Multimodal interpretable deep learning for transcriptome-informed precision oncology and drug mechanism analysis

Precision oncology faces critical challenges in interpreting complex cellular signals and predicting drug responses across heterogeneous cancer environments. Here, we present BioGDR, a multimodal interpretable deep learning framework that integrates structure-based predicted biological features, including differential gene expression and kinase inhibition profiles, eliminating the need for experimental measurements. By modeling tumor transcriptomic states through pathway-informed graph neural networks and employing a drug-guided attention strategy, BioGDR enables mechanistic insights into drug sensitivity across compound and cellular contexts. Comprehensive evaluations demonstrate that BioGDR outperforms existing methods in compound screening relevant to early-stage drug discovery and in predicting cell line sensitivity across heterogeneous cellular states characteristic of precision oncology, while analyses on clinical patient cohorts further confirm its practical utility and generalization capability. Experimental validation with a novel ALDH1B1 inhibitor confirms its ability to identify sensitive cell populations and reveal underlying mechanisms. This work establishes a robust, biologically informed framework that bridges preclinical drug development and clinical applications, advancing precision oncology through integrative, multimodal learning and interpretable mechanism analysis.