Metapath-guided transfer learning with clinical validation for identifying herb-drug interactions

Drug co-administration can alter metabolism and cause clinically important pharmacokinetic interactions. Herb-drug interactions (HDIs) are particularly hard to identify because validated cases are scarce and herbal products are chemically complex.

We aimed to develop Meta-HDI, a metapath-guided transfer-learning framework that utilizes large drug-drug interaction graphs to improve HDI prediction and interpretability, and to prospectively confirm its predictions at the clinical pharmacokinetics level.

This study integrates the development of a computational deep learning framework with a prospective clinical crossover trial involving 18 participants and in vitro human liver microsome assays.

Meta-HDI combines a GCN encoder and shortest-path LSTM with hierarchical attention mechanisms to generate interpretable mechanistic chains. We benchmarked the model against baselines across three in vivo HDI classes and clinical cases. The predicted interaction between donepezil and the herbal formulas Gami-soyosan and Ojeok-san was evaluated in the crossover study, followed by mechanistic validation using microsome assays.

Meta-HDI improved the micro-averaged AUROC to 0.95 (from ≤0.60) and correctly classified all eight clinical cases. Ablation studies highlighted the essential role of protein-protein and drug-protein edges. In the clinical trial, co-administration increased donepezil exposure (1.6-fold C_max,ss; 1.5-fold AUC_tau,ss) without serious adverse events. Attention weights and microsome assays identified falcarinol and glabranin as CYP2D6 inhibitors (IC₅₀ 45 µM and 4.5 µM).

Meta-HDI mitigates HDI data scarcity while providing mechanistic, clinically interpretable predictions. Our framework demonstrated its potential for decision support in herb-drug co-administration, though broader validation across diverse drugs and herbal products is needed.