Computational Techniques for Predicting Response to Therapy for Cancer

Personalizing cancer therapies is a key goal in oncology research. This study proposes a novel approach to tackling the problem of predicting drug response in the form of IC50 values from gene expression data of cancer cell lines. Working on publicly available data, we applied various dimensionality reduction techniques are used to address the problem of high dimensionality, including biological data reduction methods using datasets such as LINCS L1000 and COSMIC CGC and statistical data reduction methods such as LIME and SHAP. We ran multiple versions of the reduced data on machine learning models (Elastic Net and XGBoost). Finally, we evaluated the utility of our models based on clinical research in the field of oncology by performing post-hoc analysis on key genes indicated by statistical modelling. Our results show that machine learning models can accurately predict drug response and provide insights to the underlying biological mechanisms driving drug response even after a reduction of 99% of genes. This study has important implications for personalized cancer therapy. It provides a framework for predicting drug response based on gene expression data, even when working on a reduced gene set.