Large language model-driven knowledge graph learning for herb-macromolecule interaction prediction and functional...

The rapid accumulation of biological macromolecule data has created an urgent need for computational methods capable of characterizing the functional roles of macromolecular targets in complex therapeutic systems. Traditional Chinese medicine represents a typical scenario, in which multiple herbal components may coordinately regulate diverse protein targets through multi-component and multi-target mechanisms. However, existing methods often fail to jointly integrate traditional medicine knowledge, modern biomedical associations, and the functional context of biological targets, limiting systematic identification and mechanistic interpretation of herb-macromolecular target interactions (HTIs). Here, we propose LLM-HTI, a large language model (LLM)-driven graph learning framework with adaptive parameter updating based on a Traditional Chinese and Western medicine knowledge graph (TWKG), for predicting potential interactions between herbs and biological macromolecular targets. LLM-HTI extracts latent semantic representations of herb-target associations using LLMs and integrates them with knowledge graph topology, enabling joint characterization of herbs, macromolecular targets, and their functional relationships. The framework further introduces a gradient-guided adaptive parameter updating mechanism to dynamically select important parameter subsets, improving training stability and optimization efficiency. In addition, gastrointestinal surgeons independently assessed the top-ranked candidate targets for disease relevance, biological plausibility, clinical actionability, and translational priority to provide a clinician-guided evaluation of their medical relevance. These benchmark and case-study results demonstrate that LLM-HTI supports macromolecular target function analysis and herb-target mechanism discovery, while clinician-guided assessment further highlights medically relevant candidates for translational follow-up. Our data and code are publicly available at https://github.com/sisyyuan/LLM-HTI.