About DPCexplorer

An interactive platform for exploring DPCfam and DPCstruct protein domain classifications.

The Annotation Gap

Every organism, from a simple bacterium to a human being, depends on proteins to carry out the chemical reactions that make life possible. Understanding what a protein does, in other words, knowing its function, is therefore one of the most important goals in modern biology. The gold standard for assigning function is direct experimental characterisation. That method is reliable, but it is also slow and expensive, and it cannot keep pace with the speed of discovery.

High-throughput sequencing has deposited millions of new sequences every year. UniProtKB now holds over 245 million sequences, yet only a tiny fraction carry experimentally verified annotations. The launch of AlphaFold2 made things even more striking: reliable 3D structure predictions now exist for over 200 million proteins. The data is there. The annotations are not.

Domain Family Classification

One practical shortcut is protein domain family classification. The idea is simple: proteins (or protein domains) that belong to the same family typically share the same evolutionary origin and, usually, a similar function. If one member is well studied, that knowledge can be carefully extended to the rest. This is the logic behind Pfam, the most widely used protein family database. Pfam builds families from Multiple Sequence Alignments and profile Hidden Markov Models, which lets it detect distant relatives even when raw sequence similarity is low.

This approach works well. Expert curation improves annotation quality, and recent deep learning additions push coverage further still. Pfam-N, for instance, uses transformer-based models and convolutional neural networks to detect distant homologs that classical methods had missed, successfully increasing UniProtKB coverage by 8.8%. But there is a fundamental limit: quality does not scale with quantity. Expert curation is inherently limited by human bandwidth. Deep learning models, however powerful, are bounded by their training data. A genuinely novel family, one with no known relatives, generally stays invisible to both.

A fully unsupervised alternative: DPCfam and DPCstruct

DPCfam and DPCstruct take a different approach. Instead of starting from curated families or training on known annotations, they start from raw pairwise similarity between proteins and let the data organise itself.

For sequences, this means BLASTp all-versus-all comparisons across 23 million UniRef50 sequences (UniRef50 v. 2017_07), producing roughly 55 billion alignments. For structures, the same idea is applied to 15 million AlphaFold2-predicted structures (the Foldseek Clusters representative set of the AlphaFold Database, AlphaFoldDB v4.0) using Foldseek and MMseqs2, which translate 3D shapes into a structural alphabet for large-scale comparison.

The clustering itself is handled by the Density Peak Clustering (DPC) algorithm (Rodriguez & Laio, 2014). DPC identifies cluster centres as points with high local density and high isolation from other dense regions, no prior assumption about the number or shape of clusters needed. A two-scale design first segments individual proteins into domain-like units (primary clusters), then groups those units globally into evolutionary families called metaclusters.

Because no annotation database was consulted during clustering, the resulting metaclusters are fully independent hypotheses about evolutionary relationships. To ensure that the outputs of DPCfam and DPCstruct were biologically meaningful, these outputs were compared against expert-curated databases, including Pfam, ECOD, SCOP, and CATH.

On one side, DPCfam was found to recover approximately 81% of medium-to-large Pfam families and 72% of ECOD families, demonstrating a high degree of correspondence with established sequence classifications. More importantly, the pipeline identified 14,190 metaclusters in the standard subset and 18,989 in the DPCfamB subset that completely lack Pfam mappings, representing putative novel protein families.

On the other side, DPCstruct was also compared with Pfam at the clan level to evaluate its consistency with sequence-based annotations; the analysis showed that 70% of the 14,423 metaclusters with available Pfam labels achieved a perfect consistency score of 1, meaning all members shared the same Pfam annotation. Additionally, DPCstruct successfully recovered 91% of SCOP folds and 83% of CATH folds, confirming its effectiveness in identifying well-characterized structural domains. Furthermore, 24% of DPCstruct metaclusters show no significant similarity to any known protein families or folds, pointing toward a massive pool of potential novel structural folds. The average intra-cluster sequence identity is only 34.5 %, placing most metaclusters well within the twilight zone where standard sequence-based tools become unreliable. Structure-based clustering reveals evolutionary relationships that sequence alone cannot detect.

In conclusion, a significant portion of metaclusters in both datasets carry no known Pfam annotation: specifically, 47,002 metaclusters (33,179 in DPCfam and 13,823 in DPCstruct). These are labeled UNKNOWN in DPCexplorer. They are not errors or noise; rather, they are structurally and sequentially coherent protein families that simply have not been named yet. Some may represent novel folds, while others may be ancient families overlooked by curation-based approaches. The evidence for their biological relevance is concrete: 63 DPCfam UNKNOWN metaclusters were adopted as official new entries in Pfam release 35.0 (e.g., MC202620 → PF20147, MC15137 → PF20146).

What DPCexplorer does

Both datasets were published on Zenodo in line with FAIR data principles. However, exploring them directly required downloading and parsing gigabyte-scale archives, including an 81 GB XML file for DPCfam. This presents a significant barrier for most researchers.

DPCexplorer removes that barrier. The raw archives were cleaned and preprocessed on the ORFEO HPC cluster, then loaded into a PostgreSQL database optimized for fast searches across millions of rows. A Django 6.0.1 backend serves this data through a responsive web interface, with no login required.

Each metacluster detail page is organized into four tabs. Summary displays key biological properties and Pfam annotations. Sequences lists the seed domains with UniProt links, paginated ten rows at a time. 3D Viewer: for DPCstruct, it renders AlphaFold2-predicted structures directly in the browser via PDBe-Mol*, colored by per-residue pLDDT confidence; for DPCfam, it links to the representative entry in AlphaFoldDB where available. Downloads provides one-click access to FASTA seed sequences, MSA profiles, and HMM models for DPCfam, as well as representative FASTA sequences and zipped PDB files for DPCstruct.

All preprocessed files used to build this database are openly available on Zenodo (DOI: 10.5281/zenodo.20159208), so any researcher can reproduce a running instance of DPCexplorer without re-running the heavy HPC preprocessing steps.

REST API

Alongside the web interface, DPCexplorer now exposes a public, read-only REST API under /api/. It lets you send a DPCfam or DPCstruct metacluster ID, or a comma-separated list of IDs, and get back its properties and members (protein ID and coordinate range) as JSON, without downloading any archive.

No login or API key is required. The full endpoint reference, error behaviour, and a ready-to-run Python script and Jupyter notebook are documented in DPCexplorer_API_Documentation.md on GitHub.

Credits & Funding

References

If you use DPCexplorer, please cite the software (below). Should you want to learn how it was built, see the thesis; for a concise summary, see the eScience 2026 paper. Both appear under "Project documentation".

1. Nyandu Kagarabi, E., Saadat, E., & Piomponi, V. (2026). DPCexplorer: A Django Web Application for Interactive Exploration of DPCfam and DPCstruct Protein Domain Classifications [Software]. Zenodo. 10.5281/zenodo.20575268 (concept DOI, always the latest release).

You do not need to cite these; they are here if you want to dig deeper. To learn how DPCexplorer was built, see the MDMC thesis (full documentation). For a concise summary, see the eScience 2026 paper.

2. Russo ET, Barone F, Bateman A, Cozzini S, Punta M, Laio A. DPCfam: Unsupervised protein family classification by density peak clustering of large sequence datasets. PLOS Computational Biology, 18(10), e1010610, 2022. DOI: 10.1371/journal.pcbi.1010610
3. Russo, E. T., & Barone, F. (2022). Metaclusters by DPCfam clustering of UniRef50 v 2017_07 [Data set]. Zenodo. DOI: 10.5281/zenodo.6900559

4. Barone F, Laio A, Punta M, Cozzini S, Ansuini A, Cazzaniga A. Unsupervised domain classification of AlphaFold2-predicted protein structures. PRX Life, 3(2), 023009, 2025. DOI: 10.1103/PRXLife.3.023009
5. Barone, F., Laio, A., Punta, M., Cozzini, S., Ansuini, A., & Cazzaniga, A. (2024). DPCstruct classification of AlphaFold2-predicted protein structures [Data set]. Zenodo. DOI: 10.5281/zenodo.13334296

6. Rodriguez A, Laio A. Clustering by fast search and find of density peaks. Science, 344(6191), 1492–1496, 2014. DOI: 10.1126/science.1242072

Source code: github.com/RitAreaSciencePark/dpc_fam_and_struct_webapp  (MIT License).