Machine Learning Reveals Quantitative Amino Acid Preferences in Bifidobacterium longum Growth

Bifidobacterium longum is a prevalent human gut symbiont whose carbohydrate metabolism is well characterized, whereas the quantitative contribution of amino acids to growth remains unclear. Here, we combined genome-based pathway analysis, growth phenotyping in chemically defined media, and iterative machine-learning-guided medium design to quantify amino acid preferences in B. longum subsp. longum JCM 1217^T. Genome analysis predicted cysteine as the sole auxotrophy, and experiments confirmed that cysteine alone supported growth but did not restore the high maximum cell density and short lag time achieved with a complete amino acid mixture. Regression models and genetic algorithms identified amino acid formulations in the selected optimized compositions that reduced total amino acid input by 66.5% under glucose and 77.2% under lactose while maintaining growth comparable to complete medium. SHAP analysis highlighted tyrosine as the main determinant of maximum cell density, whereas glutamate, leucine, and valine consistently shortened lag time. These results show that amino acid requirements in B. longum extend beyond binary auxotrophy and provide a machine-learning framework for designing streamlined defined media.