NEIGHBORHOOD CLUSTERING TO ANALYSE ANTIMICROBIAL RESISTANCE IN BACTERIAL GENOMES

Abstract

Valuable insight into gene function and evolution can be obtained by analysing the order of genes in prokaryotic genomes, as neighboring genes often share related functions and evolutionary histories. Obtaining precise functional predictions is particularly important in the case of antimicrobial resistance (AMR) genes, as subtle differences in similarity patterns can reflect the potential for an organism to be treatable or resistant to one or more antibiotics. Databases such as the Comprehensive Antibiotic Resistance Database (CARD) provide high-quality predictions, but there is a significant gray area (“Loose hits” according to CARD) where genes differ in sequence from the reference sequence and may or may not confer AMR.

We introduce an approach to compare the genomic neighborhoods of AMR genes in genomes with different degrees of relatedness, to provide additional insight into their potential function. Our approach uses a technique to identify candidate AMR, then applies novel similarity measures and application of the UPGMA, MCL and DBSCAN graph-clustering techniques to identify patterns of similarity among gene neighborhoods. This analysis is complemented by phylogenetic analysis to assess the similarity of identified genes as well as their neighborhoods. We also provide a graphical tool to visualize the gene content in sets of neighborhoods.

AMR gene neighborhoods were observed to be very similar within closely related members of species including Salmonella Heidelberg. The proximity of some Loose hits to other AMR genes in many neighborhoods provided additional evidence for their function, whereas in other cases the CARD Loose hits were isolated and likely not associated with AMR. We also considered a set of genomes that encompassed several enteric pathogens. In this set, we found cases where seemingly poor Loose predictions were associated with clusters of AMR genes, and instances where gene order was surprisingly similar across distantly related genomes which may indicate recent transmission of AMR genes between pathogenic organisms. Our method provides new insights into the function of candidate AMR genes, and these refined predictions can be used to predict resistance and identify candidate evolutionary events.