MTB-ImmunogenKG: An LLM-assisted knowledge graph for antigen selection in tuberculosis vaccine research

Tuberculosis (TB) vaccine design relies on the selection of optimal antigens. However, evidence on candidate antigens is scattered across the literature and rarely linked to critical vaccine parameters, such as immune responses, antigen combinations, and adjuvants. Here, we introduce Mycobacterium tuberculosis (MTB)-ImmunogenKG, a provenance-linked, antigen-centric knowledge graph for MTB, that is constructed via an optimized information-extraction pipeline and integrated with a large language model (LLM) system to enable knowledge-augmented reasoning. From over 77,000 publications indexed in PubMed (as of July 2024), this graph consolidates 1.48 million sentence-level statements spanning 14 entity types. In practical applications, MTB-ImmunogenKG enables contradiction-aware antigen profiling for 3,154 proteins (representing about 77% of the total annotated proteins) and improves protective-efficacy prediction with a Matthews correlation coefficient (MCC) gain of 0.19 over sequence-based tools and 0.45 over an LLM-only baseline. By delivering a traceable synthesis of fragmented findings, MTB-ImmunogenKG facilitates the construction of antigen panels and the pairing of adjuvants, thereby streamlining the experimental design cycle for next-generation TB vaccines.