Explainable Protein Language Model Embeddings for Rapid Antibiotic Resistance Gene Classification

Abstract Antimicrobial resistance (AMR) is a major public-health threat because resistant infections narrow therapeutic options, delay effective treatment, and increase the clinical and economic burden of bacterial disease. Conventional phenotypic susceptibility testing remains essential but requires 24–48 hours after isolate growth, while sequence-based interpretation requires curated databases and bioinformatics expertise. This paper presents ResiScan, an explainable machine-learning framework for rapid classification of antibiotic resistance genes from protein sequences. The proposed system uses the Comprehensive Antibiotic Resistance Database (CARD) protein homolog model data and extracts contextual protein representations with the pre-trained ESM2 model. Each protein sequence is tokenized, truncated to 512 residues for computational consistency, passed through ESM2, and represented by a 320-dimensional mean-pooled embedding. A logistic regression classifier is trained on these embeddings, with the Synthetic Minority Oversampling Technique (SMOTE) applied only to the training partition to address severe class imbalance across five target resistance classes: beta-lactam, aminoglycoside, macrolide, tetracycline, and vancomycin. The framework further integrates SHAP-based feature attribution to expose which embedding dimensions contribute most strongly to each prediction, and provides an interactive Gradio web interface for non-specialist use. The prototype achieves 99.25% accuracy and a macro-F1 score of 0.9507 on a held-out test set of 400 protein sequences, demonstrating a practical balance among speed, interpretability, and deployability.