Fast and sensitive taxonomic assignment to metagenomic contigs

1Metagenomic studies shine a light on previously unstudied parts of the tree of life. However, unraveling taxonomic composition accurately and quickly remains a challenge. While most methods label short metagenomic reads (reviewed in ), only a handful (e.g. ) assign entire contigs, even though this should lead to improved accuracy.

2Recently, developed CAT, a tool for taxonomic annotation of contigs based on protein homologies to a reference database. It combines Prodigal for predicting open reading frames (ORFs), DIAMOND to search with the translated ORFs, and logic to aggregate individual ORF annotations. CAT achieved higher precision than state-of-the-art tools on bacterial benchmarks. Despite its advantage over existing methods, CAT has limitations: (i) Prodigal was designed for prokaryotes and not eukaryotes; (ii) Prodigal runs single-threaded, limiting applicability to metagenomics; (iii) CAT’s r parameter determines the cut-off score below each ORF’s top-hit above which hits are included in the ORF’s lowest common ancestor (LCA) computation. Although the authors provide guidelines to set r, it is unclear how general they are.

3Here, we present MMseqs2 taxonomy, a novel protein-search-based tool for taxonomy assignment to contigs. It overcomes the aforementioned limitations by extracting all possible protein fragments, covering the coding repertoire of all domains of life. It quickly eliminates fragments that do not bear minimal similarity to the reference database, and searches with the remaining ones. MMseqs2 taxonomy uses an approximate 2bLCA strategy to assign translated fragments to taxonomic nodes (Supplementary Material). The hits for the approximate 2bLCA computation are determined automatically, saving the need to tune an equivalent of CAT’s r parameter. It outperforms CAT on bacterial and eukaryotic datasets.

4Input. Contigs are provided as (compressed) FASTA/Q files. As reference, the databases workflow can download and prepare various public taxonomy databases, such as, nr, UniProt or GTDB. Alternatively, users can prepare their own taxonomic reference database (see MMseqs2 wiki).

5Algorithm. The four main steps are described in Figure 1A.

6Output. MMseqs2 taxonomy returns the following eight fields for each contig accession: (i) the taxonomic identifier (taxid) of the assigned label, (ii) rank, (iii) name, followed by the number of fragments: (iv) retained, (v) taxonomically assigned, and (vi) in agreement with the contig label (i.e. same taxid or have it as an ancestor), (vii) the support the taxid received and, optionally, (viii) the full lineage. The result can be converted to a TSV-file, and to a Kraken report or a Krona visualization (Supplementary Material).

7Bacterial dataset. The CAMI-I high-complexity challenge and its accompanying RefSeq 2015 reference database were given to MMseqs2 and CAT. AMBER v2 was used to assess the taxonomic assignment by computing the average completeness (Fig. 1B) and purity (Supplementary Fig. S1) bp using its taxonomic binning benchmark mode. At similar assignment quality, MMseqs2 taxonomy is 18× faster than CAT. Using the nr, MMseqs2 is 10× faster (Supplementary Fig. S2).

8Eukaryotic dataset. All 57 SAR (taxid 2698737) RefSeq assemblies and their taxonomic labels were downloaded from NCBI in 08/2020. To resemble metagenomic data, their scaffolds were randomly divided following the length distribution of contigs assembled for sample ERR873969 of eukaryotic Tara Oceans, resulting in 2.7 million non-overlapping contigs with a minimal length of 300 bp. Using nr from 08/2020, MMseqs2 classified more contigs than CAT (62% versus 47%). For 36%, CAT extracted a fragment that did not hit the reference, suggesting fragments extracted by MMseqs2 are more informative for eukaryotic taxonomic annotation (Fig. 1C, Supplementary Fig. S3).

9MMseqs2 taxonomy is as accurate as CAT on a bacterial dataset while being 3–18× faster and requiring fewer parameters. Its extracted fragments make it suitable for analyzing eukaryotes. It is accompanied by several taxonomy utility modules to assist with taxonomic analyses.

10E.L.K. is a FEBS long-term fellowship recipient and an EMBO nonstipendiary long-term fellow. The work was supported by the BMBF CompLifeSci project horizontal4meta; the ERC’s Horizon 2020 Framework Programme [‘Virus-X’, project no. 685778]; the National Research Foundation of Korea grant funded by the Korean government (MEST) [2019R1A6A1A10073437, NRF-2020M3A9G7103933]; and the Creative-Pioneering Researchers Program through Seoul National University.

11Conflict of Interest: none declared.

12The data used to benchmark MMseqs2 taxonomy in this study are openly available from https://data.cami-challenge.org/participate at the Databases section. The SAR assemblies were downloaded from NCBI in 08/2020 and processed as described.