Computational paradigms for antimicrobial resistance prediction: integrating multi-omics, structural modeling, and foundation artificial intelligence systems

Antimicrobial resistance (AMR) poses an escalating threat to global health, as multidrug-resistant pathogens undermine therapeutic efficacy and surveillance systems. Although whole-genome sequencing and phenotypic drug susceptibility testing have strengthened resistome profiling, translating multi-omics data into reliable, clinically deployable intelligence remains computationally fragmented. Following PRISMA 2020 guidelines, we systematically reviewed 156 records published between 2016 and 2025, of which 93 studies were included in the final synthesis. We organize AMR modeling into three methodological strata: (i) classical and interpretable machine-learning frameworks, (ii) structural and deep genomic architectures, and (iii) transformer-based and applied large language model systems that integrate genomic, clinical, and epidemiological signals. Across studies, we identify four converging integrative directions: embedding-level multimodal fusion, knowledge-graph-guided causal reasoning, evolutionary and temporal forecasting, and agentic artificial intelligence systems enabling autonomous, evidence-grounded workflows. Comparative analysis reveals substantial heterogeneity in dataset scale, frequent reliance on internal validation, limited assessment of cross-site robustness, and vulnerability to distribution shift, particularly for minority resistance phenotypes. We argue that future AMR intelligence must integrate uncertainty-aware modeling, standardized validation protocols, and FAIR-compliant infrastructures to transition from static genomic classification toward interpretable, temporally adaptive, and clinically actionable decision systems within One Health surveillance ecosystems.