Detecting protein domains in metagenomic data

We will be cite-able soon! Please check back for updates on the NCBI Virus Discovery codeathon paper submission.

Rationale

Metagenomic data can be difficult to interpret. Which types of bacteria and viruses are present? Are specific sequences present that might be indicative of pathogenicity? One way to address these questions is to assess the protein domain content of a metagenome. Our tools allow users to perform a comprehensive annotation of domains within their metagenomic assembly (using 6-frame translation RPS-BLAST), allowing follow-up on specific occurrences of domains of interest. We also provide a set of domain matches found in metagenomics assemblies, that can be used as a "gold-standard" (or at least a silver one:-) for future experiments. Additionally, we provide an alternative version of our pipeline that we began testing (but did not deploy on our full dataset) that uses MASH on translated reads to perform a more quick & rough scan of the domain content of their metagenomic data.

Potential uses of these results include:

• Filtering your metagenome to just the contigs that contain a particular virus of interest
• Summarizing the gene/protein content of your metagenome
• Estimating taxonomy from domains or mapping your metagenome's domains onto known taxonomy (for more details see the Taxonomy Domain Integration repo)

Schematic description

How to use this pipeline

This pipeline takes the following as inputs:

(1) Query sequences, which should be contigs of a metagenome assembly*,

and also

(2) Domain models, which represent existing domain models - e.g., from CDD, PFAM, POGs/PVOGs, etc., in PSSM format.

We provide tools for the user to perform Domain search using 6-frame translation of Reverse Position-Specific BLAST (RPS-BLAST) (sometimes unofficially referred to as "RPS-tBLASTn"), or a non-optimized implementation of Mash.

*There are several ways to generate a metagenomic assembly; we built the one for our use case with SKESA.

Output

If using RPS-BLAST, a tab delimited file will be generated with added information on the source SRR sample:

Example table:

contig_id	CDD	pident	length	mismatch	gapopen	qstart	qend	sstart	send	evalue	bitscore	SRR	contig_id_only	contig_length
Contig_321_123.726:1.12237	CDD:222861	38.528	231	131	3	5068	5739	1	227	1.04e-47	162.0	SRR4451607	Contig_321_123.726	12237
NC_019445.1_3:1.15349	CDD:283078	91.379	58	5	0	14004	14177	1	58	3.46e-27	97.9	SRR4451607	NC_019445.1_3	15349
Contig_17_145.65:1.11778	CDD:165469	47.489	219	106	4	9292	8645	99	311	6.62e-73	238.0	SRR4451607	Contig_17_145.65	11778
Contig_256_126.089:1.6592	CDD:164995	47.674	172	84	4	1614	2120	1	169	7.67e-54	177.0	SRR4451607	Contig_256_126.089	6592
Contig_30_137.493:1.78458	CDD:222896	28.342	374	232	13	34344	33238	17	359	2.93e-43	158.0	SRR4451607	Contig_30_137.493	78458
Contig_11_130.746:1.25303	CDD:222896	23.82	445	272	19	19072	17810	8	409	7.54e-34	129.0	SRR4451607	Contig_11_130.746	25303

If using Mash, a file will be generated containing the following information:

• Domain ID
• Query sequence
• Estimated distance
• P-value
• Number of sketches in the query found in number of sketches for the domain.

Initial use case

We used the following data to assess the runtime, scalability, and accuracy of this pipeline:

(1) Query sequences are from assembled contigs found in Codeathon1. For RPS-BLAST, we used the translated ORFs from all 3,000+ datasets, while for MASH, we only used a subset of 700 of them.

(2) Domain models are from CDD.

Parallelization is built into the pipeline in a hard-coded form. We initially parallelized across 64 nodes on 10 cloud instances.

Example outputs