AI for Bio · Benchmark 12

DUD-E / LIT-PCBA

Two generations of virtual-screening enrichment sets: constructed decoys (DUD-E) and PubChem inactives (LIT-PCBA); both have known bias or leakage.

bench-12Type BEvidence L2Small molecules◐ Low-frequency Third-party中文

Overview

NameDUD-E / LIT-PCBA · aliases: DUD-E, DUDE, LIT-PCBA, DUDE-Z, Directory of Useful Decoys, Enhanced
Full nameDirectory of Useful Decoys, Enhanced / Literature-derived unbiased PubChem set for virtual screening
Ambiguous names (not auto-linked)DUD, MUV, DEKOIS
FamilyFamily: MUV, DUD-E, LIT-PCBA and successor DUDE-Z; DEKOIS 2.0 is only a related set; the original 2006 DUD was not checked or listed.
  • MUV · †
  • DUD-E ·
  • LIT-PCBA ·
  • DUDE-Z · †
DomainSmall molecules
InclusionOriginal 30
TypeB · Static virtual-screening dataset suite, no official board
Evidence levelL2 · Retrospective, random or clustered splits
PublisherUniversity of California, San Francisco (academic group), University of Strasbourg (academic group) (Academic lab)
First release
Basis: DUD-E paper (JMC online) source
Other milestones
  • Journal online: · DUD-E paper online in JMC source
Latest update† partly verified · maintenance only
Basis: DUDE-Z page update note
DUDE-Z page: recommended grids and ligand/decoy files re-checked against the paper with DOCK3.8; no new DUD-E or LIT-PCBA data version found.
Note: The 2025 third-party LIT-PCBA audit is not a dataset update source
The most recent sourced event, not a claim that nothing changed since.
Next expectedNeither dataset has an announced versioning plan.
Status◐ Low-frequency · Static; LIT-PCBA audited in 2025 and found to contain duplicates across splits
ScaleDUD-E: 102 targets, 22,886 actives; LIT-PCBA: 15 targets, 7,761 actives, 382,674 inactives (official site count).
  • Targets: 102 (DUD-E)
  • Actives: 22,886 (DUD-E)
  • Decoys per active: 50 (DUD-E)
  • Targets: 15 (LIT-PCBA)
  • Actives: 7,761 (LIT-PCBA (official site; some sources give 7,844))
  • Unique inactives: 382,674 (LIT-PCBA (official site; some sources give 407,381))
  • Targets: 43 (DUDE-Z)
Wet-lab refreshStatic data · None; DUD-E uses constructed decoys, LIT-PCBA uses PubChem measured data.
Own wet lab needed to evaluate a new modelYes, to claim real generalization
Models reported2 counted (retrospective 2, unconfirmed 3)
Unverified fields† latest_update, family.members[0], family.members[3], family.boundary, models_reported[2], models_reported[3], models_reported[4]
As of2026-11-16

What it measures

  • Task: Virtual screening enrichment: ranking true actives above decoys or measured inactives.
  • Truth: DUD-E: ChEMBL actives plus property-matched ZINC decoys; LIT-PCBA: PubChem dose-response labels.
  • Primary metric: AUROC, EF1%
  • Secondary metrics: BEDROC; logAUC
  • Version traps: Do not mix DUD-E, DUDE-Z and LIT-PCBA; LIT-PCBA's 'validation' is a held-out test needing a separate internal split; two count versions exist.

Current ranking

Third-party

MetricMedian EF1% (LIT-PCBA)
Eval setLIT-PCBA, 15 targets (DUD-E: no current site-model numbers)
Snapshot
BoardNo official board
Retrieved2026-11-16

Caveat. No official board for either set; source is a non-peer-reviewed preprint; AlphaFold 3, Protenix, Boltz-2 have only second-hand qualitative claims; known bias and leakage.

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
—Supervised ML re-ranking (best)median EF1% 4.49Median EF1% (LIT-PCBA)
LIT-PCBA, 15 targets (DUD-E: no current site-model numbers)
Third-party
source Not a site model; best in that study
—AutoDock-GPU + GNINA rescoringmedian EF1% 2.14Median EF1% (LIT-PCBA)
LIT-PCBA, 15 targets (DUD-E: no current site-model numbers)
Third-party
source Not a site model; best single scorer
—DiffDockmedian EF1% 1.17Median EF1% (LIT-PCBA)
LIT-PCBA, 15 targets (DUD-E: no current site-model numbers)
Third-party
source The DiffDock-GNINA variant scores lower; all variants score zero on OPRK1
—AutoDock Vinamedian EF1% 1.3, median AUROC 0.61Median EF1% (LIT-PCBA)
LIT-PCBA, 15 targets (DUD-E: no current site-model numbers)
Third-party
source Cited from literature in that preprint, not run by its authors; their own AutoDock-GPU run differs

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 2. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Retrospective dataset use (2)

  • DiffDock Third-party LIT-PCBA virtual-screening study by a third party · Third-party LIT-PCBA study (arXiv 2605.01681) †partly verified · evidence: link
  • Vina Third-party LIT-PCBA, rerun by a third-party study · Third-party LIT-PCBA study ran AutoDock/GNINA †partly verified · evidence: link

Unconfirmed (3, not counted)

  • AlphaFold 3 Unconfirmed source · Only a second-hand pose claim in an RSC 2026 paper; no numbers; the benchmark paper does not evaluate AF3 low confidence †disputed · evidence: no source link
  • Boltz-2 Unconfirmed source decoy-set screening in a third-party paper · Third-party paper uses DEKOIS 2.0 etc., not clearly DUD-E/LIT-PCBA; second-hand low confidence †disputed · evidence: link
  • Protenix Unconfirmed source · Second-hand LIT-PCBA mention, no numbers low confidence †disputed · evidence: no source link

Caveats

  1. DUD-E has decoy and analogue biases that inflate ML results. A 2025 preprint reports cross-split duplicates and near-neighbour leakage in LIT-PCBA (single preprint, not peer-reviewed).
  2. The audit claims a no-learning memorisation baseline matches or beats SOTA; no response from the LIT-PCBA authors was found.
  3. A 2026 preprint finds docking and ML re-ranking only slightly better than random on LIT-PCBA (not peer-reviewed).
  4. LIT-PCBA has two count versions (official site vs some archives); this entry follows the official site.
  5. Retrospective; no systematic evidence that score gains raise experimental hit rates.

How to use it

Download and run statically; no wet lab. Before LIT-PCBA, deduplicate train/validation, run a near-neighbour analysis and carve an internal validation split; report EF1% with intervals and per-target spread plus ligand-only and memorisation baselines. Novel-scaffold claims need prospective assays or CACHE; pair with CASF-2016 and PLINDER.

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

Sources

  1. DUD-E paper (J Med Chem 2012) (journal)
  2. DUD-E official site (official site)
  3. DUDE-Z official site (official site)
  4. LIT-PCBA paper (JCIM 2020) (journal)
  5. LIT-PCBA data page (official site)
  6. LIT-PCBA audit preprint (arXiv 2507.21404) (third-party paper)
  7. LIT-PCBA docking benchmark preprint (arXiv 2605.01681) (third-party paper)
  8. DUD-E bias analysis (PLoS ONE 2019) (third-party paper)
  9. Co-folding models in virtual screening (Chem Sci 2026, cited second-hand) (second-hand) second-hand

Corrections

None recorded.

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