Ok-k-mer: Selecting k-mer length ensembles for calculating genomic distance

Current methods for computing sequence similarity across large groups of genome sequences tend to do so at the level of k-mers since this is far more efficient than base-by-base alignment. While these methods use default k-mer lengths, recent work shows that the particular choice of k-mer length greatly affects accuracy, and that using multiple values of k can boost accuracy relative to the best single choice of k, particularly when the goal is to estimate genomic distances that reflect true phylogenetic relationships. Here, we introduce Ok-k-mer, a tool to systematically select weighted k-mer length ensembles for pairwise genomic distance estimation, thereby increasing the accuracy of those estimates. Given a phylogenetic tree, Ok-k-mer ranks genome pairs by Jaccard distance across a range of k and uses a progressive search to maximize the Spearman correlation, ρ, between weighted-average k-mer ranks and branch distances. Across eight Genome Taxonomy Database datasets, Ok-k-mer ensembles improved correlation over standard default k-mer lengths. In the Chlamydiales taxonomy, a three-length ensemble k=(9,17,22) achieved a 23.9% relative improvement over the best single length (k=20) and a 36.7% improvement over a commonly used default of 31-mers. We demonstrate that Ok-k-mer k-mer length ensembles remain stable (i.e. lengths and weights change only moderately) under genome subsampling within a taxonomy but vary across datasets. Ok-k-mer can annotate all subtrees of a phylogeny, which reveals distinct patterns of ensembles reflecting potential differences in genome evolution. Ok-k-mer provides a framework for moving beyond a "one-k-fits-all" k-mer analysis, improving genomic distance estimation in the era of massive-scale genomics.