To develop a machine-learning framework integrating TCMBank-derived liver disease seed curation with structured herb-level annotations to prioritize herbs with potential relevance to liver disease.

Liver-focused non-tumor disease seeds were curated from TCMBank to identify 356 annotation-supported positive herbs. Herb-level features were derived from structured TCMBank annotations. Six unweighted machine-learning models were trained using the original 356-positive/8835-background positive-unlabeled dataset. Performance was assessed using the area under the receiver operating characteristic curve and the area under the precision-recall curve (PR-AUC), with PR-AUC interpreted relative to the baseline prevalence of 0.039. Candidate herbs were further refined through consensus prioritization, cross-model concordance, and translational evidence evaluation.

A total of 124 curated liver disease seeds and 9191 TCMBank herb records were retained. The final modeling dataset comprised 356 annotation-defined positive herbs and 8835 unlabeled background herbs, corresponding to a positive prevalence of 0.039. Model performance was evaluated using the 356-positive positive-unlabeled dataset. PR-AUC baselines, candidate rankings, and validation-priority scores were generated within and aligned with the final modeling framework.

This study establishes a reproducible machine-learning framework for prioritizing candidate herbs in liver disease research and provides a data-driven strategy for translating large-scale database resources for Chinese medicine into experimentally-testable hypotheses. The prioritized candidates should be regarded as computational hypotheses requiring staged pharmacological, hepatobiliary, and safety validation rather than as evidence of established clinical efficacy.