Neuro-Symbolic AI Grounds AMR Prediction in Biology

This research introduces KG-TRACE, a new neuro-symbolic framework designed to enhance antimicrobial resistance (AMR) prediction by providing mechanistic biological grounding. While existing whole-genome sequencing (WGS)-based AMR models achieve high accuracy, they often lack transparency regarding the biological pathways underpinning their predictions. KG-TRACE addresses this by integrating the World Health Organization (WHO) mutation knowledge graph (KG) as a structured biological constraint directly into a neural genomic model. The framework fuses genomic features with RotatE-based KG embeddings through a learned epistemic trust gate, which dynamically balances neural evidence with established symbolic biological knowledge. Evaluated on the CRyPTIC M. tuberculosis cohort, KG-TRACE achieved an AUROC of 0.9760 for isoniazid, demonstrating competitive accuracy. Its primary value, however, lies in its symbolic grounding, quantified by the new Biological Grounding Ratio (BGR) metric. KG-TRACE achieved 92.5% symbolic coverage for isoniazid-resistant predictions and can identify multi-drug resistance co-occurrence artifacts, flagging 'UNCERTAIN' cases for laboratory follow-up, thereby bridging the gap between predictive accuracy and clinical trust.