Abstract
The prediction of cancer drug responses (CDRs) holds considerable relevance for the guidance of patient-specific clinical treatments. However, the inherent cellular heterogeneity among cancers introduces substantial challenges to such predictions. This paper presents a novel molecular graph convolutional model, named DVFGCDR, developed for predicting CDRs. This innovative model amalgamates both 2D chemical and 3D geometric drug properties to derive a more representative drug embedding. It also includes the capacity to integrate gene expression data from single-cell and bulk RNA sequencing data associated with hundreds of cancer cell lines, thereby enhancing the accuracy of drug response predictions. Experimental comparisons highlight the superior performance of the DVFGCDR model in CDR classification and regression tasks, achieving a Pearson score of 0.959, surpassing current state-of-the-art models. This superiority underscores the model's powerful predictive capability.