AI for Bio · Benchmark 05

PoseBusters

Protein-ligand pose benchmark: an RMSD threshold plus 18 physical and chemical validity checks.

bench-05Type C+BEvidence L3Structure & co-folding● Active Not comparable中文

Overview

NamePoseBusters · aliases: PoseBusters, PoseBusters V1, PoseBusters V2, PoseBusters Benchmark set
Full nameThe name is not an abbreviation, so there is no expanded form.
Ambiguous names (not auto-linked)PB-valid
FamilyNot a family entry. V1 and V2 are two sets of one tool: versions, not family members.
DomainStructure & co-folding
InclusionOriginal 30
TypeC B · Validity-checker tool plus two frozen static complex sets
Evidence levelL3 · Retrospective with a time split, not blind
PublisherUniversity of Oxford, OPIG group (Academic lab)
First release · paper v1
Basis: Preprint arXiv 2308.05777 v1; the first public package version was uploaded to PyPI two days later. source
Other milestones
  • Preprint v1: · arXiv 2308.05777 v1 source
  • Code release: · First public PyPI version, 0.2.2 source
Latest update · maintenance only
Basis: Commit history on the official repository's main branch
Unreleased commits on main (an HTML report, plus a version-bump commit) with no tag or release object; the last published version is still 0.6.5 and both sets are unchanged. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedThe maintainers have not announced a new set version or a new evaluation round.
Status● Active · The checker is widely used; the sets are frozen since V2 and near saturation
ScaleV1 has 428 complexes, V2 has 308, with an 85-complex easy control set; 18 checks.
  • Complexes: 428 (V1)
  • Complexes, crystal-contact ligands removed: 308 (V2)
  • Complexes in the easy control set: 85 (Astex Diverse set)
  • Validity checks: 6 chemical, 6 intramolecular, 6 intermolecular: 18
Wet-lab refreshStatic data · No new experiments; reuses public PDB complexes released after 2021.
Own wet lab needed to evaluate a new modelYes, to claim real generalization
Models reported8 counted (self 4, modified protocol 4, cited only 3)
Unverified fields† models_reported[3], models_reported[4], models_reported[6], models_reported[7], ranking.entries[0], ranking.entries[3]
As of2026-11-16

What it measures

  • Task: Predict the ligand pose for a protein-ligand pair, judging both geometric closeness to the crystal and physical and chemical plausibility.
  • Truth: Public PDB crystal structures, limited to post-2021 releases outside the common training set.
  • Primary metric: Share with RMSD under 2 Å and all checks passed
  • Secondary metrics: Success rate on RMSD under 2 Å alone; Pass rates of the individual checks
  • Version traps: Any number must state V1 or V2, whether validity checks are included, and the sampling rule (best@N or top confidence).

Current ranking

Not comparable

MetricShare with RMSD under 2 Å and checks passed; some sources give RMSD only
Eval setMixed: PoseBusters V2 (308) vs Pearl post-2021-10-01 subset (297); best@5 vs top-confidence
Snapshot stale
BoardNo official board
Retrieved2026-11-16

Caveat. No official leaderboard; all numbers come from papers with different subsets, sampling rules and training cutoffs, and some entries overlap the training window. One model was excluded for its cutoff and two others were not re-verified.

No headline entry: There is no official board and the sources are not comparable, so no headline number is given.

The entries below use different protocols, so they are listed, not ranked, and the overview table shows no score.

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
—PearlWithheld: Not comparable: subsets and sampling rules differ, so the ranking row shows no score.Share with RMSD under 2 Å and checks passed; some sources give RMSD only
Mixed: PoseBusters V2 (308) vs Pearl post-2021-10-01 subset (297); best@5 vs top-confidence
Self-reported
source Self-reported on a post-2021-10-01 subset; no independent reproduction found
—Boltz-1xWithheld: Not comparable: subsets and sampling rules differ, so the ranking row shows no score.Share with RMSD under 2 Å and checks passed; some sources give RMSD only
Mixed: PoseBusters V2 (308) vs Pearl post-2021-10-01 subset (297); best@5 vs top-confidence
Third-party
source Run by the team behind another model's paper, not self-reported
—Chai-1Withheld: Not comparable: subsets and sampling rules differ, so the ranking row shows no score.Share with RMSD under 2 Å and checks passed; some sources give RMSD only
Mixed: PoseBusters V2 (308) vs Pearl post-2021-10-01 subset (297); best@5 vs top-confidence
Third-party
source Run by the team behind another model's paper, not self-reported
—AlphaFold 3Withheld: Not comparable: subsets and sampling rules differ, so the ranking row shows no score.Share with RMSD under 2 Å and checks passed; some sources give RMSD only
Mixed: PoseBusters V2 (308) vs Pearl post-2021-10-01 subset (297); best@5 vs top-confidence
Third-party
source Pose-selection rule differs from best@5, and the training cutoff overlaps part of the set
—NeuralPLexer3Withheld: Not comparable: subsets and sampling rules differ, so the ranking row shows no score.Share with RMSD under 2 Å and checks passed; some sources give RMSD only
Mixed: PoseBusters V2 (308) vs Pearl post-2021-10-01 subset (297); best@5 vs top-confidence
Self-reported
source Self-reported on the full V2 set

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 8. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Developer report, official protocol (4)

  • AlphaFold 3 Self-reported PoseBusters V1 (428), official set · Self-report on the official V1 set; V2 numbers are third-party recomputations · evidence: link
  • Chai-1 Self-reported PoseBusters V1 (428), official set · V1 self-report (preprint, quoted via reviews); the V2 numbers are third-party · evidence: link
  • NeuralPLexer3 Self-reported · V2 set, with validity checks included · evidence: link
  • Protenix Self-reported official set, early version only (via the Mini paper) · Only an early version's table; current versions report no score on this set, so the matrix cell overstated it low confidence †unverified · evidence: link

Developer report, modified protocol (4)

  • IntelliFold-2 Self-reported PoseBusters V2 (README, steering option) · Approximate values in the repository README with a steering option; no table †partly verified · evidence: link
  • Nesso-1 Self-reported · A missing candidate added and confirmed in the review Resolution: The same sentence names this benchmark; its poses are coarse-grained, so the physical-validity checks are not run as on all-atom predictions, hence modified protocol. · evidence: link
  • Pearl Self-reported PoseBusters V2, best@5 with PB-valid filter · Mixes its own best@5 with another model's top-confidence result; one model excluded for its cutoff †partly verified · evidence: link
  • TerraBind Self-reported · Merged into a chart with two other benchmarks; no number for this benchmark alone †partly verified · evidence: link

Cited as comparison only (3, not counted)

  • Boltz-1 / Boltz-1x Third-party PoseBusters V2, best@5 · Numbers come from another paper, not a self-report †partly verified · evidence: link
  • DiffDock Third-party PoseBusters V2 comparison table in a third-party paper · Quoted from the comparison table of another model's paper †partly verified · evidence: link
  • Vina Third-party PoseBusters V2 with known pocket, from a third-party comparison table · Quoted from the comparison table of another model's paper †partly verified · evidence: link

Caveats

  1. Some structures fall inside newer models' training windows, so about 80% success cannot be read as generalization to new pockets.
  2. Docking and co-folding differ in inputs and sampling rules, so cross-comparisons are not like-for-like.
  3. With only 308 systems in V2, differences of a few percentage points at the top do not support a ranking.
  4. The sets are frozen since V2 and near saturation, and most new models already satisfy the validity checks.
  5. Pose accuracy is not affinity or activity, and no public study was found linking these scores to wet-lab hit rates.

How to use it

Install the official package, predict a pose for each complex and let the tool compute RMSD and the 18 checks; it runs statically with no wet lab of your own. Report the set version, whether validity checks are included, the sampling rule, the training cutoff and any overlap with the set. Generalization claims need similarity-stratified splits or a newer time split. Pair with PLINDER and PoseX.

Pairs well with: PLINDER / PINDER / Runs N' Poses · PoseX · FoldBench · CASF-2016 · CrossDocked2020

Sources

  1. Original preprint (preprint)
  2. Official repository (repository)
  3. Official package on PyPI (package registry)
  4. Third-party model paper using a subset of this benchmark (third-party paper)
  5. Model paper recomputing results on the full V2 set (third-party paper)
  6. Co-folding model paper reporting V1 results (journal)

Corrections

5 corrections
FieldWasNowBasisFound
latest_update_kindnew-versioncode-maintenance-onlysource2026-11-16
latest_update_descRepo added HTML diagnostic report (#102); 2026-09-17 version bumped 0.6.5 to 0.7.0. PyPI still shows 0.6.5 (2026-03-07).Unreleased commits on main: HTML report (2026-10-03) and a 0.7.0 version-bump commit (2026-09-17); no tag or release object. Last tagged release and PyPI version: v0.6.5 (2026-03-07). Benchmark sets V1/V2 unchanged.source2026-11-16
domainsmall-moleculestructuresource2026-11-16
n_models_reported108 counted under the new relation taxonomy, with comparison-only citations excludedsource2026-11-16
models_reported (one co-folding model)listed as reporting this benchmark in its current versionsversion-qualified: only an early version's table reports a score on this setsource2026-11-16

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