KG-TRACE: Neuro-symbolic framework grounds AMR predictions in biology

Current whole-genome sequencing (WGS) models predict antimicrobial resistance (AMR) with high accuracy but lack mechanistic grounding—they don't align predictions with known biological pathways. KG-TRACE, developed by a team including Naman Garg, Sarika Jain, and colleagues, solves this by fusing a neural genomic model with the WHO mutation knowledge graph. It uses RotatE embeddings to represent symbolic biological knowledge and an epistemic trust gate that dynamically weights neural evidence against this knowledge. The result is a neuro-symbolic framework that doesn't just predict resistance but explains why, grounding each attribution in established biology.

Evaluated on the CRyPTIC M. tuberculosis cohort, KG-TRACE achieved an AUROC of 0.976 for isoniazid, demonstrating competitive accuracy. More importantly, it introduced the Biological Grounding Ratio (BGR), a dataset-level metric that quantifies how well neural attributions align with known biology. The framework achieved a 92.5% symbolic coverage for isoniazid-resistant predictions and could flag uncertain cases for laboratory follow-up, helping identify multi-drug resistance (MDR) artifacts. This bridges the gap between predictive accuracy and clinical trust, offering a verifiable audit trail for clinicians.