The Pfam protein families database: embracing AI/ML.
Paysan-Lafosse T, Andreeva A, Blum M, Chuguransky SR, Grego T, Pinto BL, Salazar GA, Bileschi ML, Llinares-López F, Meng-Papaxanthos L, Colwell LJ, Grishin NV, Schaeffer RD, Clementel D, Tosatto SCE, Sonnhammer E, Wood V, Bateman A. · 2025
Read The Pfam protein families database: embracing AI/ML. by Paysan-Lafosse T, Andreeva A, Blum M, Chuguransky SR, Grego T, Pinto BL, Salazar GA, Bileschi ML, Llinares-López F, Meng-Papaxanthos L, Colwell LJ, Grishin NV, Schaeffer RD, Clementel D, Tosatto SCE, Sonnhammer E, Wood V, Bateman A. (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc11701544.txt.
Introduction
Paysan-Lafosse and a large team update the Pfam protein families database in this 2025 paper. Pfam groups proteins into families and domains, the reusable parts of proteins, to aid genome annotation and protein study. Since 2020 the old Pfam website was shut and the resource folded into InterPro, a larger hub.
Curation grew for families from metagenomes, meaning DNA read straight from the environment, and for microproteins and repeat-rich families. The scheme was harmonised with the ECOD structural classification. AlphaFold structure predictions now help refine domain boundaries and spot new domains. Large-scale scans of AlphaFold models found new families. A deep learning add-on, Pfam-N, widened coverage, lifting coverage of the UniProtKB protein archive by 8.8 percent. More frequent releases are planned.
Catalog record
| CREATOR | Paysan-Lafosse T, Andreeva A, Blum M, Chuguransky SR, Grego T, Pinto BL, Salazar GA, Bileschi ML, Llinares-López F, Meng-Papaxanthos L, Colwell LJ, Grishin NV, Schaeffer RD, Clementel D, Tosatto SCE, Sonnhammer E, Wood V, Bateman A. |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The Pfam protein families database: embracing AI/ML By Typhaine Paysan-Lafosse, Antonina Andreeva, Matthias Blum, Sara Rocio Chuguransky, Tiago Grego, Beatriz Lazaro Pinto, Gustavo A Salazar, Maxwell L Bileschi, Felipe Llinares-López, Laetitia Meng-Papaxanthos, Lucy J Colwell, Nick V Grishin Published in: Nucleic Acids Research DOI: 10.1093/nar/gkae997 ABSTRACT Abstract The Pfam protein families database is a comprehensive collection of protein domains and families used for genome annotation and protein structure and function analysis (https://www.ebi.ac.uk/interpro/). This update describes major developments in Pfam since 2020, including decommissioning the Pfam website and integration with InterPro, harmonization with the ECOD structural classification, and expanded curation of metagenomic, microprotein and repeat-containing families. We highlight how AlphaFold structure predictions are being leveraged to refine domain boundaries and identify new domains. New families discovered through large-scale sequence similarity analysis of AlphaFold models are described. We also detail the development of Pfam-N, which uses deep learning to expand family coverage, achieving an 8.8% increase in UniProtKB coverage compared to standard Pfam. We discuss plans for more frequent Pfam releases integrated with InterPro and the potential for artificial intelligence to further assist curation. Despite recent advances, many protein families remain to be classified, and Pfam continues working toward comprehensive coverage of the protein universe. INTRODUCTION Pfam is a comprehensive database of protein families and domains that is widely used for the analysis of novel genomes and metagenomes. It serves as a crucial tool for guiding experimental research on specific proteins and systems. Each family entry within the Pfam database contains a seed alignment containing a representative set of sequences, from which a profile hidden Markov model (HMM) is generated. This HMM is then used to search against a sequence database called pfamseq (containing sequences from the UniProtKB reference proteomes [RPs]) using the HMMER software (http://hmmer.org/) (1). Sequence regions that meet a family-specific curated threshold, known as the gathering threshold, are aligned to the profile HMM to create a full alignment. Pfam entries are annotated by our curators with functional information sourced from the literature, where available. Although the underlying sequence database used to build Pfam alignments is based on the UniProtKB RP, all the profile HMMs are also searched against the whole of UniProtKB, and the resulting matches are made available on the InterPro website (https://www.ebi.ac.uk/interpro) and in flat file format on the Pfam ftp (https://ftp.ebi.ac.uk/pub/databases/Pfam/releases). When a Pfam entry is created or updated, we search it against pfamseq, trying to identify more distant homologs. Pfam entries are designed to avoid overlaps, ensuring that the same region of a sequence does not match more than one family. This non-overlap rule is an important quality control measure, helping to prevent the inclusion of false positive matches within a family. However, since Pfam 28.0, we relaxed this rule to allow small overlaps between families, as resolving all overlaps during each Pfamseq update has become increasingly time-consuming (2). There are many situations, such as incorrect gene predictions due to skipped exons or fusions of adjacent genes that cause overlapping of models, that are extremely difficult to resolve. Sets of Pfam entries that are believed to be evolutionarily related are grouped into clans (3). Relationships between entries are identified through sequence similarity, structural similarity, functional similarity, and/or profile-profile comparisons using software such as HHsearch (4). Where possible, a single, comprehensive profile HMM is built to detect all members of a family. For some of the larger superfamilies, where this is not feasible, multiple-profile HMMs are built and placed within the same clan. Since families within a clan are evolutionarily related, they are allowed to overlap with other members of the same clan. The development of advanced structure prediction methods has profoundly impacted the construction of Pfam entries. High-accuracy structure predictions, particularly those generated by tools such as AlphaFold (5), have significantly enhanced our ability to identify and characterize protein families. These predictions provide crucial insights into the 3D conformations of proteins, which in turn inform the delineation of evolutionary relationships and functional annotations. By incorporating structural data, we can more accurately define domain boundaries, detect distant homologs and resolve ambiguities that arise from sequence-based analyses alone. Here, we describe the work undertaken since Pfam release 33.1, including the decommissioning of the Pfam website, our collaboration with the ECOD database (6), the utilization of structure predictions to construct new Pfam entries and enhance existing ones, and some of the family updates implemented in Pfam release 37.0. We also detail the development of Pfam-N, which uses deep learning to expand the coverage of Pfam families. Decommission of the Pfam website In 2022, we took the decision to retire the Pfam website. The Pfam website codebase was first released over 20 years ago, and although it has been updated from time to time, some of its core functionality still dates to its origins. The Perl codebase represented a significant technical debt, and it was becoming progressively harder to maintain. Additionally, the most time-consuming part of the Pfam release pipeline was the generation of data exclusively related to the website rather than the core data of Pfam, i.e. its alignments and models. To address this issue, we decided to use the InterPro website as the primary way to view Pfam data. This would mean that European Molecular Biology Laboratory, European Bioinformatics Institute (EMBL-EBI), hosted a single website for protein family-related information rather than two. The InterPro website (https://www.ebi.ac.uk/interpro) has been redesigned using up-to-date technologies, including a modern framework (React) (7), and Pfam data and different viewing features are now available on the InterPro website. The decommission was carried out over the course of a year. The first step consisted of expanding the InterPro website to incorporate several features from the Pfam website that were lacking, including a compact protein domain view with a set of non-overlapping representative domains for each protein. This work is detailed in the paper describing InterPro in this issue (Blum et al., 2025). The second phase consisted of announcing the decommission to encourage Pfam users to start their transition to the InterPro website. Multiple announcements were made on the Pfam website, Pfam blog and social media platforms. We also ran a webinar highlighting where to find Pfam data on the InterPro website (Figure 1) to allow a smooth transition of Pfam users toward using the InterPro website. [Figure: Example of a Pfam family page for the Ras domain (Pfam: PF00071) on the Pfam legacy website (A) and on the InterPro website (B).] With the decommission of the Pfam website, we have also slightly changed the MySQL database schema, as some tables were only used by the Pfam website. The change has been made from Pfam release 36.0 onward. Although we tried to carefully explain our approach to users, there has been a persistent misunderstanding that because the Pfam website was being shut down, the Pfam database itself was being terminated. As Mark Twain said after erroneous press reports of his death, ‘The report of my death was an exaggeration’. We hope that this paper will help to quell the exaggerated rumors of our demise. Update of the Pfam documentation and online training materials With the decommissioning of the Pfam website, we have undertaken an extensive update of the Pfam documentation (https://pfam-docs.readthedocs.io/en/latest/index.html) to assist users in