Explainable artificial intelligence for genomic prediction of antimicrobial resistance in Escherichia coli: a leakage-controlled, cross-model, and biologically validated pilot framework
Year: 2026
Type: article
Abstract: Antimicrobial resistance threatens modern medicine, and Escherichia coli is a leading resistant pathogen for which genome based prediction can complement slow culture testing. To be clinically useful, a predictor must be accurate, explainable, biologically coherent, and free from the optimistic bias caused by bacterial population structure. This pilot study presents a transparent framework that combines leakage controlled modelling with multi method explainability and biological validation, and is intended as a methodological guide rather than a new laboratory finding. Using laboratory confirmed susceptibility phenotypes for 58,750 to 76,364 E. coli genomes obtained from the BV-BRC database, resistance to five antibiotics was predicted from a resistance gene presence and absence matrix. Eight classifiers were tuned with Optuna under clone aware cross validation, evaluated with the Matthews correlation coefficient and complementary metrics, and interpreted with SHAP and permutation importance. Attribution stability was quantified by the cross model Jaccard overlap, and the most important genes were tested for resistance mechanism and KEGG/BRITE pathway over representation. Tree ensembles performed best, reaching a Matthews correlation coefficient of 0.94 for gentamicin and 0.92 for ampicillin, whereas ciprofloxacin remained the hardest task despite a deceptively high ROC area (area under the receiver operating characteristic curve of 0.89 versus a Matthews correlation coefficient of 0.58). Attributions were stable across architecturally different models and concentrated on the established determinants of each drug, which mechanism and pathway enrichment confirmed. The ciprofloxacin case shows that, when resistance is driven by point mutations, the data representation can matter more than the model, an interpretive lesson central to deploying artificial intelligence for resistance prediction.
Source: Journal of Scientific Reports-A
Author: Mesut Ersin Sönmez
Institution: Karamanoğlu Mehmetbey University
Language: English
Subfield: Molecular Medicine
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7217648967