LLM 기반 생체 모델 검증 프레임워크 BELL 공개
생체 모델 구축 시 수동 검증의 비효율성을 극복하기 위해 LLM 기반 자동화 프레임워크 BELL이 개발되었습니다. BELL은 증거 검색, 점수화, 설명 생성을 자동화하여 49.5% 상위 신뢰도, 72.4% 긍정적 추천 결과를 보였습니다.
Building accurate and predictive mechanistic models requires careful biological interaction curation and verification against existing knowledge. When done manually, these tasks become impractical, especially with massive extraction of interactions facilitated by advanced natural language processing methods and large language models (LLMs). We present BELL (Biomodel Evidence and LLM-based Logic), a biocuration support framework that automates evidence retrieval, scoring, and explanation for interaction-level verification. BELL processes each interaction through a five-step pipeline: entity grounding, database ranking, evidence retrieval from seven biological databases, a heuristic four-dimension programmatic scoring, and chain-of-thought explanation with a recommended curator action generated by LLMs. We applied BELL on 210 protein-protein interactions from a curated Glioblastoma Multiforme (GBM) model. Results show that 49.5% of interactions achieved HIGH confidence and 72.4% received a positive curator recommendation, while qualitative flags precisely directed curator attention to evidence gaps. BELL is integrated into the KALIMBA curation platform and available at www.boheme.pit