Intra-genomic genes-to-genes correlation enables genome representation

Abstract

AbstractThe pan-genome consists of core genes shared by all members of a taxonomy and accessory genes found in only a subset, holding the keys to advancing our understanding of evolution and tackling medical challenges. Here, we discovered a strong intra-genomic correlation among bacterial genes within each ofEscherichia coli,Listeria monocytogenes,Staphylococcus aureus, andCampylobacter jejuni. With a convolutional neural network assisted core genome knock-out simulation, we found that different combinations of fewer than 20 highly variable core genes could recover the sub-species type classified by complete core genome with accuracy >95%. This observation led us to test the genes-to-genes predictability: with more than 52,000 assemblies from each species, combinations of highly variable core genes could predict the sequence variants of other core genes (average accuracy >94%) within the same genome and could also predict sequence variants (average accuracy >91%) as well as the presence (average AUROC >0.91) of some accessory genes. Furthermore, combinations of highly variable core genes could also predict multiple antibiotic resistances (AUROC >0.80) in large published datasets ofE. coli,S. aureus, andMycobacterium tuberculosis. Collectively, we propose that genes within the same genome can strongly correlate with each other. Therefore, the strain phylogeny and the stauts of other genes could be uniformly represented by combinations of highly variable core genes, which could further represent certain phenotypes includingin vitroresistance.