Machine learning predicts and provides insights into milk acidification rates of Lactococcus lactis

• Published in: PloS one Vol. 16; no. 3; p. e0246287
• Main Authors: Karlsen, Signe Tang, Vesth, Tammi Camilla, Oregaard, Gunnar, Poulsen, Vera Kuzina, Lund, Ole, Henderson, Gemma, Bælum, Jacob
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 15.03.2021; Public Library of Science (PLoS)
• Subjects: Acidification; Algorithms; Antibiotic resistance; Antibiotics; Bacteria; Biology and Life Sciences; Carbon sources; Cheese; Computer and Information Sciences; Decision trees; Drug resistance; Gene sequencing; Genetic aspects; Genomes; Genotype & phenotype; Lactococcus; Learning algorithms; Machine learning; Medicine and Health Sciences; Microorganisms; Milk; Phenotypes; Polysaccharides; Prediction models; Regression analysis; Saccharides; Strains (organisms); Denmark
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0246287

Lactococcus lactis strains are important components in industrial starter cultures for cheese manufacturing. They have many strain-dependent properties, which affect the final product. Here, we explored the use of machine learning to create systematic, high-throughput screening methods for these properties. Fast acidification of milk is such a strain-dependent property. To predict the maximum hourly acidification rate (Vmax), we trained Random Forest (RF) models on four different genomic representations: Presence/absence of gene families, counts of Pfam domains, the 8 nucleotide long subsequences of their DNA (8-mers), and the 9 nucleotide long subsequences of their DNA (9-mers). Vmax was measured at different temperatures, volumes, and in the presence or absence of yeast extract. These conditions were added as features in each RF model. The four models were trained on 257 strains, and the correlation between the measured Vmax and the predicted Vmax was evaluated with Pearson Correlation Coefficients (PC) on a separate dataset of 85 strains. The models all had high PC scores: 0.83 (gene presence/absence model), 0.84 (Pfam domain model), 0.76 (8-mer model), and 0.85 (9-mer model). The models all based their predictions on relevant genetic features and showed consensus on systems for lactose metabolism, degradation of casein, and pH stress response. Each model also predicted a set of features not found by the other models.