Inter-species metabolic interactions in cheese flavour formation

Abstract

Cheese fermentation and flavour formation are governed by complex biochemical reactions driven by polymicrobial activity. While the compositional dynamics of cheese microbiomes is relatively well mapped, the mechanistic role of microbial interactions in flavour formation is yet unknown. We microbially and metabolically characterised a year-long Cheddar cheese process using a commonly used starter culture containing Streptococcus thermophilus and Lactococcus strains. By using an experimental strategy whereby certain strains were left out from the starting mixture, we identified the critical role of S. thermophilus in boosting Lactococcus growth and in shaping flavour compound profile. Controlled milk fermentations with systematic exclusion of single Lactococcus strains, combined with genomics, genome-scale metabolic modelling, and metatranscriptomics, indicated that proteolytic activity of S. thermophilus relieves nitrogen limitation for Lactococcus and boosts de novo nucleotide biosynthesis. While S. thermophilus had large contribution to the flavour profile, L. cremoris also played a role by limiting diacetyl and acetoin formation which leads to off-flavour when in excess. This off-flavour control could be attributed to different metabolic re-routing of citrate between L. cremoris and other L. lactis strains. Further, closely related L. lactis strains exhibited different interaction patterns with S. thermophilus highlighting the importance of strain-specificity in cheese-making. Overall, our results bring forward the critical role of competitive and cooperative microbial interactions shaping cheese flavour profile.

Improving genome-scale metabolic models of incomplete genomes with deep learning

“…Understanding these factors can show the strengths of the network, show possible areas of improvement, and provide new insights into the gap-filling problem. We identified two important factors that affect the accuracy of predictions by the neural network: (1) the frequency of reaction across all bacteria and (2) the phylogenetic distance between the organisms in the testing and training dataset.…”

“…Simulating microbial metabolism is an effective method to understand bacterial physiology and interactions within their communities (1)(2)(3). The functions and interactions of bacteria can be inferred from their genome sequences using genomescale metabolic models (GSMMs) (3)(4)(5)(6).…”

“…The DNNGIOR neural network learns weights for missing reactions in incomplete GSMMs from patterns in the presence and absence of metabolic reactions in genomes spanning the bacterial domain. We identified two important factors contributing to prediction accuracy: (1) the frequency of reaction across all bacteria, and (2) the phylogenetic distance between the query and the genomes in the training dataset. Reactions that occur in > 30% of the training genomes can be accurately predicted (Mean F1 score = 0.85).…”