The research investigated the ability of large language models (LLMs) to generate biomedical hypotheses, specifically examining whether they were genuinely reasoning from scientific evidence or simply producing plausible ideas. The study combined the Kyoto Encyclopedia of Genes and Genomes (KEGG), Rhea, and UniProt databases into a unified biochemical knowledge graph. This created a benchmark consisting of 550 paths connecting enzyme sources to rare disease endpoints, resulting in 13,200 hypotheses generated by six different LLMs. The experiments varied the biological information provided to each model, testing conditions including source enzyme only, full biological path, and source and disease endpoint only.
Hypotheses were evaluated using an expert-derived five-criterion rubric, assigning a score from 1 to 5 for each criterion. The results indicated that models given both the source and disease endpoint frequently produced the highest-scoring hypotheses. However, these hypotheses demonstrated less grounding in the underlying evidence. Conversely, models supplied with the full biological path generated hypotheses that more closely aligned with established mechanistic relationships, a process termed ‘evidence-disciplined reasoning’.
To confirm this effect, the researchers conducted a controlled experiment involving the shuffling of intermediate path steps while maintaining fixed endpoints. This manipulation resulted in a significant drop in evidence grounding (delta = -0.793, p < 0.001), confirming that the models genuinely utilized the path structure during their reasoning process. This suggests that the LLMs were not simply generating hypotheses based on superficial correlations but were leveraging the structured knowledge within the biochemical pathways.
The findings demonstrate that knowledge graphs support hypothesis generation in two distinct ways: they can identify biological endpoint pairs not previously documented in the literature, and the mechanistic paths within the graph guide the LLMs’ reasoning between these endpoints. This approach offers a method for evaluating and potentially improving the reliability of LLM-generated hypotheses in biomedical research.
Source: https://arxiv.org/abs/2609.12260
