🐍🟡♦️🟦 PyHMMER

Cython bindings and Python interface to HMMER3.

🗺️ Overview

HMMER is a biological sequence analysis tool that uses profile hidden Markov models to search for sequence homologs. HMMER3 is developed and maintained by the Eddy/Rivas Laboratory at Harvard University.

pyhmmer is a Python package, implemented using the Cython language, that provides bindings to HMMER3. It directly interacts with the HMMER internals, which has the following advantages over CLI wrappers (like hmmer-py):

• single dependency: If your software or your analysis pipeline is distributed as a Python package, you can add pyhmmer as a dependency to your project, and stop worrying about the HMMER binaries being properly setup on the end-user machine.
• no intermediate files: Everything happens in memory, in Python objects you have control on, making it easier to pass your inputs to HMMER without needing to write them to a temporary file. Output retrieval is also done in memory, via instances of the pyhmmer.plan7.TopHits class.
• no input formatting: The Easel object model is exposed in the pyhmmer.easel module, and you have the possibility to build a DigitalSequence object yourself to pass to the HMMER pipeline. This is useful if your sequences are already loaded in memory, for instance because you obtained them from another Python library (such as Pyrodigal or Biopython).
• no output formatting: HMMER3 is notorious for its numerous output files and its fixed-width tabular output, which is hard to parse (even Bio.SearchIO.HmmerIO is struggling on some sequences).
• efficient: Using pyhmmer to launch hmmsearch on sequences and HMMs in disk storage is typically as fast as directly using the hmmsearch binary (see the Benchmarks section). pyhmmer.hmmer.hmmsearch uses a different parallelisation strategy compared to the hmmsearch binary from HMMER, which can help getting the most of multiple CPUs when annotating smaller sequence databases.

This library is still a work-in-progress, and in an experimental stage, but it should already pack enough features to run biological analyses or workflows involving hmmsearch, hmmscan, nhmmer, phmmer, hmmbuild and hmmalign.

🔧 Installing

pyhmmer can be installed from PyPI, which hosts some pre-built CPython wheels for x86-64 Linux, as well as the code required to compile from source with Cython:

$ pip install pyhmmer

Compilation for UNIX PowerPC is not tested in CI, but should work out of the box. Other architectures (e.g. Arm) and OSes (e.g. Windows) are not supported by HMMER.

A Bioconda package is also available:

$ conda install -c bioconda pyhmmer

📖 Documentation

A complete API reference can be found in the online documentation, or directly from the command line using pydoc:

$ pydoc pyhmmer.easel
$ pydoc pyhmmer.plan7

💡 Example

Use pyhmmer to run hmmsearch to search for Type 2 PKS domains (t2pks.hmm) inside proteins extracted from the genome of Anaerococcus provencensis (938293.PRJEB85.HG003687.faa). This will produce an iterable over TopHits that can be used for further sorting/querying in Python. Processing happens in parallel using Python threads, and a TopHits object is yielded for every HMM passed in the input iterable.

import pyhmmer

with pyhmmer.easel.SequenceFile("pyhmmer/tests/data/seqs/938293.PRJEB85.HG003687.faa", digital=True) as seq_file:
    sequences = list(seq_file)

with pyhmmer.plan7.HMMFile("pyhmmer/tests/data/hmms/txt/t2pks.hmm") as hmm_file:
    for hits in pyhmmer.hmmsearch(hmm_file, sequences, cpus=4):
      print(f"HMM {hits.query_name.decode()} found {len(hits)} hits in the target sequences")

Have a look at more in-depth examples such as building a HMM from an alignment, analysing the active site of a hit, or fetching marker genes from a genome in the Examples page of the online documentation.

💭 Feedback

⚠️ Issue Tracker

Found a bug ? Have an enhancement request ? Head over to the GitHub issue tracker if you need to report or ask something. If you are filing in on a bug, please include as much information as you can about the issue, and try to recreate the same bug in a simple, easily reproducible situation.

🏗️ Contributing

Contributions are more than welcome! See CONTRIBUTING.md for more details.

⏱️ Benchmarks

Benchmarks were run on a i7-10710U CPU running @1.10GHz with 6 physical / 12 logical cores, using a FASTA file containing 4,489 protein sequences extracted from the genome of Escherichia coli (562.PRJEB4685) and the version 33.1 of the Pfam HMM library containing 18,259 domains. Commands were run 3 times on a warm SSD. Plain lines show the times for pressed HMMs, and dashed-lines the times for HMMs in text format.

Raw numbers can be found in the benches folder. They suggest that phmmer should be run with the number of logical cores, while hmmsearch should be run with the number of physical cores (or less). A possible explanation for this observation would be that HMMER platform-specific code requires too many SIMD registers per thread to benefit from simultaneous multi-threading.

To read more about how PyHMMER achieves better parallelism than HMMER for many-to-many searches, have a look at the Performance page of the documentation.

🔍 See Also

Building a HMM from scratch? Then you may be interested in the pyfamsa package, providing bindings to FAMSA, a very fast multiple sequence aligner. In addition, you may want to trim alignments: in that case, consider pytrimal, which wraps trimAl 2.0.

If despite of all the advantages listed earlier, you would rather use HMMER through its CLI, this package will not be of great help. You can instead check the hmmer-py package developed by Danilo Horta at the EMBL-EBI.

⚖️ License

This library is provided under the MIT License. The HMMER3 and Easel code is available under the BSD 3-clause license. See vendor/hmmer/LICENSE and vendor/easel/LICENSE for more information.

This project is in no way affiliated, sponsored, or otherwise endorsed by the original HMMER authors. It was developed by Martin Larralde during his PhD project at the European Molecular Biology Laboratory in the Zeller team.