LIDR-TB: Large Language Model-Integrated Platform for Traceable Drug Repurposing in Tuberculosis

Tuberculosis (TB) remains a global challenge exacerbated by the ongoing emergence of drug-resistant strains. Although drug repurposing offers a cost-effective strategy to expedite therapeutic discovery, its efficacy is impeded by the lack of unified, disease-specific frameworks that integrate fragmented chemical and biological data from heterogeneous sources. Herein, we developed LIDR-TB (publicly available at http://lidrtb.sysbio.org.cn), a TB-focused, LLM-integrated informatics platform for traceable drug repurposing. The platform integrates three core modules: a curated knowledge base that consolidates heterogeneous chemical and biological evidence into a structured graph topology; interactive network visualization for exploring multi-layered molecular and biological relationships; and a RAG-based question-answering model that anchors LLM reasoning within the curated knowledge base. Their coordinated integration enables multi-level verification, allowing generative responses to be traced directly back to raw experimental evidence. Benchmarking on expert-curated suites demonstrated high performance in semantic understanding and answer fidelity. Notably, we observed a "semantic compensatory effect", wherein the LLM leverages latent semantic context to partially offset structural parsing misalignments, maintaining reasoning reliability in complex biomedical scenarios.