Metapath2vec and attention-driven heterogeneous graph learning for prioritizing TCM-derived small molecules in gastric cancer

Traditional Chinese Medicine (TCM) represents a multi-component therapeutic system with substantial chemical complexity. This complexity makes it difficult to directly elucidate anti-gastric cancer mechanisms from macroscopic herbal formulae. To support the systematic prioritization of potential small-molecule candidates, this study focuses on screening bioactive small-molecule constituents from TCM. We constructed a heterogeneous network integrating Chinese herbal pieces (CHPs), Chinese patent medicines (CPMs), genes, diseases, and small molecules, and incorporated metapath2vec representations and attention mechanisms into a graph neural network framework, the model is designed to learn relational patterns among heterogeneous nodes. Network pharmacology and in vitro validation in AGS and MKN-45 gastric cancer cell lines show that Icaritin and Arundine inhibit cell proliferation and induce apoptosis. This study presents an AI-assisted pipeline for candidate prioritization, providing methodological support for the systematic prioritization and preliminary validation of bioactive TCM molecules in gastric cancer.