Abstract
Average Nucleotide Identity (ANI) is becoming a standard measure for bacterial species delimitation. However, its calculation can take orders of magnitude longer than similarity estimates based on sampling of short nucleotides, compiled into so-called sketches. These estimates are widely used. However, their variable correlation with ANI has suggested that they might not be as accurate. For a where-the-rubber-meets-the-road assessment, we compared two sketching programs, mash and dashing, against ANI, in delimiting species among Esterobacterales genomes. Receiver Operating Characteristic (ROC) analysis found Area Under the Curve (AUC) values of 0.99, almost perfect species discrimination for all three measures. Subsampling to avoid over-represented species reduced these AUC values to 0.92, still highly accurate. Focused tests with ten genera, each represented by more than three species, also showed almost identical results for all methods. Shigella showed the lowest AUC values (0.68), followed by Citrobacter (0.80). All other genera, Dickeya, Enterobacter, Escherichia, Klebsiella, Pectobacterium, Proteus, Providencia and Yersinia, produced AUC values above 0.90. The species delimitation thresholds varied, with species distance ranges in a few genera overlapping the genus ranges of other genera. Mash was able to separate the E. coli + Shigella complex into 25 apparent phylogroups, four of them corresponding, roughly, to the four Shigella species represented in the data. Our results suggest that fast estimates of genome similarity are as good as ANI for species delimitation. Therefore, these estimates might suffice for covering the role of genomic similarity in bacterial taxonomy, and should increase confidence in their use for efficient bacterial identification and clustering, from epidemiological to genome-based detection of potential contaminants in farming and industry settings.