Introduction
Predicting drug responses is crucial for cancer therapy and drug discovery. However, due to the vastness of chemical space and the high cost and time-consuming nature of wet-lab experiments, models often exhibit limited predictive capability for unseen combinations of drugs and cell lines. To address these challenges, we propose KMVF, a progressive drug response prediction framework that integrates multimodal domain knowledge. KMVF generates sequential features including drug scaffolds and bioisosteres, graph structures, and multi-state molecular conformations to enrich the model's chemical knowledge. Leveraging these multimodal features, KMVF learns molecular biological relationships based on drug-target knowledge and transfers this insight to the task of drug response prediction. Experimental results demonstrate that KMVF achieves superior performance in both warm-start and cold-start scenarios. KMVF is also applicable to drug-target interaction prediction. A case study on phenotypic screening further validates KMVF's effectiveness in identifying responses to cancer drugs. KMVF provides a valuable tool for drug response prediction and drug discovery, highlighting its potential to advance cancer drug research.
Figure 1. The model architecture of KMVF, which is used for drug response prediction and precision oncology.