Data coherence over data volume drives generalisable genome-based prediction of microbial carbon utilisation

Concordance filtering of genotype–phenotype agreement enabled genome-based prediction of microbial carbon utilisation across four datasets, recovering reproducible signals for 12 of 15 shared carbon sources.

• Why it matters: Genome-based machine-learning models often generalise poorly across datasets, limiting prediction of microbial carbon utilisation from genomes when training sets are small or biased.
• What they did: The study integrated binary growth phenotypes for 819 strains across 240 carbon sources from four datasets, then restricted training to samples where experimental labels matched annotated pathway mechanisms.
• The result: Cross-dataset balanced accuracy dropped from 0.86 within datasets to 0.62, but in-clade restriction recovered only 0.03 of this loss; concordance filtering doubled reproducible sources (6 to 12) and improved efficiency versus augmentation with up to 1800 BacDive genomes per phenotype.