Machine Learning-Based Drug Discovery: Predicting Drug-Target Interactions for Accelerated Pharmaceutical Research

Abstract

Drug discovery remains one of the most resource-intensive stages in pharmaceutical research, often requiring years of experimental work and significant financial investment. Identifying effective drug–target interactions is a major bottleneck, traditionally addressed through high-throughput screening and biochemical assays. However, the availability of large-scale pharmacogenomic data presents new opportunities to accelerate this process using computational techniques. This study proposes a machine learning-based framework to predict drug–target interactions by leveraging the Genomics of Drug Sensitivity in Cancer (GDSC) dataset. The dataset includes drug response values (IC50) and genomic features such as gene expression, mutation status, and copy number variations across various cancer cell lines. An extensive exploratory data analysis was conducted to assess data distribution, identify feature correlations, and resolve issues such as missing values and skewness. After appropriate preprocessing, predictive models were trained to learn the complex relationships between genomic features and drug sensitivity. The best-performing model achieved a strong predictive performance, with a coefficient of determination (R²) of 0.715 and a low root mean squared error, indicating robust generalization. Model interpretability was enhanced using SHAP (Shapley Additive Explanations), which identified biologically relevant genes such as TP53, EGFR, and BRAF as significant contributors to drug response variation. This research highlights the potential of computational approaches to complement traditional drug discovery methods. By enabling accurate and interpretable predictions of drug efficacy, the proposed framework supports advancements in AI-assisted pharmacology, with promising implications for precision medicine, drug repurposing, and targeted therapy development.