Predicting variable gene content in Escherichia coli using conserved genes.

Having the ability to predict the protein-encoding gene content of an incomplete genome or metagenome-assembled genome is important for a variety of bioinformatic tasks. In this study, as a proof of concept, we built machine learning classifiers for predicting variable gene content in genomes using only the nucleotide k-mers from a set of 100 conserved genes as features. Protein families were used to define orthologs and a single classifier was built for predicting the presence or absence of each protein family occurring in 10%-90% of all genomes. The resulting set of 3,259 extreme gradient boosting classifiers had a per-genome average macro F1 score of 0.944 [0.943-0.945, 95% CI]. We show that the F1 scores are stable across multi-locus sequence types and that the trend can be recapitulated by sampling a smaller number of core genes or diverse input genomes. Surprisingly, the presence or absence of poorly annotated proteins, including "hypothetical proteins" was accurately predicted (F1 = 0.902 [0.898-0.906, 95% CI]). Models for proteins with horizontal gene transfer-related functions had slightly lower F1 scores but were still accurate (F1s = 0.895, 0.872, 0.824, and 0.841 for transposon, phage, plasmid, and antimicrobial resistance-related functions, respectively). Finally, using a holdout set of 419 diverse genomes that were isolated from freshwater environmental sources, we observed an average per-genome F1 score of 0.880 [0.876-0.883, 95% CI], demonstrating the extensibility of the models. Overall, this study provides a framework for predicting variable gene content using a limited amount of input sequence data. IMPORTANCE Having the ability to predict the protein-encoding gene content of a genome is important for assessing genome quality, binning genomes from shotgun metagenomic assemblies, and assessing risk due to the presence of antimicrobial resistance and other virulence genes. In this study, we built a set of binary classifiers for predicting the presence or absence of variable genes occurring in 10%-90% of all publicly available genomes. Overall, the results show that a large portion of the variable gene content can be predicted with high accuracy, including genes with functions relating to horizontal gene transfer. This study offers a strategy for predicting gene content using limited input sequence data.