A Systematic Machine Learning Framework for Evaluating and Ranking Omics Layers in Cancer Drug Response Prediction

This study introduces a top-down framework for evaluating the utility of multi-omics features to predict the response of 309 drugs in cancer cell lines. This was done by taking a multi-omics approach where data from proteomic, transcriptomic, genomic, metabolomic, and miRNA were integrated with drug sensitivity (area under the curve, AUC) data. We performed modular dimensionality reduction using t-SNE (t-distributed Stochastic Neighbor Embedding), followed by K-Means clustering to stratify cell lines into data-driven molecular subgroups, and applied a Random Forest model to refine the drug list, selecting only those with a prediction accuracy exceeding 75%. Our findings show that among the evaluated single-omics features, transcriptomics is the most informative; however, multi-omics integration significantly enhances predictive capability compared to single-omics analysis, with a combination of transcriptomic, proteomic, and miRNA data achieving the best predictive performance across both primary and validation datasets. Cluster analysis showed the importance of well-defined clusters, indicating that while silhouette scores were linked to prediction success, biological variability also played a critical role. This study advances personalized oncology treatment strategies and provides a foundation for future studies focused on ranking omics features based on their predictive capabilities, eventually contributing to better therapeutic outcomes. Predictive performance is used here to evaluate omics feature strength, rather than as an objective to optimize predictive models.