level: research
kg-trace is a new method for predicting antimicrobial resistance from genomic data. it uses a neuro-symbolic approach that combines a neural network with a structured knowledge graph of mutations from the world health organization. the neural part learns from raw genomic features, while the knowledge graph provides established biological facts. a learned gate dynamically balances the two sources, giving more weight to symbolic knowledge when the neural evidence is uncertain.
the framework was tested on mycobacterium tuberculosis data from the cryptic project. for the drug isoniazid, it achieved an auroc of 0.9760, which is competitive with existing models. however, the main advantage is not higher accuracy but better biological grounding. the authors introduce the biological grounding ratio, a metric that measures how well the model's attributions match known resistance mechanisms. this helps verify that the model is focusing on relevant mutations rather than spurious patterns.
by grounding predictions in biological knowledge, kg-trace makes model decisions more interpretable and trustworthy. this is important for clinical applications where understanding why a prediction was made is as critical as the prediction itself. the approach could be extended to other pathogens and drugs, potentially aiding in the development of more reliable diagnostic tools.
why it matters: it provides a way to build ai models for drug resistance that are not just accurate but also biologically plausible, increasing trust for clinical use.