AI for Bio · Benchmark 30

PLINDER / PINDER / Runs N' Poses

Docking and co-folding infrastructure that makes training-set similarity a first-class split criterion.

bench-30Type C+BEvidence L3Structure & co-folding● Active Board中文

Overview

NamePLINDER / PINDER / Runs N' Poses · aliases: PLINDER, PINDER, Runs N' Poses, RnP, PLINDER-PL50, PLINDER-TIME, PLINDER-ECOD, plinder, pinder
Full namePLINDER (protein-ligand interaction dataset and evaluation), PINDER (protein-protein interaction dataset and evaluation) and Runs N' Poses (co-folding generalization test set)
Ambiguous names (not auto-linked)runs-n-poses
FamilyAll three are leakage-aware infrastructure from one organization: PLINDER and PINDER provide train/test splits; Runs N' Poses stratifies trained co-folding models post hoc.
  • PLINDER · †
  • PINDER · †
  • Runs N' Poses · †
DomainStructure & co-folding
InclusionOriginal 30
TypeC B · Split tooling and datasets, plus a stratified static test set
Evidence levelL3 · Strict OOD retrospective splits, no wet lab
PublisherVantAI, NVIDIA, MIT CSAIL, University of Basel, SIB Swiss Institute of Bioinformatics (Mixed)
First release† partly verified · paper v1
Basis: bioRxiv v1 of PLINDER, the earliest member; PINDER is recorded as 2024-07-19 and Runs N' Poses as 2025-02-07.
Note: Member dates rely on dynamic pages; see unverified_fields source
Other milestones
  • Preprint v1: † · PLINDER bioRxiv v1 source
  • Journal online: · Runs N' Poses final version in Nature Structural & Molecular Biology source
  • Code release: · First PyPI upload of the plinder package (pinder first upload 2024-07-19) source
Latest update† partly verified · new version
Basis: plinder 1.0.0 released on PyPI and GitHub
plinder 1.0.0 released (Metaflow removed); later commits are docs-only. pinder 0.6.0 (2026-09-14) moved dataset hosting to R2. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedThe maintainers have announced no date for a next release or a new challenge round.
Status● Active · Active; plinder 1.0.0 released 2026-10-09
ScalePLINDER: 449,383 systems; PINDER: 2,319,564 dimers; Runs N' Poses: 2,073–2,600 systems.
  • PLINDER protein-ligand systems: 449,383 systems
  • PLINDER annotations per system: >500 annotations
  • PINDER dimers: 2,319,564 dimers
  • PINDER apo monomers: 47,717 monomers
  • PINDER AFDB monomers: 42,827 monomers
  • PINDER test set: 1,955 systems
  • Runs N' Poses systems (preprint): 2,600 systems (preprint)
  • Runs N' Poses systems (Polaris release): 2,073 systems (Polaris)
  • Systems in the lowest-similarity bin (0–20): 76 systems
Wet-lab refreshGrows from public sources · No new experimental data; follows weekly PDB updates and reuses experimental structures.
Own wet lab needed to evaluate a new modelNo, run it as a download
Models reported5 counted (self 3, modified protocol 2, cited only 4)
Unverified fields† first_release, latest_update, family.members[0], family.members[1], family.members[2], release_milestones.paper_v1, models_reported[0], models_reported[1], models_reported[2], models_reported[4], models_reported[5], models_reported[6], models_reported[7], models_reported[8]
As of2026-11-16

What it measures

  • Task: Docking and co-folding generalization to systems dissimilar from training data in protein, pocket, interaction and ligand, reported by similarity bin.
  • Truth: Published experimental PDB structures; Runs N' Poses uses high-resolution systems released after 2021-09-30.
  • Primary metric: Pose success rate (LDDT-PLI > 0.8 and BiSyRMSD < 2 Å)
  • Secondary metrics: PINDER: DockQ and iRMSD-type metrics; Results stratified into 8 training-similarity bins
  • Version traps: The PL50 split's system definition differs between v1 (2024-04) and the v2 used in the MLSB challenge (2024-06), which relies on ligand distance only; numbers across versions are not comparable.

Current ranking

Board

Metricsuccess_rate: share of predictions with LDDT-PLI > 0.8 and BiSyRMSD < 2; the page also lists rmsd_success_rate, mean_lddt_pli and coverage
Eval setRuns N' Poses (Polaris preview benchmark, plinder-org/runs-n-poses)
Snapshot
Boardhttps://polaris-hub.vercel.app/benchmarks/plinder-org/runs-n-poses
Retrieved2026-11-16

Caveat. PLINDER and PINDER have no board; the only board is a Polaris preview of Runs N' Poses whose values sit far below the paper and whose bin counts conflict, so it is not a true ordering. Newer models are self-reported or second-hand on own subsets, unranked.

Shown in the overview table:

success_rate 15.068; rmsd_success_rate 17.808; mean_lddt_pli 0.304; coverage 96.053
success_rate: share of predictions with LDDT-PLI > 0.8 and BiSyRMSD < 2; the page also lists rmsd_success_rate, mean_lddt_pli and coverage · Runs N' Poses (Polaris preview benchmark, plinder-org/runs-n-poses)
Board ·

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
1AlphaFold3success_rate 15.068; rmsd_success_rate 17.808; mean_lddt_pli 0.304; coverage 96.053success_rate: share of predictions with LDDT-PLI > 0.8 and BiSyRMSD < 2; the page also lists rmsd_success_rate, mean_lddt_pli and coverage
Runs N' Poses (Polaris preview benchmark, plinder-org/runs-n-poses)
Board
source Board top-1; benchmark created 2025-02-03, figure updated 2025-02-04; no per-entry date
2Boltz-1success_rate 5.556; rmsd_success_rate 9.259; mean_lddt_pli 0.250; coverage 71.053success_rate: share of predictions with LDDT-PLI > 0.8 and BiSyRMSD < 2; the page also lists rmsd_success_rate, mean_lddt_pli and coverage
Runs N' Poses (Polaris preview benchmark, plinder-org/runs-n-poses)
Board
source
3Chai-1success_rate 4.688; rmsd_success_rate 18.750; mean_lddt_pli 0.281; coverage 84.211success_rate: share of predictions with LDDT-PLI > 0.8 and BiSyRMSD < 2; the page also lists rmsd_success_rate, mean_lddt_pli and coverage
Runs N' Poses (Polaris preview benchmark, plinder-org/runs-n-poses)
Board
source
4Protenixsuccess_rate 1.471; rmsd_success_rate 11.765; mean_lddt_pli 0.251; coverage 89.474success_rate: share of predictions with LDDT-PLI > 0.8 and BiSyRMSD < 2; the page also lists rmsd_success_rate, mean_lddt_pli and coverage
Runs N' Poses (Polaris preview benchmark, plinder-org/runs-n-poses)
Board
source
—Pearl85.2% (<2 Å, PoseBusters-valid, best@5); 70.0% at <1 ÅShare with RMSD < 2 Å and PoseBusters-valid (best@5)
Runs N' Poses (702-structure filtered subset, Pearl paper)
Self-reported
source Different protocol (best@5, filtered 702-structure subset); not comparable with the Polaris board
—AlphaFold374.4% (<2 Å, best@5); 61.5% at <1 ÅShare with RMSD < 2 Å (best@5)
Runs N' Poses (702-structure filtered subset, Pearl paper)
Third-party
source second-hand Second-hand, quoted from the Pearl paper's comparison table, which also lists Boltz, Chai-1 and Protenix

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 5. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Developer report, official protocol (3)

  • IsoDDE Self-reported · Tech report claims a large gain in the hardest Runs N' Poses bin; no numeric page †partly verified · evidence: link
  • OpenFold3 Self-reported Runs N' Poses (protein-ligand) · Runs N' Poses appears in a README bar chart; no numbers in text †partly verified · evidence: link
  • Pearl Self-reported Runs N' Poses, best@5 with PB-valid filter · Runs N' Poses best@5; baselines rerun by the Pearl authors †partly verified · evidence: link

Developer report, modified protocol (2)

  • Nesso-1 Self-reported · A missing candidate added and confirmed in the review Resolution: Third structural benchmark named in the same sentence, with the same metrics, baseline sourcing and filtering caveat; classed as a modified-protocol self-report. · evidence: link
  • Protenix Self-reported v1 self-report on the subset where all compared models have results (figure only) · v1 self-report on the subset where all models have results, figure only; plus a third-party figure in the Pearl paper †partly verified · evidence: link

Cited as comparison only (4, not counted)

  • AlphaFold 3 Third-party Runs N' Poses, best@5 · Numbers quoted second-hand from the Pearl paper †partly verified · evidence: link
  • Boltz-1 / Boltz-1x Third-party Runs N' Poses, best@5 · Numbers from the Pearl paper †unverified · evidence: no source link
  • Boltz-2 Third-party Runs N' Poses · Quoted in the Pearl paper and flagged not directly comparable; its official evaluation docs omit this benchmark †unverified · evidence: no source link
  • Chai-1 Third-party Runs N' Poses, best@5 · Quoted from the Pearl paper †unverified · evidence: no source link

Mentioned, not a pairing

  • · This family appears only as protein-ligand training data, which is not a pairing no source link

Caveats

  1. Binned numbers are mostly developer-reported with differing definitions; system definitions changed between versions; the lowest-similarity bin has only 76 systems.
  2. PLINDER splits only constrain models trained on them; fixed-training co-folding models can only be stratified post hoc against an approximate shared training set, with slight mismatch.
  3. The Polaris preview board diverges widely from the paper and has inconsistent bin counts, so it should not be cited as a true ordering.
  4. Several subsets and best@5/best@15 protocols coexist under one name, so numbers across papers are not comparable.
  5. High-similarity bins are near saturation, and there is no direct evidence that score gains translate into wet-lab success.

How to use it

Runs statically with no wet lab. Train own models on the PLINDER-PL50 split and report by similarity bin, not just overall; stratify trained co-folding models with Runs N' Poses. State data version and split files, seeds, pose selection (top-1 confidence or best-of-N) and pretraining cutoffs; add PoseBusters validity checks. Not for screening enrichment, affinity ranking or ADMET.

Pairs well with: PoseBusters · FoldBench · PoseX · CAPRI · CASF-2016

Sources

  1. PLINDER preprint (bioRxiv) (preprint)
  2. PINDER preprint (bioRxiv) (preprint)
  3. Runs N' Poses preprint (bioRxiv v3) (preprint)
  4. Runs N' Poses final version (Nature Structural & Molecular Biology 2026) (journal)
  5. PLINDER dataset documentation (official site)
  6. PLINDER official repository (repository)
  7. plinder on PyPI (package registry)
  8. pinder on PyPI (package registry)
  9. Runs N' Poses preview benchmark on Polaris (leaderboard)
  10. MLSB 2024 challenge page (official site)
  11. Pearl model paper (Runs N' Poses comparison table) (third-party paper)
  12. DockGen paper (earlier ECOD leakage-aware split) (third-party paper)

Corrections

3 corrections
FieldWasNowBasisFound
n_models_reported85 counted under the new relation taxonomy (4 comparison-only citations excluded)source2026-11-16
models_reported (Nesso-1)not listedself-reported-modified-protocolsource2026-11-16
non_pairings (Proteina-Complexa)candidate pairingnon-pairing: used only as training datasource2026-11-16

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