A hierarchical attention network integrating multi-scale relationship for drug response prediction

Abstract

Anticancer drug response prediction with deep learning technology has become the foundation of precision medicine. It is essential for anticancer drug response prediction to incorporate multi-scale relationships within feature items and biomedical entities. Therefore, we propose MultiDRP that develops the hierarchical attention networks integrating multi-scale relationship for drug response prediction. MultiDRP can fuse both internal correlation of feature items and external relationship of biomedical entities by hierarchically integrating graph attention and self-attention networks to improve the anticancer drug response prediction. A variety of results showed that MultiDRP generated the great representation by integrating multi-scale relationships, and achieved higher performance compared to existing methods on various prediction scenarios. The results of network proximity, gene ontology biological process (GOBP) enrichment, and drug pathway association analysis show that MultiDRP can accurately screen the sensitive and resistant drugs for cancer cell lines. In vitro experiments, eight novel drugs predicted by MultiDRP exhibited high sensitivity to lung cancer cell line NCI-H23, seven of which showed IC50 values of less than 10 nM. These results further suggest that MultiDRP can serve as a powerful tool for anticancer drug response prediction. The source data and code are available at https://github.com/pengsl-lab/MultiDRP.git.