A robust and fast clustering method for amplicon-based studies.

The purpose of swarm is to provide a novel clustering algorithm that handles massive sets of amplicons. Results of traditional clustering algorithms are strongly input-order dependent, and rely on an arbitrary global clustering threshold. swarm results are resilient to input-order changes and rely on a small local linking threshold d, representing the maximum number of differences between two amplicons. swarm forms stable, high-resolution clusters, with a high yield of biological information.

To help users, we describe a complete pipeline starting from raw fastq files, clustering with swarm and producing a filtered occurrence table.

swarm 3.0 introduces:

• a much faster default algorithm,
• a reduced memory footprint,
• binaries for Windows x86-64, macOS ARM64, GNU/Linux ARM64, and GNU/Linux POWER8,
• an updated, hardened, and thoroughly tested code (806 carefully crafted black-box tests),
• strict dereplication of input sequences is now mandatory,
• --seeds option (-w) now outputs results sorted by decreasing abundance, and then by alphabetical order of sequence labels.

swarm 2.0 introduced several novelties and improvements over swarm 1.0:

• built-in breaking phase now performed automatically,
• possibility to output cluster representatives in fasta format (option -w),
• fast algorithm now used by default for d = 1 (linear time complexity),
• a new option called fastidious that refines d = 1 results and reduces the number of small clusters.

swarm is a single-linkage clustering method, with some superficial similarities with other clustering methods (e.g., Huse et al, 2010). swarm's novelty is its iterative growth process and the use of sequence abundance values to delineate clusters. swarm properly delineates large clusters (high recall), and can distinguish clusters with as little as two differences between their centers (high precision).

swarm uses a local clustering threshold (d), not a global clustering threshold like other algorithms do. Users may be tempted to convert a 97%-global similarity threshold into a number of differences, and to use large d values. This is not a correct use of swarm. Clusters produced by swarm are naturally larger than d, and tests have shown that using the default d value (d = 1) gives good results on most datasets. Using the new fastidious option further improves the quality of results. For long amplicons or shallow sequencing, higher d values can be used (d = 2 or d = 3, very rarely more).

swarm produces high-resolution results, especially when using d = 1. Under certain rare conditions though, a given marker may not evolve fast enough to distinguish molecular taxa. If it concerns abundant sequences, swarm may form a cluster with a large radius, whereas classic clustering methods will cut through randomly, forcing delineation where the 97%-threshold falls. So, keep in mind that molecular markers have limitations too.

swarm most simple usage is:

./swarm amplicons.fasta

That command will apply default parameters (-d 1) to the fasta file amplicons.fasta. The fasta file must be formatted as follows:

>seqID1_1000
acgtacgtacgtacgt
>seqID2_25
cgtcgtcgtcgtcgt

where sequence identifiers are unique and end with a value indicating the number of occurrences of the sequence (e.g., _1000). Alternative format is possible with the option -z, please see the user manual. Swarm requires each fasta entry to present a number of occurrences to work properly. That crucial information can be produced during the dereplication step.

Use swarm -h to get a short help, or see the user manual for a complete description of input/output formats and command line options.

The memory footprint of swarm is roughly 0.6 times the size of the input fasta file. When using the fastidious option, memory footprint can increase significantly. See options -c and -y to control and cap swarm's memory consumption.

Get the latest binaries for GNU/Linux, macOS or Windows from the release page. Get the source code from GitHub using the ZIP button or git clone, and compile swarm with GCC (version 4.8.5 or more recent) or with clang (version 9 or more recent):

git clone https://github.com/torognes/swarm.git
cd swarm/
make

# or, with clang
make CC="clang-9" CXX="clang++-9"

If you have administrator privileges, you can make swarm accessible for all users. Simply copy the binary ./bin/swarm to /usr/local/bin/ or to /usr/bin/. The man page can be installed this way:

cd ./man/
gzip -c swarm.1 > swarm.1.gz
mv swarm.1.gz /usr/local/share/man/man1/
# or
mv swarm.1.gz /usr/share/man/man1/

Once installed, the man page can be accessed with the command man swarm.

(thanks to GitHub user Gian77 for reporting this procedure)

Assuming you already have a conda set-up (anaconda or miniconda), start by activate an environment with python 3:

conda activate py3

Make sure you have all the necessary channels for the bioconda packages:

conda config --add channels defaults
conda config --add channels bioconda
conda config --add channels conda-forge

List the different versions of swarm available and install one:

conda search -c bioconda swarm
conda install -c bioconda swarm=3.0.0=hc9558a2_0
swarm --version  # check

To facilitate the use of swarm, we provide examples of shell commands that can be use to format and check the input fasta file. Warning, these examples may not be suitable for very large files.

We assume your SFF or FASTQ files have been properly pair-assembled (with vsearch for example), trimmed from adaptors and primers (with cutadapt for example), and converted to fasta.

In a sample, or collection of sample, a given sequence may appear several times. That number of strictly identical occurrences represents the abundance value of the sequence. Swarm requires all fasta entries to present abundance values to be able to produce high-resolution clusters, like this:

>seqID1_1000
acgtacgtacgtacgt
>seqID2_25
cgtcgtcgtcgtcgt

were seqID1 has an abundance of 1,000 and seqID2 has an abundance of 25 (alternative formats are possible, please see the user manual).

The role of the dereplication step is to identify, merge and sort identical sequences by decreasing abundance. Here is a command using vsearch v1.3.3 or superior:

vsearch \
    --derep_fulllength amplicons.fasta \
    --sizeout \
    --relabel_sha1 \
    --fasta_width 0 \
    --output amplicons_linearized_dereplicated.fasta

The command performs the dereplication, the linearization (--fasta_width 0) and the renaming with hashing values (--relabel_sha1). If you can't or don't want to use vsearch, here is an example using standard command line tools:

grep -v "^>" amplicons_linearized.fasta | \
grep -v [^ACGTacgt] | sort -d | uniq -c | \
while read abundance sequence ; do
    hash=$(printf "${sequence}" | sha1sum)
    hash=${hash:0:40}
    printf ">%s_%d_%s\n" "${hash}" "${abundance}" "${sequence}"
done | sort -t "_" -k2,2nr -k1.2,1d | \
sed -e 's/\_/\n/2' > amplicons_linearized_dereplicated.fasta

Amplicons containing characters other than "ACGT" are discarded. The dereplicated amplicons receive a meaningful unique name (hash values), and are sorted by decreasing number of occurrences and by hash values (to guarantee a stable sorting). The use of a hashing function also provides an easy way to compare sets of amplicons. If two amplicons from two different sets have the same hash code, it means that the sequences they represent are identical.

If for some reason your fasta entries don't have abundance values, and you still want to run swarm (not recommended), you can specify a default abundance value with swarm's --append-abundance (-a) option to be used when abundance information is missing from a sequence.

Here is a typical way to use swarm:

./swarm \
    --fastidious \
    --threads 4 \
    --seeds cluster_representatives.fasta \
    amplicons.fasta > /dev/null

swarm will partition your dataset with the finest resolution (local number of differences d = 1 by default, built-in elimination of potential chained clusters, fastidious processing) using 4 CPU-cores. cluster representatives will be written to a new fasta file, other results will be discarded (/dev/null).

See the user manual for details on swarm's options and parameters.

To facilitate the use of swarm, we provide examples of options or shell commands that can be use to parse swarm's output. We assume that the amplicon fasta file was prepared as describe above (linearization and dereplication).

The chain-breaking, which used to be performed in a second step in swarm 1.0, is now built-in and performed by default. It is possible to deactivate it with the --no-otu-breaking option, but it is not recommended. The fastidious option is recommended when using d = 1, as it will reduce the number of small clusters while maintaining a high clustering resolution. The principle of the fastidious option is described in the figure below:

You might want to check the size distribution of clusters (number of amplicons in each cluster), and count the number of singletons (clusters containing only one amplicon). It can be easily done with the --statistics-file filename option. Each line in the output file represents a cluster and provides different metrics. See the manual for a complete description.

It is frequent for subsequent analyses to keep only one representative amplicon per cluster (usually the seed) to reduce the computational burden. That operation is easily done with swarm by using the -w filename option.

For each cluster, get the fasta sequences for all amplicons. Warning, this loop can generate a very large number of files. To limit the number of files, a test can be added to exclude swarms with less than n elements. See this wiki page for more examples.

INPUT_SWARM="amplicons.swarms"
INPUT_FASTA="amplicons.fasta"
OUTPUT_FOLDER="swarms_fasta"
AMPLICONS=$(mktemp)
mkdir "${OUTPUT_FOLDER}"
while read CLUSTER ; do
    tr " " "\n" <<< "${CLUSTER}" | sed -e 's/^/>/' > "${AMPLICONS}"
    seed=$(head -n 1 "${AMPLICONS}")
    grep -A 1 -F -f "${AMPLICONS}" "${INPUT_FASTA}" | \
        sed -e '/^--$/d' > "./${OUTPUT_FOLDER}/${seed/>/}.fasta"
done < "${INPUT_SWARM}"
rm "${AMPLICONS}"

To cite swarm, please refer to:

Swarm: robust and fast clustering method for amplicon-based studies.
Mahé F, Rognes T, Quince C, de Vargas C, Dunthorn M. (2014)
PeerJ 2:e593 doi: 10.7717/peerj.593

Swarm v2: highly-scalable and high-resolution amplicon clustering.
Mahé F, Rognes T, Quince C, de Vargas C, Dunthorn M. (2015)
PeerJ 3:e1420 doi: 10.7717/peerj.1420

Swarm v3: towards tera-scale amplicon clustering.
Mahé F, Czech L, Stamatakis A, Quince C, de Vargas C, Dunthorn M, Rognes T. (2021)
Bioinformatics doi: 10.1093/bioinformatics/btab493

Many thanks to the following people for their valuable contributions:

• Lucas Czech
• Etienne Platini

You are welcome to:

• submit suggestions and bug-reports at: https://github.com/torognes/swarm/issues
• send a pull request on: https://github.com/torognes/swarm/
• compose a friendly e-mail to: Frédéric Mahé [email protected] and Torbjørn Rognes [email protected]

swarm is available in third-party pipelines:

• FROGS: a Galaxy/CLI workflow designed to produce a cluster count matrix from high depth sequencing amplicon data.
• LotuS (v1.30): extremely fast cluster building, annotation, phylogeny and abundance matrix pipeline, based on raw sequencer output.
• QIIME (v1.9): a multi-purpose pipeline for performing microbiome analysis from raw DNA sequencing data.

If you want to try alternative free and open-source clustering methods, here are some links:

• vsearch
• Oligotyping
• DNAclust
• Sumaclust
• Crunchclust

swarm adheres to semantic versioning 2.0.0:

Given a version number MAJOR.MINOR.PATCH, increment the:

MAJOR version when you make incompatible API changes, MINOR version when you add functionality in a backwards compatible manner, and PATCH version when you make backwards compatible bug fixes.

swarm 3.1.6:

• use more C++11 and STL features,
• replace pthreads with std::thread,
• eliminate most of clang-tidy's warnings,
• refactor to reduce cyclomatic complexity (simpler and shorter functions),
• reduce/eliminate linuxisms to improve portability,
• measure the effect of code modernization on run-time performances

swarm 3.2.0:

• swarm can be compiled natively on a BSD or a Windows system

swarm 4.0.0:

• rename option -n to --no-cluster-breaking (API change),
• faster similarity search? (AVX2, GPU, symmetric deletion lookup algorithm, ...),
• drop compatibility with GCC 4 and GCC 5 (late-2024 GCC 8 will become the new de facto standard for HPC centers),
• start using C++14 and C++17 features

swarm 3.1.5 fixes four minor bugs, improves code and documentation, and eliminates compilation warnings and static analysis warnings:

• change: when using the fastidious and the ceiling options, minimal ceiling value is now 40 megabytes (instead of 8 megabytes),
• add: more compilation checks (shadow, useless-cast, conversion, sign-conversion),
• add: 42 new black-box tests,
• fix: two minor bugs introduced in version 3.1.1 (alloc-dealloc-mismatch, and allocating too much memory; bugs had no impact on clustering results),
• fix: a minor bug introduced in version 3.1.4 (misaligned memory under certain conditions; bug had no impact on clustering results),
• fix: minor bug in the way available memory is estimated (buffer underflow; bug had no impact on clustering results),
• fix: 50 warnings triggered by newly added compilation checks,
• fix: 1,677 clang-tidy warnings (from 2,629 warnings, down to 952),
• performance: generally stable but compiler-dependent, with the exception of a 5 to 10% increase in total memory footprint when d >= 2 (we expect these performance regressions to be temporary and to be fixed with further refactoring),
• improve: documentation for output option --network_file (advanced users),
• improve: build target platform detection,
• improve: code-coverage of our test-suite,
• improve: code modernization for long-term maintenance,
• revert: due to inconsistent test results, reading from named pipes (fifo) is marked as experimental for now

swarm 3.1.4 fixes a minor bug, eliminates compilation warnings and static analysis warnings, and improves code:

• fix: add checks to prevent silent overflow of short unsigned integers,
• fix: compilation warnings with GCC 13 and clang 18,
• fix: 1,040 clang-tidy warnings (from 3,669 warnings, down to 2,629),
• improve: code modernization for long-term maintenance,
• improve: double the maximal number of threads (from 256 threads to 512),
• improve: make -DNDEBUG the default compilation behavior,
• performance: stable for all modes, except a 6 to 10% increase in memory footprint when d > 2

swarm 3.1.3 fixes a few minor bugs, removes warnings, and improves code and documentation:

• fix: bug introduced in version 3.1.1, that caused swarm to allocate way too much memory when d > 1 (bug had no impact on clustering results),
• fix: off-by-one error when allocating memory for a Bloom filter (bug had no impact on clustering results),
• fix: compilation warning with GCC 12 (and more recent) when using link-time optimization,
• fix: compilation warning with clang 13 (and more recent): unused set variable,
• fix: five clang-tidy warnings (readability-braces-around-statements),
• fix: minor code refactoring,
• improve: more uniform vocabulary throughout swarm's documentation (code, help, manpage, README, companion scripts and wiki),
• improve: code coverage of our test suite (swarm-tests).

swarm 3.1.2 fixes a bug with fastidious mode introduced in version 3.1.1, that could cause Swarm to crash. Probably due to allocating too much memory.

swarm 3.1.1 eliminates a risk of segmentation fault with extremely long sequence headers. Documentation and error messages have been improved, and code cleaning continued.

swarm 3.1 includes a fix for a bug in the 16-bit SIMD alignment code that was exposed with a combination of d>1, long sequences, and very high gap penalties. The code has also been been cleaned up, tested and improved substantially, and it is now fully C++11 compliant. Support for macOS on Apple Silicon (ARM64) has been added.

swarm 3.0 is much faster when d = 1, and consumes less memory. Strict dereplication is now mandatory.

swarm 2.2.2 fixes a bug causing Swarm to wait forever in very rare cases when multiple threads were used.

swarm 2.2.1 fixes a memory allocation bug for d=1.

swarm 2.2.0 fixes several problems and improves usability. Corrected output to structure and uclust files when using fastidious mode. Corrected abundance output in some cases. Added check for duplicated sequences and fixed check for duplicated sequence IDs. Checks for empty sequences. Sorts sequences by additional fields to improve stability. Improves compatibility with compilers and operating systems. Outputs sequences in upper case. Allows 64-bit abundances. Shows message when waiting for input from stdin. Improves error messages and warnings. Improves checking of command line options. Fixes remaining errors reported by test suite. Updates documentation.

swarm 2.1.13 removes a bug in the progress bar when writing seeds.

swarm 2.1.12 removes a debugging message.

swarm 2.1.11 fixes two bugs related to the SIMD implementation of alignment that might result in incorrect alignments and scores. The bug only applies when d>1.

swarm 2.1.10 fixes two bugs related to gap penalties of alignments. The first bug may lead to wrong aligments and similarity percentages reported in UCLUST (.uc) files. The second bug makes Swarm use a slightly higher gap extension penalty than specified. The default gap extension penalty used have actually been 4.5 instead of 4.

swarm 2.1.9 fixes a problem when compiling with GCC version 6.

swarm 2.1.8 fixes a rare bug triggered when clustering extremely short undereplicated sequences. Also, alignment parameters are not shown when d=1.

swarm 2.1.7 fixes more problems with seed output. Ignore CR characters in FASTA files. Improved help and error messsages.

swarm 2.1.6 fixes problems with older compilers that do not have the x86intrin.h header file. It also fixes a bug in the output of seeds with the -w option when d>1.

swarm 2.1.5 fixes minor bugs.

swarm 2.1.4 fixes minor bugs in the swarm algorithm used for d=1.

swarm 2.1.3 adds checks of numeric option arguments.

swarm 2.1.2 adds the -a (--append-abundance) option to set a default abundance value to be used when abundance information is missing from the input file. If this option is not specified, missing abundance information will result in a fatal error. The error message in that case is improved.

swarm 2.1.1 fixes a bug with the fastidious option that caused it to ignore some connections between heavy and light swarms.

swarm 2.1.0 marks the first official release of swarm 2.

swarm 2.0.7 writes abundance information in usearch style when using options -w (--seeds) in combination with -z (--usearch-abundance).

swarm 2.0.6 fixes a minor bug.

swarm 2.0.5 improves the implementation of the fastidious option and adds options to control memory usage of the Bloom filter (-y and -c). In addition, an option (-w) allows to output cluster representatives sequences with updated abundances (sum of all abundances inside each cluster). This version also enables dereplication when d = 0.

swarm 2.0.4 includes a fully parallelized fastidious option.

swarm 2.0.3 includes a working fastidious option.

swarm 2.0.2 fixes SSSE3 problems.

swarm 2.0.1 is a development release that partially implements the fastidious option.

swarm 2.0.0 simplifies the usage of swarm by using the fast algorithm and the built-in cluster breaking by default. Some options are changed and some new output options are introduced.

swarm 1.2.21 is supposed to fix some problems related to the use of the SSSE3 CPU instructions which are not always available.

swarm 1.2.20 presents a production-ready version of the alternative algorithm (option -a), with optional built-in cluster breaking (option -n). That alternative algorithmic approach (usable only with d = 1) is considerably faster than currently used clustering algorithms, and can deal with datasets of 100 million unique amplicons or more in a few hours. Of course, results are rigourously identical to the results previously produced with swarm. That release also introduces new options to control swarm output (options -i and -l).

swarm 1.2.19 fixes a problem related to abundance information when the sequence identifier includes multiple underscore characters.

swarm 1.2.18 reenables the possibility of reading sequences from stdin if no file name is specified on the command line. It also fixes a bug related to CPU features detection.

swarm 1.2.17 fixes a memory allocation bug introduced in version 1.2.15.

swarm 1.2.16 fixes a bug in the abundance sort introduced in version 1.2.15.

swarm 1.2.15 sorts the input sequences in order of decreasing abundance unless they are detected to be sorted already. When using the alternative algorithm for d = 1 it also sorts all subseeds in order of decreasing abundance.

swarm 1.2.14 fixes a bug in the output with the swarm breaker option (-b) when using the alternative algorithm (-a).

swarm 1.2.13 updates the citation.

swarm 1.2.12 improves speed of new search strategy for d = 1.

swarm 1.2.11 corrects the number of differences reported in the break swarms output.

swarm 1.2.10 allows amplicon abundances to be specified using the usearch style in the sequence header (e.g. >id;size=1) when the -z option is chosen. Also fixes the bad URL shown in the previous version of swarm.

swarm 1.2.9 includes a parallelized variant of the new search strategy for d = 1. It seems to be fairly scalable up to about 16 threads for longer reads (~400bp), while up to about 8 threads for shorter reads (~150bp). Using about 50% more threads than available physical cores is recommended. This version also includes the d parameter in the beginning of the mothur-style output (e.g., swarm\_1). Also, in the break_swarms output the real number of differences between the seed and the amplicon is indicated in the last column.

swarm 1.2.8 fixes an error with the gap extension penalty. Previous versions effectively used a gap penalty twice as large as intended. This version also introduces an experimental new search strategy in the case where d = 1 that appears to be almost linear and faster at least for datasets of about half a million sequences or more. The new strategy can be turned on with the -a option.

swarm 1.2.7 incorporates a few small changes and improvements to make it ready for integration into QIIME.

swarm 1.2.6 add an option (-r or --mothur) to format the output file as a mothur-compatible list file instead of the native swarm format. When swarm encounters an illegal character in the input sequences it will now report the illegal character and the line number.

swarm 1.2.5 can be run on CPUs without the POPCNT feature. It automatically checks whether the CPU feature is available and uses the appropriate code. The code that avoids POPCNT is just slightly slower. Only basic SSE2 is now required.

swarm 1.2.4 changes the name of the new option from --break_swarms to --break-swarms for consistency with other options, and also adds a companion script swarm_breaker.py to refine swarm results (scripts folder).

swarm 1.2.3 adds an option (-b or --break_swarms) to output all pairs of amplicons to stderr. The data can be used for post-processing of the results to refine the swarms. The syntax of the inline assembly code is also changed for compatibility with more compilers.

swarm 1.2.2 fixes an issue with incorrect values in the statistics file (maximum generation and radius of swarms). This version is also a bit faster.

swarm 1.2.1 removes the need for a SSE4.1 capable CPU and should now be able to run on most servers, desktops and laptops.

swarm 1.2.0 introduces a pre-filtering of similar amplicons based on k-mers. This eliminates most of the time-consuming pairwise alignments and greatly improves speed. The speedup can be more than 100-fold compared to previous swarm versions when using a single thread with a large set of amplicons. Using multiple threads induces a computational overhead, but becomes more and more efficient as the size of the amplicon set increases.

swarm now works on Apple computers. This version also corrects an issue in the pairwise global alignment step that could lead to sub-optimal alignments. Slightly different alignments may result relative to previous version, giving slightly different swarms.

swarm 1.1.0 introduces new optimizations and is 20% faster than the previous version on our test dataset. It also introduces two new output options: statistics and uclust-like format.

By specifying the -s option to swarm it will now output detailed statistics about each swarm to a specified file. It will print the number of unique amplicons, the number of occurrences, the name of the seed and its abundance, the number of singletons (amplicons with an abundance of 1), the number of iterations and the maximum radius of the swarm (i.e. number of differences between the seed and the furthermost amplicon). When using input data sorted by decreasing abundance, the seed is the most abundant amplicon in the swarm.

Some pipelines use the uclust output format as input for subsequent analyses. swarm can now output results in this format to a specified file with the -u option.