AnnoNet-TCMs: An annotation network-driven discovery of novel active compounds and mechanisms in traditional Chinese...

Understanding the mechanistic basis of Traditional Chinese Medicines (TCMs) remains challenging due to its multi-component, multi-target nature and the lack of systematic interpretability in current analytical frameworks. Here, we propose AnnoNet-TCMs, an annotation network-driven discovery of novel active components and mechanisms in TCMs, which is a multi-dimensional heterogeneous network model that integrates herbs, compounds, protein targets, Gene Ontology (GO) terms, and diseases to decode pharmacological mechanisms underlying TCMs treatments.

This method enables the prediction of positive compound-disease and herb-disease treatments and further priorities potential treatment mechanisms and active components of herbs.

Upon the network structure of AnnoNet-TCMs, a random walk with restart (RWR) algorithm was conducted to obtain effect propagation profiles (EPPs). The distances of compound-disease and herb-disease were calculated for activity prediction. Mechanistic insights were acquired by interpreting the key molecular targets and biological processes effects highlighted by the profiles.

AnnoNet-TCMs achieved mean area under curve (mAUC) of 0.8235 and 0.6644, mean average precision (mAP) of 0.0084 and 0.0855, mean recall at 20% (mR@20%) of 0.6529 and 0.5706 for compound-disease and herb-disease association prediction, respectively. As a case, the model successfully identified effective anti-cancer compounds such as obacunone for colon carcinoma, predicted the anxiolytic herb (Apocynum venetum L.), and uncovered a previously unreported synergistic compound pair including tanshinone I and astragaloside IV for myocardial infarction. Subgraph-based analysis and pathway enrichment confirmed the involvement of apoptosis, NF-κB signaling cascades in these interventions.

The AnnoNet-TCMs represents a generalizable framework for interpretable drug discovery from traditional medicine systems and offers a foundation for future integration with AI-driven target prediction.