Large language model inference of macromolecular complex composition via model consensus and experimental data integration
Large language models (LLMs) are poised to reshape how biologists retrieve specialized knowledge at scale. Yet their performance on deep, domain-specific queries is poorly defined because much biological information resides in structured databases or large experimental datasets rather than in a free text format. One such gap in cellular biology lies in identifying major macromolecular complexes, conserved biological units essential to many cellular processes. Cataloging large complexes, such as the ribosome or RNA polymerase, along with their constituent genes, presents a significant challenge for LLMs because of their tendency to hallucinate and to produce incomplete or inconsistent lists of components. Here, we systematically evaluate six state-of-the-art LLMs on the task of retrieving the gene components of 91 protein complexes and develop an integrative framework that combines LLM output consensus with experimental multi-omics data to reconcile and filter model responses. We found that two extensions of a basic single-LLM baseline, (i) aggregating LLM outputs into a consensus and (ii) integrating LLM predictions with the experimental data, each improved retrieval accuracy. Furthermore, a consensus of LLM outputs integrated with the incomplete experimental data using a graph-theoretic approach achieved the highest accuracy (F1 score of 82.5%), compared to the best stand-alone singe LLM (F1 score of 76.4%). These results show that optimized integration of predictions from multiple LLMs and high-throughput experimental data can support scalable, semi-automated curation of specialized biological resources, providing a general template for benchmarking and deploying LLMs for the structured knowledge retrieval tasks in molecular biology.