Machine learning (ML) is now widely applied in microbiology, but its reliability varies markedly across domains. In this review, we analysed data from 254 scientific articles that evaluates ML through three linked dimensions including data readiness, model suitability, and deployment readiness across diagnostics and pathogen identification, virology, microbiome research, industrial and environmental microbial biotechnology. This framework helps distinguish robust progress from performance inflated by methodological limitations. Our review shows that pathogen identification and antimicrobial resistance prediction consistently achieve strong performance when supported by curated datasets, reliable labels, and comprehensive reference databases. However, their practical value remains limited by internal validation, lineage confounding, and uneven transfer across strains, institutions, and regions. In virological studies, predictive stability is further challenged by incomplete reference databases, changing taxonomy, and temporal drift during outbreaks. In microbiome research, ML classifiers can detect disease and environmental signals, but their generalization across cohorts remains weak because of compositional data structure, technical bias, and incomplete metadata. Industrial bioprocessing and environmental applications show promise when process data are rich and controlled, but deployment beyond laboratory or site-specific settings remains limited. Across structured microbiological datasets, classical supervised models often remain competitive with deep learning while being easier to interpret and validate. Detailed quantitative benchmarks supporting these comparisons are synthesized in the main text and summary tables. Overall, progress will depend less on algorithmic novelty than on interoperable and well-annotated datasets, representative sampling, standardized benchmarking, reproducible workflows, and prospective multi-site validation.
Barcan et al. (Mon,) studied this question.