Who hasn’t dreamed of predicting single bacterium or bacterial consortium metabolic capabilities directly from genomes? While genome annotation tools exist, genome-based inference of metabolic functions remains incomplete, especially in food matrices. Traditional phenotyping is costly, highlighting the need for predictive, data-driven approaches. Our project aims to bridge genomics, AI, and open science to improve microbial phenotype prediction in fermented foods and stimulate community driven efforts. We will curate a benchmark dataset of 300 LAB/PAB strains, integrating public (BacDive, metaTraits, DSMZ) and internal (CIRM, STLO) data, enriched with kinetic phenotypes (e.g., sugar consumption rates). Unlike existing datasets, ours will focus on food-relevant traits, ensuring broad species coverage and phenotypic diversity. We propose two complementary modeling approaches: (1) interpretable pan-genome ML using biologically grounded features (genes, orthologs) and (2) AI-based models leveraging foundation DNA model (Evo2) for high-accuracy predictions.
The project will deliver a FAIR-compliant public database, pre-trained models, and R/Python packages that will serve as the basis to host a data challenge, in partnership with DataIA, to evaluate model generalization, including cross-kingdom predictions (e.g., yeast). This will facilitate rational strain selection and advance genome-to-phenotype prediction.