AI for Bio · Model

DiffDock

Diffusion-based blind docking, now used mainly as the deep-learning docking baseline

Structure & co-foldingAcademic labUnited StatesMaintained only中文

Overview

NameDiffDock · aliases: DiffDock, DiffDock-L
Full nameOfficial materials use DiffDock only; no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. The record covers DiffDock and its in-repo successor DiffDock-L without separate family members
OrganizationMIT CSAIL
Organization typeAcademic lab
CountryUnited States
Model classBlind docking (diffusion plus confidence ranking)
DomainStructure & co-folding (also: Small molecules)
First release · preprint
Basis: arXiv v1 (conference version the following year) source
Other milestones
  • First public evidence: · arXiv v1 source
  • Preprint v1: · arXiv v1 source
  • Code release: · GitHub v1.1 (DiffDock-L) source
  • Weights release: · DiffDock-L weights shipped with v1.1 source
Latest update · new version
Basis: GitHub v1.1: DiffDock-L release
DiffDock-L was trained with larger augmented data and confidence bootstrapping and appeared at that year's conference; v1.1.1-v1.1.3 (to 2024-09-04) are fixes and the last commit is 2025-05-02 source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · No newer generation; later activity is maintenance
OpennessCode: Open
Weights: Open
Code and weights are public on GitHub, with DiffDock-L as the repository default
Benchmarks reported2 counted (retrospective 2, cited only 1)
Wet-lab validationNone recorded
Unverified fieldsNone
As of2026-11-16

What it does

Treats docking as generation: it diffuses over ligand translation, rotation and torsion, then ranks candidate poses with a confidence model, and DiffDock-L is the larger-data successor. In the co-folding era it serves as the deep-learning docking baseline against the physics-scoring line of Vina and against AlphaFold3 and NeuralPLexer3.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 2 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
12 DUD-E / LIT-PCBARetrospective dataset useLIT-PCBA virtual-screening study by a third partyThird-partyThird-party LIT-PCBA study (arXiv 2605.01681) †partly verified · evidence
33 PoseXRetrospective dataset usePoseX, rerun by the benchmark authorsThird-partyEvaluated by the benchmark paper. †partly verified · evidence
05 PoseBustersCited as comparison only (not counted)PoseBusters V2 comparison table in a third-party paperThird-partyQuoted from the comparison table of another model's paper †partly verified · evidence

Other evaluations (outside the 44)

  • Self-reported blind-docking sets (self report) Self-reported: The DiffDock-L paper self-reports top-1 results on two blind-docking sets under a fixed sampling budget, a different protocol from physical-validity checks (no figures)
  • Second-hand reviews (second hand) Third-party: The claim that it beats other blind-docking methods partly reaches readers through reviews rather than independent reruns (second-hand)
  • Baseline role (third party evaluation) Third-party: Several co-folding and docking papers rerun it as the deep-learning baseline, with conclusions that vary by task

Developers report: Developers report that DiffDock-L, trained with larger augmented data and confidence bootstrapping, beats other blind-docking methods (self-reported, partly relayed by reviews).

Independent evidence: Third-party papers and benchmark authors have rerun it; the conclusion is a baseline role, with no edge in physical validity or screening (no figures).

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPDBBind, with DiffDock-L adding other PDB complexes and synthetic augmentation that cuts ligands from real structures source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An academic group with no company: teams at MIT CSAIL and an affiliated clinic, whose authors were then mostly doctoral students and faculty at the institute.
  • Funding: Academic project; funding amounts undisclosed
  • Products and partnerships: The successor DiffDock-L is the repository default; it shares an academic lineage with Boltz-2 but no institutional link is evident.
  • Related models: Vina, Boltz-2, NeuralPLexer3, AlphaFold 3

Reading notes

  1. Most of its benchmark figures come from other papers under mixed protocols and should be read as a baseline only.
  2. Its self-reported blind-docking protocol differs from physical-validity checks; do not swap them.
  3. It has no edge in virtual screening; strong pose generation does not imply strong screening.
  4. Classical docking scores higher when the pocket is known, but the settings differ.

Caveats

  1. The developers self-report on none of the site's benchmarks; all three relations come from third parties.
  2. The PoseBusters line is cited as comparison only and does not count as an established relation.
  3. The last substantive update was 2024-02, followed only by fixes and maintenance.

Sources

  1. DiffDock paper (preprint) ·
  2. DiffDock-L paper (preprint)
  3. DiffDock GitHub releases (repository)
  4. Third-party comparison table in a co-folding paper (third-party paper)
  5. Third-party virtual-screening study (third-party paper)
  6. Evaluation paper by the benchmark authors (third-party paper)

Corrections

None recorded.

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.