AI for Bio · Benchmark 31

FoldBench

All-atom co-folding evaluation on low-homology PDB assemblies; the antibody-antigen task discriminates most.

bench-31Type BEvidence L3Structure & co-folding◐ Low-frequency Board中文

Overview

NameFoldBench · aliases: FoldBench, FoldBench-AB
Full nameFoldBench all-atom biomolecular structure prediction benchmark
Ambiguous names (not auto-linked)None
FamilyNot a family entry. One dataset with one maintainer board; the paper and repository are milestones of a single entry, not family members.
DomainStructure & co-folding
InclusionSupplementary entry
TypeB · Fixed retrospective set with a maintainer board
Evidence levelL3 · Time split with low-homology filtering, not blind
PublisherFudan University, Shanghai Jiao Tong University, The Chinese University of Hong Kong, Shanghai AI Laboratory, Lingang Laboratory (Academic lab)
First release · paper v1
Basis: bioRxiv v1 (10.1101/2025.05.22.655600); the repository creation date is not used as first release source
Other milestones
  • Preprint v1: · bioRxiv v1 source
  • Journal online: · Nature Communications paper published source
  • Code release: † · Official repository made public (MIT licence) source
Latest update · board entry
Basis: Results update on the repository board
Board updated with RF3 figures on 2025-12-31 and with Boltz-2 and the OpenFold3 preview on 2025-12-05; May 2026 commits were later reverted. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedThe maintainers have announced no new version or data expansion.
Status◐ Low-frequency · Last repository results update 2025-12-31; no newer official version seen by 2026-11-16
ScaleAbstract: 1,522 assemblies, 9 tasks; text and README task counts sum to 1,823 and 1,819.
  • Assemblies (abstract): 1,522 assemblies (paper)
  • Summed task counts (text / README): 1,823 / 1,819 (paper / repo)
  • Tasks: 9 tasks
  • Antibody-antigen interfaces: 172 targets (paper)
  • RNA / DNA monomers: 15 / 14 targets (paper)
  • PDB deposition window: 2023-01-13 – 2024-11-01
Wet-lab refreshStatic data · Fixed set of already deposited PDB structures; no wet lab and no planned expansion.
Own wet lab needed to evaluate a new modelNo, run it as a download
Models reported12 counted (board 7, self 3, modified protocol 2)
Unverified fields† release_milestones.code_release, ranking.snapshot_date, models_reported[2], models_reported[7]
As of2026-11-16

What it measures

  • Task: All-atom structure prediction on low-homology PDB assemblies deposited after the reference cutoff: three monomer tasks and six interface types, including antibody-antigen.
  • Truth: Experimental PDB structures deposited in the window, filtered for low homology to training data.
  • Primary metric: Interface success rate
  • Secondary metrics: Monomer lDDT; Joint pose and interface criteria for protein-ligand
  • Version traps: Two subsets coexist (post-2023-01 full set and post-2024-01); models with later training cutoffs are asterisked and must not be ranked against the rest; abstract and repo counts differ.

Current ranking

Board

MetricInterface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
Eval setFoldBench official repository board, full set (targets released after 2023-01)
Snapshot
Boardhttps://github.com/BEAM-Labs/FoldBench
Retrieved2026-11-16

Caveat. The repository board gives no explicit snapshot date, so README update dates are used; two subsets coexist and asterisked entries have later training cutoffs that the README says must not be ranked alongside the rest.

Shown in the overview table:

Protein-Protein 72.93% / Antibody-Antigen 47.90% / Protein-Ligand 64.90%; Protein-RNA 62.32%, Protein-DNA 79.18%, RNA Monomer 0.61, DNA Monomer 0.53
Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT · FoldBench official repo leaderboard, full set (targets released after 2023-01)
Board ·

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
1AlphaFold 3Protein-Protein 72.93% / Antibody-Antigen 47.90% / Protein-Ligand 64.90%; Protein-RNA 62.32%, Protein-DNA 79.18%, RNA Monomer 0.61, DNA Monomer 0.53Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
FoldBench official repo leaderboard, full set (targets released after 2023-01)
Board
source Board top-1 on every listed interface task; ordered here by the protein-protein column.
2OpenFold 3 (preview)69.96% / 28.83% / 44.49%; Protein-RNA 18.84%, Protein-DNA 5.88%Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
FoldBench official repo leaderboard, full set (targets released after 2023-01)
Board
source Ranked 2nd on protein-protein only; weakest on nucleic-acid interfaces.
3Chai-168.53% / 23.64% / 51.23%Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
FoldBench official repo leaderboard, full set (targets released after 2023-01)
Board
source
4Boltz-168.25% / 33.54% / 55.04%Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
FoldBench official repo leaderboard, full set (targets released after 2023-01)
Board
source The board lists Boltz-1, not the 1x variant.
5Protenix68.18% / 34.13% / 50.70%Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
FoldBench official repo leaderboard, full set (targets released after 2023-01)
Board
source Figures for HelixFold 3, not a site model, appear on the same board for comparison.
—RosettaFold3*Protein-Protein 72.44% / Antibody-Antigen 37.50% / Protein-Ligand 57.28% (post-2024-01 subset)Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
FoldBench official repo leaderboard, full set (targets released after 2023-01)
Board
source Asterisked: training cutoff later than the reference date, reference only, not directly rankable.
—Boltz-2*Protein-Protein 70.54% / Antibody-Antigen 25.00% / Protein-Ligand 53.90% (post-2024-01 subset)Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
FoldBench official repo leaderboard, full set (targets released after 2023-01)
Board
source Asterisked, reference only, not directly rankable.
—SeedFold74.03 / 53.21 / 63.12 (SeedFold-Linear 74.14 / 46.91 / 66.48)Interface success rate (% ; DockQ>=0.23, protein-ligand requires LRMSD<2 A and LDDT-PLI>0.8); monomers use LDDT
FoldBench tasks as reported in the model paper
Self-reported
source Author-reported and not on the official board; if protocols match it would exceed the board leader on protein-protein and antibody-antigen.

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 12. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Official leaderboard or results (7)

  • AlphaFold 3 Board maintainer leaderboard · Maintainer leaderboard (FoldBench repo) · evidence: link
  • Boltz-1 / Boltz-1x Board maintainer leaderboard · The maintainer board lists Boltz-1 · evidence: link
  • Boltz-2 Board maintainer run on post-2024-01 subset · Maintainer ran Boltz-2 on post-2024-01 subset (training cutoff later; reference only) †partly verified · evidence: link
  • Chai-1 Board maintainer leaderboard · Maintainer leaderboard · evidence: link
  • OpenFold3 Board maintainer leaderboard (preview) · Maintainer leaderboard (preview) · evidence: link
  • Protenix Board maintainer leaderboard · Maintainer leaderboard · evidence: link
  • RF3 Board maintainer leaderboard (2024-01+ subset) · Maintainer leaderboard (2024-01+ subset) · evidence: link

Developer report, official protocol (3)

  • ESMFold2 Self-reported FoldBench no-MSA complexes / antibody-antigen · Model-card figures are second-hand; another paper reports figures on its own antibody subset †partly verified · evidence: link
  • IntelliFold-2 Self-reported FoldBench antibody-antigen / protein-ligand · Published as a figure only; the predecessor reports numbers · evidence: link
  • SeedFold Self-reported FoldBench categories (protein-ligand, antibody-antigen, protein-protein, RNA) · Self-reported on FoldBench; arXiv 2512.24354 · evidence: link

Developer report, modified protocol (2)

  • Nesso-1 Self-reported · Added by R7/fc_zero as a missing candidate and confirmed in the review. Resolution: Technical report names the benchmark and copies baselines from another paper while admitting its filtering criteria differ, so it is self-reported on a modified protocol, not an official entry. Moves this model off the zero-pairing list. · evidence: link
  • OpenDDE Self-reported self-defined FoldBench antibody-antigen subset · Self-defined FoldBench-AB antibody-antigen subset; not comparable to official board · evidence: link

Caveats

  1. A retrospective set, not a blind test; the window starts early, so models with later training cutoffs may have seen some targets.
  2. RNA and DNA monomers have only 14-15 targets and protein-peptide 51, so small-sample rank gaps should not be read.
  3. Target counts differ across abstract, text and repository (1,522 vs 1,823 vs 1,819), so citations must state which.
  4. Self-reported entries differ in sampling settings and versions, so same-named models need not be comparable.
  5. It measures structural accuracy only, not function, affinity or developability.

How to use it

Run the repository scripts and emit the board format; models with later training cutoffs should report only the post-2024-01 subset and state their cutoff. Report subset, sampling settings, seeds and per-task target counts, and give both success rate and the DockQ distribution for antibody-antigen; generalisation claims need newer structures.

Pairs well with: PoseBusters · PLINDER / PINDER / Runs N' Poses · PoseX · CAPRI · CASP

Sources

  1. FoldBench paper (Nature Communications 2025) (journal)
  2. FoldBench preprint (bioRxiv v1) (preprint)
  3. FoldBench official repository and board (leaderboard)
  4. FoldBench dataset (Zenodo) (archive)
  5. Model paper reporting FoldBench results (arXiv 2512.24354) (preprint)
  6. Model paper reporting a custom antibody subset (arXiv 2607.03787) (preprint)

Corrections

2 corrections
FieldWasNowBasisFound
models_reported.model-nesso-1not listed (zero-pairing list)self-reported-modified-protocolsource2026-11-16
n_models_reported11recomputed under the new relation-type rules (established pairings only)source2026-11-16

Part of the AI for Bio benchmarks survey · all 44 benchmarks · open in the explorer. Record as of 2026-11-16.