henhaobio
Sign in

bioRxivBioinformaticsPreprint

BFVD v3-UniProt-complete, improved viral protein structure predictions

Seongeun KimiD, Olga Pimenova, Eli Levy KariniD, Milot MirditaiD, Martin SteineggeriD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

The Big Fantastic Virus Database (BFVD) v3 is the most comprehensive resource for ColabFold-AlphaFold2-predicted viral protein structures. It holds 5,776,417 structures of nearly all viral sequences in UniProt 2025_03, a 16.4-fold increase compared to the representative-only catalogs of BFVD v1 and v2. Structure prediction quality has also improved, with high-confidence predictions accounting for 75.3% of BFVD v3 entries. This is due to Logan's enormous sequence assembly, now mined for the entire BFVD v3 instead of only for shallow alignments, alongside continued prediction quality improvements. BFVD v3 covers 72.7% of ICTV's virus species, 1.54-fold more than BFVD v2. With this version, the fraction of fully covered viral reference proteomes increases from 1.5% to 72.6%. BFVD v3 thus brings us closer to a structural catalog of the known virosphere and is freely available at bfvd.steineggerlab.workers.dev and bfvd.foldseek.com.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 1 (18 Sept 2026) · CC BY 4.0

5,776,417 ColabFold-AlphaFold2 structures cover 98.9% of viral UniProt 2025_03 sequences, 16.4-fold more than earlier versions, with 75.3% high-confidence and 72.6% of viral reference proteomes fully covered.

Earlier BFVD releases modelled only viral UniRef30 cluster representatives, which hid structural variation within clusters, omitted viral sequences whose representative was non-viral and skewed taxonomy (62.3% of v2 structures in kingdom Heunggongvirae vs 29.33% of UniProt 2023_02). Version 3 predicts nearly every viral UniProt 2025_03 sequence individually, adds homologs from Logan50 (a 50%-identity clustering of the Logan Sequence Read Archive assembly) for all entries, and drops early stopping.

Key results

  1. ShownVersion 3 models almost all viral UniProt 2025_03 proteins and fully covers most viral reference proteomes; v2 fully covered very few.

    5,776,417 structures (98.9% of 5,841,128 viral sequences; sequences over 2,000 residues excluded) vs 351,242 in v2, a 16.4-fold increase. Viral reference proteomes fully covered by exact sequence matches: 72.6% (v3) vs 1.5% (v2). Proteomes with no coverage: 8.8% vs 47.6%. · Abstract; Acquisition of sequences and MSA generation; Proteome and taxonomic coverage; Scale and taxonomic coverage of BFVD v3

  2. ShownCoverage of International Committee on Taxonomy of Viruses (ICTV) taxa rises at every rank below kingdom, and taxonomic composition now follows UniProt instead of cluster representatives.

    72.7% of ICTV's 17,554 species vs 47.1% in v2 (1.54-fold). Distinct species-level NCBI taxids: 44,124 vs 20,847 (2.12-fold). Realms (10/10) and kingdoms (11/11) were already fully covered by v2. Taxids were mapped to ICTV species through GenBank-linked taxids, then by exact name matches. · Proteome and taxonomic coverage; Scale and taxonomic coverage of BFVD v3

  3. ShownA larger share of v3 predictions is high-confidence, and the gain at the top of the confidence range remains after sequence redundancy is removed.

    Predicted Local Distance Difference Test (pLDDT) ≥ 70 for 75.3% of entries (4,351,255) vs 62.1% in v2. After 30%-identity clustering, v3 representatives track v2 up to pLDDT about 75 and exceed it above that. 40,524 entries (0.7%) use ESMFold+ProteinTTTMSA models in place of low-confidence AlphaFold2 ones. · Abstract; Structure prediction; Sequence and structural clustering; Prediction confidence and Logan50 homologs

  4. ShownSearching Logan50 for every entry deepens alignments across the database, including for proteins that previously had few homologs.

    Before mining, 14.1% of entries (811,890) had shallow base alignments (<30 homologs). Logan50 raised their median homolog count from 8 to 15 and moved 24.8% (200,967) above that threshold. Median effective sequence number (Neff) rose from 4.1 to 4.7 overall (1.3 to 1.8 for shallow entries); 17.7% of entries crossed Neff=6. · Acquisition of sequences and MSA generation; Prediction confidence and Logan50 homologs

  5. Shownv3 nearly contains v2's sequence space, roughly doubles it, and adds many new structural clusters, though most of the new structural clusters are singletons.

    Joint 30%-identity sequence clustering: 47.0% (308,281) of clusters v3-unique, 51.6% (338,622) shared, 1.4% (8,924) v2-unique. Foldseek structural clustering: 61.5% (450,113) v3-unique, 24.2% shared, 14.2% (104,204) v2-unique. Singletons make up 80.8% of v3-unique and 99.5% of v2-unique structural clusters. Overall: 1.9-fold more sequence variance, 2.2-fold more structural variance. · Sequence and structural clustering; Growth in sequence and structural diversity

Methods, data and reproducibility

ColabFold-AlphaFold2 (v1.6.0, model 4, three recycles, no templates or early stopping) predicted 5,776,417 viral UniProt 2025_03 sequences of at most 2,000 residues from MMseqs2 alignments extended with Logan50 homologs; 40,524 low-confidence entries were replaced by ESMFold+ProteinTTTMSA models. Coverage was scored by exact matches to viral reference proteomes and taxid mapping to ICTV species.

  • Viral UniProt 2025_03 sequences (descendants of taxid 10239), filtered to at most 2,000 residues · viruses · n = 5,776,417 of 5,841,128
  • Logan50: 50% sequence identity clustering of the Logan Sequence Read Archive assembly, human-derived sequences excluded · n = 3,009,464,616 sequences
  • UniProt viral reference proteomes (release 2026_02), canonical sequences · viruses · n = over 14,000 proteomes
  • ICTV Virus Metadata Resource species list · viruses · n = 17,554 species
  • BFVD v2 (comparison baseline) · viruses · n = 351,242 structures
  • ColabFold reference databases uniref30_2302_db and colabfold_envdb_202108_db for base alignments
  • Small benchmark subset of experimentally resolved structures not seen in AlphaFold2 training, used to choose among four prediction settings
Methods & tools
ColabFold-AlphaFold2 1.6.0, MMseqs2 18.8cc5c, colabfold_search, ESMFold+ProteinTTTMSA, hhmake 3.3.0, Foldseek 10.941cd33, Mol*, PULCHRA, TM-align, Logan50, UniRef30, ICTV Virus Metadata Resource
Software & environment
Versions stated: ColabFold-AF2 1.6.0, MMseqs2 commit 6f4523 (colabfold_search) and 18.8cc5c, Foldseek 10.941cd33, hhmake 3.3.0. No container or requirements file is described.
Reproducibility gaps
Only the webserver code is linked; the prediction and analysis pipeline scripts are not. The benchmark of prediction settings and the ESMFold+ProteinTTTMSA selection criteria are in Supplementary Results, not in this text.
Limitations (5)
  • BFVD v3 inherits UniProt's uneven sampling of the virosphere; Influenza A virus alone accounts for 13.3% of viral sequences in UniProt 2025_03. · stated by the authors
  • 27% of ICTV species still have no structure in BFVD. · stated by the authors
  • Sequences longer than 2,000 residues (1.1% of viral UniProt) were left out to limit compute; including them could have added 1.3% to ICTV species coverage. · stated by the authors
  • The confidence gains are attributed to dropping early stopping and to Logan50 mining, but the main text shows no ablation separating the two, and accuracy against experimental structures is reported only in Supplementary Results. · noted in this summary
  • Quality is judged mainly by the model's own pLDDT; 80.8% of v3-unique structural clusters are singletons, which the authors note is weaker evidence of structural validity. · noted in this summary

Read this if you need predicted structures for specific viral proteins or whole viral proteomes, or want a Foldseek-searchable viral structure database for remote homology or comparative analyses across viral lineages.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISKim '24Litvin '25Yang '26Odai '25Han '26Labra '26
84% semantic matchBioinformatics
Rachel Seongeun Kim, Eli Levy Karin, Milot Mirdita+2 more·Nucleic Acids Research·Oct 28, 2024· 86 citations