Lightweight compositional analysis of metagenomes with FracMinHash and minimum metagenome covers

Key results

1. ShownOn the simulated Critical Assessment of Metagenome Interpretation (CAMI) mouse gut benchmark, taxonomic profiles built from minimum metagenome covers retained high completeness down to species level and had the highest species-level purity of the 10 profilers compared.

   Species level: 88.7% completeness; purity 95.9% with predictions under 1% abundance filtered and 97% unfiltered, highest among methods. Also second-lowest L1-norm error, most true positives and fewest false positives. Sixty-four CAMISIM samples from 791 genomes, profiled against the CAMI Jan 8, 2019 RefSeq snapshot and scored with OPAL; other methods lose completeness below genus. · Minimum metagenome covers can accurately…; Minimum metagenome covers support accurate and…; Taxonomy

2. ShownAgainst 725,339 GenBank genomes, greedy covers were small (≈100 genomes for several metagenomes; 19 for one mock) despite overlaps with tens or hundreds of thousands of genomes: from redundancy removal where k-mers are mostly identified, but for the oil well sample because much of it is absent from the database.

   Four metagenomes (podar and zymo mocks, human gut, oil well) at scaled=2000, k=31, counting matches above 100,000 estimated shared k-mers: covers of 19 and 99 genomes for one mock and one real sample that match 406k and 96k genomes; zymo mock's Salmonella resolves to one genome among over 200,000 matches; gut real matches over 75,000 E. coli genomes. · Minimum metagenome covers select small subsets…; Minimum set covers can be used for accurate…; GenBank database sketching and searches

3. ShownFracMinHash containment estimates in sourmash were about as accurate as MinHash-based containment tools on a mock community, with sketch sizes comparable to MinHash sketches at n=1000 to 10000.

   podar mock, k=21/31/51 vs exact k-mer sets: all methods within 1% of exact containment on average; CMash consistently underestimated; Mash Screen n=10000 closest at k=21 and 31, then FracMinHash scaled=1000 and Mash Screen n=1000. sourmash sketches held 431 to 9540 hashes (median 2741). · A FracMinHash implementation accurately…; FracMinHash provides efficient containment…; Comparison between CMash, mash screen, and…

4. ShownGreedy rank ordering splits a metagenome's k-mers among genomes, with k-mers shared by related strains credited to the earlier match; in the podar mock just over half of the k-mers matched GenBank.

   podar mock: 205m (54.8%) of 375m k-mers found in GenBank; the remaining 169m (45.2%) are put down to sequencing errors or unknown community members. Two Shewanella baltica strains share approximately 50% of their k-mers; the shared ones go to OS223 (rank 8), so OS185 (rank 33) is matched only on its own k-mers. · FracMinHash can be used to construct a minimum…; Figure legends

5. SuggestedMapping reads to cover genomes one at a time in rank order, removing reads already mapped, gives per-genome coverage that tracks the k-mer estimates, supporting use of the cover to pick mapping references.

   Four Table 1 metagenomes, minimap2 -x sr via genome-grist, first 36 cover genomes, unrescaled axes: close agreement for the two mocks; for the gut and oil well samples mapping coverage tends to exceed k-mer estimates, possibly because only species-level, not strain, genomes are in the database. No summary agreement statistic is reported. · Minimum metagenome covers provide representative…; Read mapping and hybrid mapping pipeline; Figure legends