An interpretable framework for cancer drug response prediction using integrated drug and multi-omics data with a...

Previous studies have shown that efficient and effective drugs and cell line representation can improve drug response prediction (DRP) in cancer therapy. Furthermore, the performance of DRP can be enhanced by integrating multi-omics data of cell lines with drugs. However, existing DRP models have shown drawbacks in extricating rich and effective features from the cell lines and representing the molecular structure of drugs available in the SMILES format. This article proposes an integrated drug and cell line representation-based hybrid deep learning architecture, HyDRP-LG (Hybrid Drug Response Predictor with Bi-LSTM-GRU), for enhancing drug response prediction. The model combines Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to effectively capture complex interactions across integrated drug and multi-omics features in pan-cancer settings for personalized cancer treatment. Firstly, the drug molecular structures (SMILES string) are encoded to distinct embeddings to encapsulate molecular substructure data instead of using complex representations. Secondly, the high-dimensional transcriptomic data in the CCLE dataset is reduced in dimension using PCA from 17,736 to 500. The mutation (Genomic) and methylation (Epigenomic) data from the cell line are represented in their original binary forms and adjusted to the maximum length (735 and 400, respectively) of a sample required in the respective cell line. Lastly, the drugs and the cell line representations are integrated and put through a customized hybrid deep Bi-LSTM-GRU network for predicting the IC50 value. Experimental analysis using single and multi-omics features shows that the proposed HyDRP-LG model achieves improved predictive performance compared to reported DRP methods. SHAP-based feature interpretation identified eight key genes significantly associated with drug response, and pathway analysis revealed biological pathways relevant to cancer progression, providing interpretability beyond predictive accuracy.