---
title: How DiffDock results compare to the bench
description: >-
  What a DiffDock pose costs against the docking seat and the co-crystal it
  displaces, and what the confidence score does not tell you.
tags:
  - diffdock
  - validation
  - translatability
---

**A blind pose for 149 credits, against a licensed docking suite and a co-crystallization order.**

DiffDock-L places a ligand on a protein without being told where the pocket is, in 3–35 minutes, at **149 credits**. It reaches 50% of poses within 2 Å on PoseBusters as a blind method, against 58–60% for Gold and AutoDock Vina *given the binding site*. That is close to parity on accuracy, without the site and without the licence.

### What a run displaces

| Spend category | What you would otherwise buy | This run |
|---|---|---|
| **In-silico seat cost** | A docking suite licence — Gold requires a CCDC licence, Glide a Schrödinger one, each renewed annually per seat or site | 149 credits, no licence, no renewal {/* SEAT COST PENDING — a research lane is pulling Schrödinger's SEC-disclosed ACV per customer and published procurement awards for Gold/Glide/MOE. Insert the figure with its scope (seat vs site, annual vs perpetual) and source when it lands. Do NOT publish an estimate. */} |
| **CRO/CDMO order** | A co-crystallography order. ⊘ No commercial CRO publishes a price for this — ~110 vendors checked, October 2026, all quote-only | — |
| **Reagents** | Crystallization screening at **€200 per 96-well plate** (at least 376 DiffDock runs), before any collection | — |
| **Assays** | SPR at **$357.50/day** (at least 599 DiffDock runs per day) to establish engagement; ITC and MST have **no public price of any kind** | — |
| **In-vivo experiments** | ⊘ **Not displaced.** A pose hypothesis sits far upstream of animal work. Any page claiming otherwise is overreaching | — |

Wet-lab figures are academic core-facility rate cards at their commercial tier, named and dated in the [shared cost basis](/docs/translatability#what-the-wet-lab-cost-figures-mean). Read them as recharge rates and therefore as a **floor** — a CRO quote widens every gap above.

### What the result does and does not support

- **A high-confidence top pose is a geometric hypothesis** about where the ligand sits. On targets resembling the training data, roughly half of such poses land within 2 Å of the experimentally determined pose. On novel targets, far fewer do.
- **Do not rank compounds by the confidence score.** It was trained to predict pose correctness, not binding strength, and independent evaluation found it *negatively* correlated with measured potency (Spearman ρ = −0.135 on an EGFR set).
- **Confirm a pose with** X-ray crystallography of the complex, or a biophysical binding measurement (MST, SPR, ITC) if you only need to establish engagement.
- **Confirm affinity with an assay.** There is no computational affinity step here: [MM-GBSA](/docs/mmgbsa) on this platform returns no binding free energy (`delta_g_bind` is always null), and `fep_estimate` is a descriptor heuristic, not an alchemical calculation. Nothing on the platform converts a pose into a predicted affinity.

Jump to [agreement with wet lab](#agreement-with-wet-lab) or [against other in silico methods](#against-other-in-silico-methods). The general framing for every tool is in [reading an in silico result against the bench](/docs/translatability).

## How in silico is used currently

Docking is used to **generate and triage hypotheses**, not to decide them. In practice that means three things in a live program:

- **Pose hypothesis.** You have a target structure and a ligand known or suspected to bind, and you want a model of the interaction to reason about — which residues to mutate, where to grow a substituent, what to ask a crystallographer to look for.
- **Library triage.** You reduce a large compound set to a shortlist worth buying or synthesizing. This is where the spend reduction is real and compounding: every compound triaged out is a compound you do not order, synthesize or assay. The field's accepted standard here is a physics-based or consensus pipeline, not a single deep-learning method — and published campaigns that ended in confirmed inhibitors used DiffDock as *one stage among several*, alongside AutoDock Vina, GNINA, or molecular dynamics and MM-PBSA rescoring.
- **Site mapping.** Blind docking over a whole protein to ask *where* a ligand might engage, which is DiffDock's distinctive capability relative to pocket-constrained methods.

What docking is not used for, by anyone who reports their validation honestly, is deciding potency rank order.

## Against other in silico methods

On the PDBBind date-split benchmark, DiffDock's authors reported **38.2%** top-1 against GLIDE at 21.8%, GNINA at 22.9%, and TANKBind at 20.4%. On DockGen, a benchmark built to present binding-site types absent from training, DiffDock-L reached **22.6%** where the original DiffDock managed 7.1% (10 poses) and 6.0% (40 poses), and the best search-based method reached 17.5%.

Independent evaluation inverts part of that picture. On a post-2021 held-out set, both conventional methods beat every deep-learning method tested — and the study deliberately gave AutoDock Vina and Gold **larger than default search spaces** so they would be more comparable to blind methods, rather than handing them a pocket advantage. DiffDock remained the strongest deep-learning method on physical validity while being the weakest of the three on pose accuracy.

**Co-folding is the current state of the art for this task**, not docking. Where a cross-docking comparison put diffusion docking at 78.9% on its top-ranked pose, AlphaFold 3 reached 88.2% and Boltz-2 82.1%. If ligand-pose accuracy is what you need, [Boltz-2](/docs/boltz2) is the co-folding route on this platform at **13 credits** — cheaper than DiffDock and more accurate on this benchmark. [Chai-1](/docs/chai) runs here in a sequence-only configuration that accepts no ligand, so it is not an alternative for this task.

For ranking rather than posing, DiffDock's native confidence performs poorly. Across a benchmark of docking tools on an EGFR set assembled from measured IC50 values, screening-set Spearman correlations ran from **ρ = −0.135 (DiffDock)** to ρ = 0.303 (GNINA). The same work found that the large majority of DiffDock poses violated minimum protein–ligand distance checks and overlapped the receptor volume.

**On seat cost**, the comparison is the clearest in this section and the one with the least published data. Gold is licensed through the CCDC and Glide through Schrödinger; both are renewed per seat or per site, annually, independent of how many poses you generate. AutoDock Vina, GNINA and SMINA are free and open source, so against *those* the honest claim is not price but throughput and integration — you are not maintaining the install or the queue.

> Agha H, et al. Data driven selection of consensus docking pipelines for structure based hit identification. *npj Drug Discovery*. 2026. doi: [10.1038/s44386-026-00063-4](https://doi.org/10.1038/s44386-026-00063-4)

## In silico vs. wet lab

The standard wet-lab approach for the question DiffDock answers is **X-ray crystallography of the ligand–protein complex**. Co-crystallization establishes the binding mode directly: the ligand's position, orientation, and conformation in the bound state, at a resolution that resolves individual contacts.

The comparison between the two is unusually well defined, which is why this section can carry real numbers. A predicted pose and a crystal pose are compared as **heavy-atom RMSD**, and the field's conventional success threshold is **≤ 2 Å**. That gives a single, auditable agreement metric.

| | Time to answer | Cost basis | Accuracy, and its metric |
|---|---|---|---|
| **Standard wet lab** — co-crystal, plus ITC/SPR for affinity | Weeks to months; may never crystallize | €2,000 data collection (about 3,630 runs here) + €1,000/day refinement (about 1,810 runs here per day) at a core facility's commercial tier; no CRO publishes a price. [What this means](/docs/translatability#what-the-wet-lab-cost-figures-mean) | Ground truth, limited by resolution |
| **Conventional docking** — Gold, AutoDock Vina | Minutes to hours, CPU | Gold is commercially licensed per seat or site; Vina is free. No per-run basis published | 58% / 60% ≤ 2 Å on PoseBusters, given binding-site location and shape |
| **SOTA in silico** — co-folding: AlphaFold 3, [Boltz-2](/docs/boltz2) | Minutes to hours, GPU | Boltz-2 here: 13 credits | 88.2% / 82.1% cross-docking success on three Alzheimer's targets, top-ranked pose |
| **Combined tooling here** — DiffDock-L, optionally minimized by [MM-GBSA](/docs/mmgbsa) | 3–35 min for a pose, plus 1–5 min per minimized pose | 154 credits | Pose only: 50% ≤ 2 Å on PoseBusters and 43% on the PDBBind date split, both author-reported; 78.9% cross-docking on the Alzheimer's set, top-ranked pose. **No affinity figure, because no affinity is computed** |

**Read the accuracy column by cell, never down the column.** The three entries are not the same measurement. Crystallography's "accuracy" is definitional — it is the thing the other two are graded against. A docking percentage is the share of poses within a geometric threshold. An affinity method's accuracy would be an energy error in kcal/mol. Nothing meaningful comes from comparing those numbers to each other; what the table supports is comparing *time and cost at a stated accuracy*.

On that reading, the honest position for this chain is: **minutes instead of months, at a pose accuracy approaching conventional docking while requiring no binding site.** The platform runs DiffDock-L, which reaches 50% on PoseBusters as a blind method against 58–60% for Gold and Vina given the pocket.

The weaker figure often quoted for DiffDock — 12% on that set — is for the **original DiffDock v1.0**, which is not what runs here. Both numbers are real and they measure different software.

Two further asymmetries matter. Crystallography can fail outright for a given complex and no amount of effort fixes it, which is a large part of why computational pose prediction exists. And crystallography reports what *is*, while DiffDock reports what is *likely* — the two are not interchangeable at any confidence level.

## Agreement with wet lab

All figures below are RMSD ≤ 2 Å against deposited crystal structures. The spread between them is the most useful thing on this page.

**On benchmark sets drawn from the PDB**, author-reported, on 363 test structures from 2019 held out by date:

- DiffDock, 10 poses sampled: **35.0%** top-1, median RMSD 3.6 Å. Top-5 40.7%.
- DiffDock, 40 poses: **38.2%** top-1, median 3.3 Å. Top-5 44.7%.
- DiffDock-L, 10 poses: **43.0%** top-1, median 2.8 Å.
- On [ESMFold](/docs/esmfold)-predicted rather than crystal receptors, this falls to **21.7%** — a loss of roughly a third. Read that next to what happened to search-based docking on the same receptors: GNINA fell from 22.9% to **2.0%** and SMINA from 18.7% to 3.4%. DiffDock was the most robust method tested to a predicted receptor, so the lesson is not that DiffDock degrades but that conventional docking collapses.

**On an independently constructed set of complexes released from 2021 onward**, chosen so that no evaluated method had seen them in training, **the original DiffDock v1.0 reached 12%** — against 58% for AutoDock Vina and 55% for Gold. This is the most-cited criticism of diffusion docking, and it does not describe the version running here: on the same benchmark, DiffDock-L's authors report **50%**, against Gold 58% and Vina 60% in their own table, where the conventional methods were given the binding site's location and shape. On the older Astex Diverse set, where 47 of 85 complexes appear in PDBbind 2020, the same study put DiffDock at **72%** by RMSD alone, dropping to **47%** once poses were also required to be physically valid.

That progression — 72%, then 47%, then 12% — is one method evaluated three ways. It is the single most important thing to understand about a DiffDock number.

**On a cross-docking comparison across three Alzheimer's targets** (acetylcholinesterase, BACE1, GSK-3β), diffusion and co-folding methods held up under receptor mismatch where physics-based docking collapsed. Taking each method's own top-ranked pose, DiffDock reached **78.9%** and GNINA 1.3 fell to **8.0%**, with AlphaFold 3 at 88.2% and [Boltz-2](/docs/boltz2) at 82.1%. On the paper's paired analysis the ordering is the same and the values differ — AlphaFold 90.0% (95% CI 87.0–92.3), Boltz 82.3% (78.4–85.6), DiffDock 72.9% (69.8–75.9) — because, as the authors put it, the pose-ranking strategy itself moves the estimate.

**A predicted site confirmed biophysically.** The closest published case to a prospective pose confirmation used DiffDock-L to map small-molecule binding across the TREM2 structure, identifying an ectodomain hydrophobic groove; microscale thermophoresis then confirmed the interaction, and a reporter assay supported the proposed competition with Aβ. The pose itself was not crystallographically resolved.

> Corso G, Stärk H, Jing B, Barzilay R, Jaakkola T. DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking. *ICLR*. 2023. doi: [10.48550/arXiv.2210.01776](https://doi.org/10.48550/arXiv.2210.01776)
>
> Corso G, Deng A, Polizzi N, Barzilay R, Jaakkola T. Deep Confident Steps to New Pockets: Strategies for Docking Generalization. *ICLR*. 2024. doi: [10.48550/arXiv.2402.18396](https://doi.org/10.48550/arXiv.2402.18396)
>
> Buttenschoen M, Morris GM, Deane CM. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. *Chemical Science*. 2024;15(9):3130-3139. doi: [10.1039/d3sc04185a](https://doi.org/10.1039/d3sc04185a)
>
> Suri K, Yadav A, Tripathi A, Murugan NA. Cross-docking and redocking reveal distinct determinants of success in physics-based and AI-driven binding pose prediction in protein-ligand complexes. *RSC Advances*. 2026;16(43):53425-53444. doi: [10.1039/d6ra05440d](https://doi.org/10.1039/d6ra05440d)
>
> Cho S, Gabr MT. Identification of a novel Aβ-overlapping binding site on the TREM2 ectodomain engaged by the small-molecule agonist VG-3927. *Bioorganic & Medicinal Chemistry Letters*. 2026;139:130712. doi: [10.1016/j.bmcl.2026.130712](https://doi.org/10.1016/j.bmcl.2026.130712)

## What the figure doesn't cover

- **Training overlap dominates the headline numbers.** The PDB is both the benchmark source and the training source. Stratifying by sequence identity to the training set shows deep-learning docking methods performing worse as identity falls, with the sub-30% band worst — the published conclusion is that these methods overfit to the proteins they were trained on. If your target has no close relative in the PDB, the 12% figure describes your situation better than the 43%.
- **RMSD is blind to physical validity.** A pose can sit within 2 Å and still clash with the receptor or contain strained geometry. The 72% → 47% drop on Astex is entirely this effect.
- **No published figure covers prospective pose confirmation.** No study has reported a DiffDock pose predicted in advance and then resolved by a newly solved co-crystal structure.
- **Campaign hit rates are not DiffDock's.** Published screens that produced confirmed inhibitors ran DiffDock inside multi-stage pipelines. Those IC50 values belong to the pipeline. For context on what the physics-based class has achieved prospectively, large-scale docking campaigns have reported confirmed hit rates in the 11–41% range against targets including D4, AmpC, MT1, KEAP1 and A2A — none of which used DiffDock.
- **Version matters, and it is the most common error in reading these numbers.** Nearly all independent evaluation tested DiffDock v1.0. The platform runs DiffDock-L, whose stronger generalization figures are so far author-reported — no independent group has yet published a DiffDock-L evaluation on a held-out set. So the harshest published figure is for software we do not run, and the better figure has not been independently checked.
- **The cost comparison is not a like-for-like against free software.** Against Gold or Glide the saving is a licence. Against AutoDock Vina, GNINA or SMINA — all free and open source — there is no licence to save, and the honest claim is throughput and not maintaining the install.
- **No affinity, at all — and no downstream step supplies one.** DiffDock returns no binding energy, and neither does the platform's MM-GBSA, whose `delta_g_bind` is always null. Treat every pose here as a geometric hypothesis only. See [outputs and ranking](/docs/diffdock/results) and [MM-GBSA translatability](/docs/mmgbsa/translatability).

## Go / no-go

**Go — act on the result when:**

- You want a binding-mode hypothesis to guide mutagenesis, SAR, or a crystallography request, and your target has close structural relatives in the PDB.
- You are mapping candidate sites across a protein where you do not know the pocket, which is the case pocket-constrained methods cannot address.
- You are triaging a library before committing to purchase, synthesis or assay — the stage where the spend reduction actually compounds.
- You treat the pose as geometry and take any affinity question to an assay. [MM-GBSA](/docs/mmgbsa) here minimizes the complex and reports its energy; it does not return a binding free energy, so it cannot settle an affinity question.
- You are using the pose as one input among several, in the consensus pattern that published campaigns actually used.

**No-go — do not act on the result when:**

- You are ranking or prioritizing compounds by confidence score. Use a scoring function validated for that, and expect near-random enrichment from confidence alone.
- Your target is structurally novel, with no close PDB relative, and you intend to treat a single top pose as correct.
- You need the binding mode established rather than proposed. That requires the co-crystal.
- You need the most accurate pose available here. That is [Boltz-2](/docs/boltz2), at a tenth the cost — DiffDock-L is the blind-docking option, not the accurate one.
- The pose will not be inspected. Check for clashes and strained torsions before anything downstream — physical implausibility is this method's documented failure mode, and the confidence score does not catch it.

**Caveats on reading this page:** every figure above is quoted from the cited source with its original metric and sample. Figures from different studies are not directly comparable, because test sets, search spaces, and validity criteria differ between them.