MMseqs2

What is MMseqs2?

MMseqs2 (Many-against-Many sequence searching) is an open-source sequence search and clustering suite for protein and nucleotide data. Its k-mer prefilter and vectorized alignment stages make large database searches much faster than conventional pairwise tools while retaining sensitive local alignment.

Search mode finds homologs between query and target sets. Cluster mode groups related sequences with a cascaded workflow, while Linclust uses a linear-time prefilter for very large collections. For multiple sequence alignment after collecting homologs, Clustal Omega can align the selected sequences.

How to use MMseqs2 online

ProteinIQ runs MMseqs2 18 online from one FASTA or FASTQ query set. Search mode also takes a FASTA or FASTQ target set and returns the native 12-column m8 file. Cluster and Linclust modes return native membership, representative-sequence, and all-sequence files, together with run diagnostics and reproducibility metadata.

Inputs

Input	Description
Query sequences	One protein, DNA, or RNA sequence set in FASTA or FASTQ format. Clustering requires at least two records.
Target sequences	One protein, DNA, or RNA sequence set in FASTA or FASTQ format. Required only for search mode.

Sequence type uses MMseqs2's native auto mode by default. An explicit protein or nucleotide choice maps to the source search or database-type option. Auto mode is safer than inferring the alphabet locally, especially for nucleotide records containing ambiguity symbols such as N.

Modes and defaults

Setting	Search	Cluster	Linclust
Sensitivity	5.7	4.0	Determined by Linclust
Minimum sequence identity	Not used	0	0
Minimum coverage	0	0.8	0.8
Coverage mode	0	0	0
Maximum hits per query	300	Not used	Not used
E-value threshold	0.001	Not used	Not used

Search and cluster sensitivity can range from 1 to 7.5. Higher values examine more candidate matches and take longer. Linclust does not accept the same sensitivity control.

Coverage modes

The Minimum coverage value is interpreted according to Coverage mode:

• Mode 0: Require the threshold for both query and target coverage.
• Mode 1: Apply the threshold to target coverage.
• Mode 2: Apply the threshold to query coverage.
• Mode 3: Require the target length to be at least the threshold multiplied by the query length.
• Mode 4: Require the query length to be at least the threshold multiplied by the target length.
• Mode 5: Require the shorter sequence to be at least the threshold multiplied by the longer sequence length.

Modes 3 to 5 are length-ratio filters. They are not bidirectional alignment-coverage calculations.

Clustering algorithms

• Set-cover (mode 0): Greedily selects representatives that cover many sequences. This is the default.
• Connected component (mode 1): Groups transitively connected sequences.
• Greedy by sequence length (mode 2): Uses the CD-HIT-like length-ordered strategy.
• Greedy by sequence length, low memory (mode 3): Uses the same length-ordered family with lower memory use.

In standard Cluster mode, Reassign clusters performs the native post-clustering reassignment step. Linclust does not support that option. Every run uses the four CPUs allocated to it and the native automatic memory split.

Understanding search results

The primary search file is MMseqs2's native, headerless BLAST-TAB output. ProteinIQ does not replace its identity unit or insert a synthetic header.

Position	Native field	Meaning
1	query	Query identifier
2	target	Target identifier
3	fident	Fraction of identical aligned residues, from 0 to 1
4	alnlen	Alignment length
5	mismatch	Mismatch count
6	gapopen	Gap-opening count
7-8	qstart, qend	Query alignment coordinates
9-10	tstart, tend	Target alignment coordinates
11	evalue	Expected number of chance matches
12	bits	Bit score

Lower E-values and higher bit scores indicate stronger evidence for homology. Identity and coverage should be interpreted together: a high identity over a short segment does not establish full-length similarity.

Understanding clustering results

MMseqs2 creates three native files:

File	Contents
*_cluster.tsv	Headerless representative-member pairs
*_rep_seq.fasta	One representative sequence per cluster
*_all_seqs.fasta	All input sequences grouped by cluster

The default identity threshold is 0, not 30%. To remove near-duplicates, a threshold such as 0.9 or 0.95 is a common starting point. Lower thresholds can merge more remote relationships, but the biological meaning depends on sequence length, coverage mode, and protein family.

Linclust is preferable when the input contains hundreds of thousands or millions of sequences. Standard cluster mode performs a more sensitive cascaded search and is generally better for smaller sets where cluster quality matters more than linear scaling.