Species-Agnostic and Salmonella-Specific Models for Antimicrobial Resistance Prediction Using FCGR and ResNet-18

Antimicrobial resistance (AMR) prediction from whole genome sequencing is a major challenge for clinical microbiology and surveillance. We developed deep learning models combining Frequency Chaos Game Representation (FCGR) with a ResNet-18 architecture to classify resistance phenotypes from bacterial genomes. Using a homology-aware clustering strategy to prevent genomic data leakage, we compared a species-specific model trained on Salmonella enterica to a species-agnostic model trained across S. enterica, Escherichia coli, and Staphylococcus aureus. The Salmonella model achieved reliable performance on the independent test set (Balanced Accuracy = 0.82, MCC = 0.66), though results varied by antibiotic: cephalosporins reached near-perfect levels (Balanced Accuracy up to 0.98, MCC up to 0.92), while tetracycline and ampicillin remained more difficult (Balanced Accuracy = 0.7-0.79, MCC = 0.55-0.58). In contrast, the species-agnostic model showed weaker generalization overall, reflecting dataset heterogeneity and missing labels, but provided clear gains for underrepresented species. Notably, tetracycline prediction for E. coli improved compared to the corresponding single-species model. These findings demonstrate that FCGR-based deep learning can produce robust AMR predictors in species-focused settings, and that species-agnostic models, while currently limited, may enhance predictive power for poorly sampled pathogens.