iCAM-Net: Interpretable herb-disease association prediction via cross-channel attention and molecular interaction...

Target identification is fundamental to drug discovery, facilitating mechanistic elucidation and therapeutic target identification. Traditional herbal medicines exert therapeutic effects through multi-component, multi-target synergistic mechanisms that systematically modulate complex diseases. However, existing herb-disease association (HDA) prediction methods operate at macro-level associations without revealing underlying molecular mechanisms.

This study aims to develop iCAM-Net, an interpretable HDA prediction model that integrates molecular-level interaction data to elucidate and accurately predict herbal medicine mechanisms of action.

iCAM-Net employs a dual-channel architecture featuring separate hypergraph structures for herb-component and disease-protein relationships. A cross-channel attention mechanism dynamically captures component-protein interaction patterns, while component-protein association (CPA) prediction serves as an auxiliary task within a multi-task learning framework.

iCAM-Net achieved F1-scores of 0.9611 and 0.9310 with AUROCs of 0.9937 and 0.9779 on the TCM-suite and HERB datasets, respectively, significantly outperforming baseline methods. Ablation studies confirmed the critical contribution of each architectural component. Case study analysis validated 9 of 10 predicted disease associations for Angelica sinensis through literature evidence, while molecular docking confirmed predicted component-protein interactions.

iCAM-Net advances traditional medicine modernization by achieving interpretable HDA prediction through cross-channel attention, demonstrating superior predictive accuracy while elucidating molecular mechanisms underlying therapeutic effects. The source codes of iCAM-Net are available at https://github.com/qunshanxingyun/iCAM-Net.