The rapid global spread of antimicrobial resistance (AMR) has placed unprecedented pressure on clinical decision-making. Machine learning predictors of antibiotic susceptibility exist, but their lack of mechanistic grounding limits credibility. We present BO_SCPLOWACTEC_SCPLOWRO_SCPLOWEASONC_SCPLOW, a reasoning large language model (LLM) that predicts bacterial susceptibility to a target antibiotic, together with a mechanistic rationale. BacteReason is obtained by fine-tuning an open-weight LLM on clinical susceptibility data augmented with rationales that explain the molecular mechanisms. These rationales are produced by a proprietary teacher LLM prompted to explain known susceptibility outcomes. The teacher is interfaced via TogoMCP with a collection of biomedical knowledge-graph databases, grounding each reasoning step in retrieved evidence. On an extrapolation benchmark, BO_SCPLOWACTEC_SCPLOWRO_SCPLOWEASONC_SCPLOW achieves a relative improvement of 43% over the untuned baseline and 38% over the same base LLM fine-tuned without rationales, demonstrating that reasoning supervision improves prediction accuracy.
Oikawa, Y. et al. · CC-BY 4.0