RESEARCH HIGHLIGHTS

MICROBIOLOGY The strain in metagenomics

Two computational tools extract strain- His Latent Strain Analysis (LSA) assesses level information from reams of micro- the covariation of short ‘k-mer’ sequences bial sequence data. found in reads, on the basis of a hashing Doctors are well aware that differences function from his search algorithm that between strains of the same bacterial spe- represents all k-mer abundance patterns in cies can have consequences at opposite ends a few gigabytes of fixed memory, regardless of the health-and-disease spectrum. “If you of data size (Cleary et al., 2015). LSA also only look at a species-level annotation, you speeds up the covariation-detection and may end up missing the fact that there are assembly steps. virulence genes in some E. coli strains,” says The advantages are manifold. LSA can Ramnik Xavier of Massachusetts General Stock Photo Alamy discriminate highly related strains, and it Hospital, the Broad Institute and Harvard Microbes from the wild are now being studied at has worked efficiently on four terabytes University. Identifying strains, along with the strain level. of gut microbiome data, whereas other the genetic potential of their genomes, has approaches top out at around 100 gigabytes. become a key goal for computational biolo- tion. Gevers and Xavier say that ConStrains It also found microbes that are missed by gists, but the challenges are compounded shines when it comes to longitudinal stud- traditional assembly because they contrib- by enormous data sets, missing reference ies. The researchers used it to track strains ute few reads to any single sample. “If you sequences and rare species. in large published data sets, including five could get all those reads put into one box, At the Broad Institute, Dirk Gevers, strains of Bifidobacterium longum in infant that box may have 30-fold coverage,” Alm Xavier and their colleagues began thinking guts associated with distinct functions that explains. One family they detected made up about the problem when they analyzed the were missed in species-level analyses. just one–10-millionth of the gut data. first Human Microbiome Project shotgun Nearby at the Massachusetts Institute of LSA assembled up to 90% of some

Nature America, Inc. All rights reserved. America, Inc. © 201 5 Nature data. “We detected a person-specific profile Technology, Eric Alm was addressing simi- genomes, enabling a deeper functional at the subspecies level but could not decon- lar questions with single-cell sequencing understanding. Very similar regions that volute it back then,” says Gevers, now at the when he was pulled back into metagenomic cannot be distinguished are often inter- Janssen Human Microbiome Institute. It analysis. “It was almost accidental,” he says. preted as separate genomes, however. npg took postdoc Chengwei Luo to ultimately A chance e-mail introduced him to Brian “That’s a neat opportunity for someone to drive development of their ConStrains Cleary, just off a post-college stint develop- develop a new tool” to put together scattered strain-detection software. ing algorithms on Wall Street. Cleary was genomes—something they are working on, In principle, unique patterns of single- working on an algorithm that searched says Alm. He estimates that LSA works best nucleotide polymorphisms (SNPs) can for words across documents, which Alm with 100 or more gut samples, a study size identify strains, but SNP-based approaches realized was similar to finding patterns in that is becoming more commonplace. rely on reference sequences or are limited sequence reads. “Within a couple of weeks The researchers will use their tools to to the SNPs that appear in a short sequence he came back and he had something that understand how strains function in their read. In the ConStrains approach, reads are was starting to actually work for metage- environments. “We want to do de novo recruited using universal core genes pres- nome assembly,” says Alm. assembly-based strain calling,” and to ent in all species with tenfold sequence Reference-free genome assembly com- link strains to function, says Gevers. After coverage. The software then uses a refer- bines short reads into longer stretches ­census-taking and correlation, “the next step ence-independent strategy to cluster and called contigs—a computationally intensive would be…to get to causality,” says Xavier. concatenate the SNPs into strain-specific step when starting from complex microbial Tal Nawy fingerprints. ConStrains brings quantita- mixtures. One way to then assign contigs tive metagenomics to the strain rather than to species ‘bins’ is to track their abundance RESEARCH PAPERS the species level, without the need for addi- across samples; those that vary in the same Luo, C. et al. ConStrains identifies microbial strains tional types of data (Luo et al., 2015). way probably belong to the same species. in metagenomic datasets. Nat. Biotechnol. 33, 1045–1052 (2015). Because fingerprints are identified from Cleary’s insight was to assign reads to Cleary, B. et al. Detection of low-abundance bacterial single samples, entire cohorts need not be species bins first, then assemble contigs strains in metagenomic data sets by eigengenome analyzed at once, which eases computa- from these much smaller piles of data. partitioning. Nat. Biotechnol. 33, 1053–1060 (2015).

NATURE METHODS | VOL.12 NO.11 | NOVEMBER 2015 | 1005