What is DiffDock?
Learning objectives
After reading this article you will be able to:
- Explain how a diffusion model generates docking poses
- Compare DiffDock with classical docking programs
- Know what DiffDock's confidence score does and does not tell you
DiffDock is a docking method introduced by researchers at MIT in 2022. Instead of searching poses and scoring them with a hand-built function, it generates poses with a diffusion model, the same family of generative AI used for image generation.
How it works
A diffusion model learns to reverse a noising process. During training, the ligand’s correct pose from experimental structures is progressively scrambled — its position, orientation and torsion angles are randomised — and the network learns how to undo each step. At prediction time DiffDock starts from random poses and iteratively “denoises” them towards plausible binding modes on the protein.
A separate confidence model then ranks the generated poses by how likely they are to be close to correct.
DiffDock compared with classical docking
| Classical (e.g. AutoDock Vina) | DiffDock | |
|---|---|---|
| Binding site | Usually a user-defined box | Can dock “blind” across the whole protein |
| Pose generation | Search plus physics-inspired scoring | Learned generative model |
| Output score | Estimated interaction energy (kcal/mol) | Confidence that the pose is right |
| Strengths | Well understood; physically interpretable | Fast blind pose prediction; no box needed |
| Watch out for | Box placement, scoring bias | Physically implausible poses; behaviour on targets unlike its training data |
Using it well
- Treat the confidence as a pose-quality signal, not as binding affinity. To rank compounds, combine it with a physics-based score or experimental data.
- Check poses for clashes and bad geometry; some versions of learned docking models can produce poses that look reasonable globally but are chemically strained.
- Blind docking is useful when the binding site is unknown. When it is known, confirm that top poses land there.
Because DiffDock is open source, it can be run on your own infrastructure, which matters when compound structures are confidential.
Last updated October 31, 2026
Put it into practice
Screen, dock and order compounds in one place
Pyxis Discovery combines macrocycle and small-molecule libraries with similarity search, docking and structure prediction in the browser.