AI-driven discovery of probiotic gene modules and their functional impact in food and health.

Introduction: How a probiotic works depends on its genes, its proteins, and the environment it lives in. Current knowledge gaps make it hard to predict how probiotics act in different foods or to find new, useful traits. To address this, we developed a new AI tool that analyzes probiotics from several angles-genes, proteins, and metabolism-to predict how they actually behave in different types of food. Methods: Using 156 fully annotated genomes, we identified 247 gene modules and uncovered latent functional clusters. Protein language models (ESM-2 and ProtT5) enabled the structural and biochemical characterization of 3,110 proteins, including lineage-restricted and previously unannotated sequences. Additionally, we simulated bacterial metabolic activity across different food matrices. Results: The framework successfully identified specific traits associated with stress tolerance, adhesion, short-chain fatty acid (SCFA) biosynthesis, and immunomodulation. Metabolic simulations demonstrated precisely where strains thrive and where metabolic bottlenecks occur, while models also revealed how these genetic traits might help boost the immune system or protect the gut lining. Notably, the AI's predictions aligned closely with biological reality, confirming that the model captures meaningful insights rather than just statistical patterns. Discussion: This multi-scale framework offers a mechanistic basis for interpreting probiotic functional potential. By bridging genomic, proteomic, and environmental data, this approach supports the rational development and optimization of probiotic strains for specific food matrices.