Genetic Features for Drug Responses in Cancer -- Investigating an Ensemble-Feature-Selection Approach

Predicting drug responses using genetic and transcriptomic features is crucial for enhancing personalized medicine. In this study, we implemented an ensemble of machine learning algorithms to analyze the correlation between genetic and transcriptomic features of cancer cell lines and IC50 values, a reliable metric for drug efficacy. Our analysis involved a reduction of the feature set from an original pool of 38,977 features, demonstrating a strong linear relationship between genetic features and drug responses across various algorithms, including SVR, Linear Regression, and Ridge Regression. Notably, copy number variations (CNVs) emerged as more predictive than mutations, suggesting a significant reevaluation of biomarkers for drug response prediction. Through rigorous statistical methods, we identified a highly reduced set of 421 critical features. This set offers a novel perspective that contrasts with traditional cancer driver genes, underscoring the potential for these biomarkers in designing targeted therapies. Furthermore, our findings advocate for IC50 values as a predictable measurement of drug responses and underscore the need for more data that can represent the dimensionality of genomic data in drug response prediction. Future work will aim to expand the dataset and refine feature selection to enhance the generalizability of the predictive model in clinical settings.