Abstract

In the field of pharmacogenomics and precision medicine, gene expression analysis has become a crucial tool in predicting patient drug response. My contributions to this field come primarily in the development and application of two bioinformatic packages: oncoPredict and scIDUC. oncoPredict is a tool based in the R programming language, primarily used to predict the response of various cancer samples (cell line, patient, etc.) to different drugs. This is made possible by incorporating machine learning to analyze the complex relationships between genomic features and drug response from pan-cancer cell lines. These relationships are learned from microarray or bulk RNA sequencing (RNA-seq) data and high-throughput drug screens, then applied to patient data to generate novel drug discovery hypotheses. In turn, oncoPredict aids in identifying potential drug candidates, understanding mechanisms of drug resistance, and predicting the effectiveness of drugs on specific cancer types. scIDUC (single-cell Integration and Drug Utility Computation) is a computational framework based in python that quickly and accurately generates predictions of drug response for cells derived from single-cell RNA sequencing (scRNA-seq) data. It is a transfer learning-based approach that learns relationships between drug sensitivities and relevant gene expression patterns based on cell line bulk RNA-seq data and high-throughput drug screens, similar to oncoPredict. The key difference, however, is that prior to training drug response models, scIDUC integrates bulk RNA-seq and target scRNA-seq data to denoise and extract shared gene expression patterns between bulk and single-cell data sources. The resulting bulk data is then used to train drug response models, whose coefficients are further applied to post-integration single-cell data to infer cellular drug sensitivity scores.