Models trained with noisy genomes extend bacterial phenotype prediction into deep time

Abstract

Predicting phenotype from genotype in extant organisms is increasingly tractable through the accumulation of genome sequences and the development of machine-learning algorithms. Here we show that machine learning can be applied to reconstructed ancestral gene content, extending these predictions into the past. We trained models on a diverse set of bacterial phenotypes and found that introducing noise into gene content profiles allows predictions to generalize over larger evolutionary distances. For phenotypes with signal spread across many genes - such as metabolic oxygen use, cell envelope architecture and optimal growth temperature - noise augmentation extends resolution back to the root of the bacterial domain, while for other phenotypes - including GC content and sporulation - the range remains more limited. We therefore conclude that the last bacterial common ancestor (LBCA) was likely an anaerobic, double-membraned, and moderately thermophilic bacterium (46-75{degrees}C). Moreover, this work provides a general approach for learning about the genomic basis of phenotypes and drawing inferences about their early evolution.