Predicting tumor drug sensitivity with multi-omics data

Abstract

The prediction of tumor drug sensitivity plays an important role in clinically guiding patients' medication. In this paper, a multi-omics data-based cancer drug sensitivity prediction model was constructed by Stacking ensemble learning method. The data including gene expression, mutation, copy number variation and drug sensitivity value (IC50) of 198 drugs were downloaded from the GDSC database. Multiple feature selection methods were applied for dimensionality reduction. Six primary learners and one secondary learner were integrated into modeling by Stacking method. The model was validated with 5-fold cross-validation. In the prediction results, 36.4% of drug models' AUCs were greater than 0.9, 49.0% of drug models' AUCs were between 0.8-0.9, and the lowest drug model's AUC was 0.682. The multi-omics model for drug sensitivity prediction based on Stacking method is better than the known single-omics or multi-omics model in terms of accuracy and stability. The model based on multi-omics data is better than the single-omics data in predicting drug sensitivity. Function annotation and enrichment analysis of feature genes revealed the potential resistance mechanism of tumors to sorafenib, providing the model interpretability from a biological perspective, and demonstrated the model's potential applicability in clinical medication guidance.