SKA: Split Kmer Analysis Toolkit for Bacterial Genomic Epidemiology

Key results

1. ShownOn 55 MRSA outbreak genomes, a reference-free SKA alignment gave a tree topology consistent with those from SNP Pipeline and from a manually curated gold standard; branch lengths differed only by each method's handful of missed or false SNPs.

   Against a curated alignment to a complete outbreak genome (GCA_001369415.1), SKA missed eight variant sites (four in repeats longer than the split k-mer, four from two pairs of variants within a k-mer length); SNP Pipeline with the HO 5096 0412 reference missed seven and called five false sites. SNP Pipeline needed nearly 17 times the CPU time, over 70 times the disk space and 7 times the memory. · Reconstructing the outbreak

2. ShownStarting from raw reads, with no reference genome chosen and no species information, SKA distance clustering recovered the 45-isolate MRSA outbreak on a special care baby unit (SCBU) from 65 sequenced isolates on one CPU.

   With defaults (20-SNP cutoff, minimum identity 0.9) the 65 samples formed 16 clusters, four with two or more isolates; the largest held the 45 isolates of the original investigation (pairwise distance 0 to 12 SNPs). The ska fastq, merge, summary and distance steps together took 105.4 CPU minutes with 1.4Gb peak memory. · Pairwise distance calculation and clustering

3. ShownSplit k-mers unique to the outbreak isolates screened 1,683 clonal complex 22 S. aureus genomes from a regional survey in 116s and singled out the ten isolates that Toleman et al. had linked to the outbreak, while ruling out the nearest other candidate.

   ska unique found 1,714 split k-mers found only in the 45 outbreak isolates (20 non-outbreak isolates as outgroup). Scanning 1,683 clonal complex 22 genomes took 116s and 7Mb: ten matched 100% with no SNPs, one matched 34% (34 SNPs), no other above 5%. The ten were within 11 SNPs of an outbreak isolate and typed ST2371; the other was 53 SNPs away and ST22. · Identifying new linked cases

4. ShownOn four foodborne outbreak benchmark datasets (Timme et al.), reference-free SKA trees resembled the published reference-mapping trees, but fixed 20-SNP clusters split the diverse L. monocytogenes outbreak in two (as its tree also did) and absorbed one S. enterica outgroup.

   Sets of 23 C. jejuni, 9 E. coli, 31 L. monocytogenes and 22 S. enterica ser. Bareilly each needed well under two hours of single-core CPU from raw reads, and less than 1.5 minutes in total once split k-mers existed. Clusters matched the expected outbreaks for C. jejuni and E. coli; outgroup CFSAN000189 joined the S. enterica cluster but sat apart on the IQ-Tree tree. · Results; Results; Description

5. SuggestedOn 96 simulated genomes, IQ-Tree trees from SKA reference-free alignments came close to the artificial gold standard of Lees et al. (RAxML on reads mapped to an artificially close reference); SKA map trees were weaker on this diverse set.

   Kendall-Colijn and Robinson-Foulds distances to the true tree sit in Supplementary Table 1, absent from the extracted text, so only rankings are available: reads beat assemblies slightly, and bioNJ on SKA Mash-like distances beat every realistic method of Lees et al. kSNP3 trees were comparable but it needed more time, memory and disk (over 50,000 intermediate files, more than 15Gb). · Results; Description; Analysis of simulated data from Lees et al. ()…