DRPO: A deep learning technique for drug response prediction in oncology cell lines

Abstract

With the invention of high-throughput screening technologies, innumerable drug sensitivity data for thousands of cancer cell lines and hundreds of compounds have been produced. Computational analysis of these data has opened a new horizon in the development of novel anti-cancer drugs. Previous deep-learning approaches to predict drug sensitivity showed drawbacks due to the casual integration of genomic features of cell lines and compound chemical features. The challenges addressed include the intricate interplay of diverse molecular features, interpretability of complex deep learning models, and the optimization of drug combinations for synergistic effects. Through the utilization of normalized discounted cumulative gain (NDCG) and root mean squared error (RMSE) as evaluation metrics, the models aim to concurrently assess the ranking quality of recommended drugs and the accuracy of predicted drug responses. The integration of the DRPO model into cancer drug response prediction not only tackles these challenges but also holds promise in facilitating more effective, personalized, and targeted cancer therapies. This paper proposes a new deep learning model, DRPO, for efficient integration of genomic and compound features in predicting the half maximal inhibitory concentrations (IC50). First, matrix factorization is used to map the drug and cell line into latent ’pharmacogenomic’ space with cell line-specific predicted drug responses. Using these drug responses, we next obtained the essential drugs using a Normalized Discounted Cumulative Gain (NDCG) score. Finally, the essential drugs and genomic features are integrated to predict drug sensitivity using a deep model. Experimental results with RMSE 0.39 and NDCG 0.98 scores on Genomics of drug sensitivity in cancer (GDSC1) datasets show that our proposed approach has outperformed the previous approaches, including DeepDSC, CaDRRes, and KMBF. These good results show great potential to use our new model to discover novel anti-cancer drugs for precision medicine.

Keywords

• Deep learning
• Cancer cell lines
• Personalized medicine
• Matrix factorization
• Sensitivity score
• Molecular features