Construction and validation of THD-GFN: A tumor heterogeneity-adaptive drug–drug interaction prediction model via multi-scale temporal encoding

Accurate tumor-drug interaction prediction is essential for personalized cancer therapy, but existing methods fail to model the dynamic temporal evolution of tumor heterogeneity and context-aware drug interaction dynamics. To tackle these limitations, we propose THD-GFN, a novel framework that integrates hierarchical temporal T-GNN, dynamic graph Transformer, and heterogeneous graph fusion network. THD-GFN captures multi-scale tumor temporal heterogeneity, enables tumor state-aware dynamic drug interaction modeling, and eliminates batch effects across multi-modal features. Tested on CCLE, GDSC, and CLUE datasets, it achieves 0.052 baseline MAE, 0.072 72h temporal MAE, and 78.3% Top-10 synergistic combination precision, surpassing DrugFormer and Taco-DDI in robustness and cross-center generalization. Despite limitations of cell-line data and model complexity, THD-GFN provides an interpretable, accurate solution for tumor-drug interaction prediction and supports translational cancer pharmacotherapy research.