AI for Bio · Benchmark 32

CrossDocked2020

The de facto default set for pocket-conditioned molecule generation, scored by docking software.

bench-32Type DEvidence L2Molecule generation○ Frozen Not comparable中文

Overview

NameCrossDocked2020 · aliases: CrossDocked2020, CrossDocked, CrossDocked2020 v1.3
Full nameThe CrossDocked2020 cross-docked pose dataset and its generative-task split
Ambiguous names (not auto-linked)CrossDocked split
FamilyThe dataset, the widely reused generative split and the unified re-implementation are three different artefacts, and any number must say which one it uses.
  • CrossDocked2020 dataset · †
  • 2021 generative split · †
  • CBGBench unified re-implementation · †
DomainMolecule generation
InclusionSupplementary entry
TypeD · Generative evaluation with docking software as the oracle, no official board
Evidence levelL2 · Retrospective evaluation against a computational oracle, no experimental truth
PublisherUniversity of Pittsburgh (Academic lab)
First release† partly verified · paper v1
Basis: The dataset paper's preprint (2020-02-20); the journal online date of 2020-08-31 is kept as a separate milestone.
Note: The preprint date was not re-retrieved from its original page this round source
Other milestones
  • Preprint v1: † · Preprint of the dataset paper source
  • Journal online: · Journal online publication date from the citation record source
  • Code release: † · First public version of the unified re-implementation source
Latest update† partly verified · data
Basis: Data-version record in a third-party database catalogue
A data-version update recorded in the catalogue, not checked against a primary release log; the evaluation code repository has since only changed documentation.
Note: The version date appears only in a third-party catalogue source
The most recent sourced event, not a claim that nothing changed since.
Next expectedNo maintainer has announced a new data version or new split.
Status○ Frozen · The dataset has not changed for years; the 2021 split is still in use
Scale22.5 million docked poses; the generative split has about 100,000 training complexes and 100 test pockets.
  • Docked poses: 22500000 (original dataset)
  • Training complexes: about 100000 (2021 generative split)
  • Test pockets: 100 (2021 generative split)
  • Split filtering criteria: RMSD<1Å; sequence identity<30% (2021 generative split)
Wet-lab refreshStatic data · No experiments; everything is computational docked poses and software scores.
Own wet lab needed to evaluate a new modelYes, to claim real generalization
Models reported1 counted (self 1)
Unverified fields† first_release, latest_update, release_milestones.paper_v1, family.members[1], family.members[2], ranking.entries[0]
As of2026-11-16

What it measures

  • Task: Generate ligands for a given protein pocket, scored by docking, drug-likeness, synthetic accessibility and diversity.
  • Truth: There is no experimental truth; the reference is docking-software scores and the reference ligands.
  • Primary metric: Docking score, including the high-affinity share
  • Secondary metrics: Drug-likeness and synthetic accessibility; Diversity; Pose validity and interaction recovery
  • Version traps: State which split and preprocessing were used: the widely reused one is looser than the dataset authors' and numbers compare only within one split.

Current ranking

Not comparable

MetricDocking scores, high-affinity share, drug-likeness, synthetic accessibility, diversity
Eval setCrossDocked2020, the 2021 NeurIPS generative split (~100k training complexes, 100 unseen test pockets)
Snapshot
BoardNo official board
Retrieved2026-11-16

Caveat. There is no official board and papers differ in split and preprocessing, so numbers do not compare; the common split is reported to be looser than the dataset authors' and may leak, and the unified re-implementation is a re-run, not a ranked board.

No headline entry: There is no official board and no comparable rank-1 entry, 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
—TamGenWithheld: Not comparable: splits and preprocessing differ, so the ranking row shows no score.Docking scores, high-affinity share, drug-likeness, synthetic accessibility, diversity
CrossDocked2020, the 2021 NeurIPS generative split (~100k training complexes, 100 unseen test pockets)
Self-reported
source The main text gives only a placing statement; per-metric values are in the supplement
—TargetDiffWithheld: Not comparable: splits and preprocessing differ, so the ranking row shows no score.Docking scores, high-affinity share, drug-likeness, synthetic accessibility, diversity
CrossDocked2020, the 2021 NeurIPS generative split (~100k training complexes, 100 unseen test pockets)
Self-reported
source Not a site model; the best-documented strong baseline on this split
—Pocket2MolWithheld: Not comparable: splits and preprocessing differ, so the ranking row shows no score.Docking scores, high-affinity share, drug-likeness, synthetic accessibility, diversity
CrossDocked2020, the 2021 NeurIPS generative split (~100k training complexes, 100 unseen test pockets)
Third-party
source Not a site model; a baseline row re-run in another paper

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 1. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Developer report, official protocol (1)

  • TamGen Self-reported · Self-reported on the 100 test pockets, with values in the supplement †partly verified · evidence: link

Caveats

  1. The common split is looser than the dataset authors' and may leak homologous information into training.
  2. Docking scores are only a software proxy with no wet-lab check.
  3. Pockets are treated as rigid and the ligand chemistry drifts from drug-like space, which has pushed later work to other sets.
  4. Docking scores are easy to optimise against, and a high score implies neither synthesizability nor activity.
  5. There is no maintainer or official results table, so cross-paper numbers compare only within one split.

How to use it

Download the data for one named split, generate ligands for each of the 100 test pockets and score them with the unified re-implementation for docking, drug-likeness, synthetic accessibility and diversity; no wet lab of your own is needed. Report the split source, preprocessing, number of generated molecules, sampling settings and pose validity. Docking is only a software proxy, so activity claims need experiments or an experimentally grounded benchmark. Pair with PoseBusters and PMO.

Pairs well with: PoseBusters · PMO (Practical Molecular Optimization) · CASF-2016 · DUD-E / LIT-PCBA · GuacaMol (+MOSES)

Sources

  1. Original dataset paper (journal)
  2. Model paper reusing and documenting the generative split (third-party paper)
  3. Unified re-implementation repository (repository)
  4. Third-party database catalogue entry (second-hand) second-hand
  5. Model paper self-reporting on this split (journal)

Corrections

4 corrections
FieldWasNowBasisFound
first_release2020-08-312020-02-20source2026-11-16
first_release_basisCrossDocked2020 dataset paper (JCIM, Europe PMC first publication date; Crossref online date not retrieved)paper_v1: ChemRxiv preprint of the CrossDocked2020 paper, 2020-02-20; JCIM journal online 2020-08-31 (Crossref) stored as a separate milestone.source2026-11-16
first_release2020-09-102020-08-31source2026-11-16
ranking.entriesper-model scores shown in the ranking rowscores withheld; splits and preprocessing differ and one entry has no main-text numberssource2026-11-16

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