AI for Bio · Benchmark 33

PoseX

Compares 23 physics, AI-docking and co-folding methods under one protocol, separating self-docking from cross-docking.

bench-33Type BEvidence L3Small molecules● Active Third-party中文

Overview

NamePoseX · aliases: PoseX, PoseX-SD, PoseX-CD
Full namePoseX: a protein-ligand docking and co-folding benchmark (self-docking and cross-docking)
Ambiguous names (not auto-linked)None
FamilyNot a family entry. Single benchmark; SD and CD are two subsets of it.
DomainSmall molecules
InclusionSupplementary entry
TypeB · Fixed test sets plus an author-run online board
Evidence levelL3 · Retrospective time-window evaluation, not blind
PublisherPoseX authors (academic consortium; code under the CataAI GitHub organisation) (Academic lab)
First release · paper v1
Basis: arXiv 2505.01700 v1 source
Other milestones
  • Preprint v1: · arXiv 2505.01700 v1 source
  • Code release: · Repository v1.0 release source
  • Official announcement: † · Accepted as an ICLR 2026 poster (year only) source
Latest update · paper revision
Basis: v3 date in the arXiv API
arXiv v3 revision (ICLR 2026 version); last repository commits are 2026-03 README/BibTeX edits; online board changes are undated. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedThe authors have announced no new version or board schedule.
Status● Active · Active; accepted at ICLR 2026; online leaderboard running, no public changelog
ScalePoseX-SD 718 entries; PoseX-CD 1,312 entries (109 targets, 371 structures, 362 ligands).
  • Self-docking entries: 718 (PoseX-SD, arXiv v3)
  • Cross-docking entries: 1,312 (PoseX-CD, arXiv v3)
  • Targets: 109 (PoseX-CD)
  • Structures: 371 (PoseX-CD)
  • Ligands: 362 (PoseX-CD)
  • Methods evaluated: 23 (arXiv v3 (5 physics, 11 AI docking, 7 co-folding))
Wet-lab refreshStatic data · Public PDB entries from 2022-2025; no wet lab.
Own wet lab needed to evaluate a new modelNo, run it as a download
Models reported6 counted (retrospective 6)
Unverified fields† ranking.snapshot_date, release_milestones.official_announcement
As of2026-11-16

What it measures

  • Task: Protein-ligand pose prediction in self- and cross-docking, with pocket-given vs blind and before/after relax settings.
  • Truth: High-quality PDB crystal structures released 2022-01 to 2025-01.
  • Primary metric: Top-1 pose success at RMSD<2Å
  • Secondary metrics: PoseBusters validity (PB-valid); Chirality errors and steric clashes
  • Version traps: State SD or CD, pocket-given or blind, before or after relax, and paper version (v1-v3 numbers may differ).

Current ranking

Third-party

MetricDocking success rate, top-1 pose RMSD<2Å
Eval setPoseX-SD (718 entries, self-docking) and PoseX-CD (1,312 entries, cross-docking), arXiv v3 (ICLR 2026) text
Snapshot† partly verified
Boardhttp://dock-lab.tech
Retrieved2026-11-16

Caveat. The online board could not be loaded, so values come from the arXiv v3 text (benchmark authors running others' models); tracks and settings are mixed, so unranked; author-maintained, newer co-folding models absent.

No headline entry: No headline number: the ranking source gives no rank-1 entry (all entries are unranked or protocol-specific); the caveat text is shown instead, as for a not-comparable cell.

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
—SurfDockSD 78.0% / CD 77.0%Docking success rate, top-1 pose RMSD<2Å
PoseX-SD (718 entries, self-docking) and PoseX-CD (1,312 entries, cross-docking), arXiv v3 (ICLR 2026) text
Third-party
source Not a site model; paper text says overall first on SD and CD (pocket-given track; relax setting not stated)
—AlphaFold3SD 60.5% overall / 60.3% blind track; CD 68.6% overall / 68.8% blind trackDocking success rate, top-1 pose RMSD<2Å
PoseX-SD (718 entries, self-docking) and PoseX-CD (1,312 entries, cross-docking), arXiv v3 (ICLR 2026) text
Third-party
source Best co-folding method in the blind track on both sets; paper sections give slightly different values
—ProtenixSD 56.3% (blind-docking track); CD not stated in textDocking success rate, top-1 pose RMSD<2Å
PoseX-SD (718 entries, self-docking) and PoseX-CD (1,312 entries, cross-docking), arXiv v3 (ICLR 2026) text
Third-party
source
—Chai-1SD 56.1%; CD 67.0% (blind-docking track)Docking success rate, top-1 pose RMSD<2Å
PoseX-SD (718 entries, self-docking) and PoseX-CD (1,312 entries, cross-docking), arXiv v3 (ICLR 2026) text
Third-party
source
—Boltz-1xCD 64.4% (blind-docking track); SD not stated in textDocking success rate, top-1 pose RMSD<2Å
PoseX-SD (718 entries, self-docking) and PoseX-CD (1,312 entries, cross-docking), arXiv v3 (ICLR 2026) text
Third-party
source The co-folding exception on chirality, thanks to its built-in physical potential

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 6. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Retrospective dataset use (6)

  • AlphaFold 3 Third-party self-docking / cross-docking, benchmark authors' run · PoseX authors ran AF3 under the same protocol (third party) †partly verified · evidence: link
  • Boltz-1 / Boltz-1x Third-party benchmark authors' own sets · The benchmark paper's authors evaluated this model on their own sets. †partly verified · evidence: link
  • Chai-1 Third-party benchmark authors' own sets · Same table as above. †partly verified · evidence: link
  • DiffDock Third-party PoseX, rerun by the benchmark authors · Evaluated by the benchmark paper. †partly verified · evidence: link
  • Protenix Third-party self-docking set, benchmark authors' run · PoseX authors ran Protenix under the same protocol †partly verified · evidence: link
  • Vina Third-party PoseX subset, evaluated by the benchmark authors · Evaluated by the benchmark paper on its Astex subset. †partly verified · evidence: link

Caveats

  1. Retrospective on PDB entries up to 2025-01, so newer models may have seen them. Leaderboard is maintained by the paper's authors. Pose accuracy does not measure affinity.
  2. The paper reports most AI methods improve after relax, but relax cannot fix chirality errors.
  3. No train-similarity stratification; memorisation must be checked with a stratified benchmark.
  4. The paper's 23 methods exclude newer models such as Boltz-2 and Pearl.

How to use it

Run the repository scripts on both SD and CD, reporting pocket-given vs blind and before/after relax, plus chirality accuracy and PB-valid, not only RMSD. Models trained past 2022 should use post-cutoff entries or flag leakage; no wet lab. Report alongside PoseBusters and Runs N' Poses similarity bins.

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

Sources

  1. PoseX paper (arXiv 2505.01700) (preprint)
  2. PoseX paper v3 full text (preprint)
  3. PoseX official repository (repository)
  4. ICLR 2026 poster page (official site)
  5. PoseX online leaderboard (could not be loaded this round) (leaderboard)

Corrections

1 correction
FieldWasNowBasisFound
latest_update / repositoryRepository has no releasev1.0 release on 2025-06-03; arXiv v3 on 2026-04-26 is the latest sourced eventsource2026-11-16

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