Chemistry & ADMET · Correia Lab / EPFL
DiffLinker
Joins 3D fragments into one molecule by generating the atoms in between.
What DiffLinker is
Fragment linking is the step after a screen finds two pieces that bind in different sub-pockets: something has to connect them without breaking either pose. DiffLinker generates the connecting atoms directly in 3D, and unlike earlier linker models it takes any number of fragments, needs no attachment points specified and predicts how long the linker should be rather than being told. A separate size model supplies that prediction. Checkpoints are released for two training sets and, for pocket-conditioned generation, for the full-atom and backbone-only representations of the protein around the fragments — the pocket-conditioned variants currently require the anchor atoms and the linker size to be given. Code is MIT; the checkpoints are published separately under their own licence.
Run it here
We list this model but cannot run it yet: it needs two checkpoints fetched from Zenodo and a GPU image to run them in — we have not built one yet.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Inputs and outputs
You provide
Fragments
Two or more disconnected fragments in 3D. Their poses are kept.
Pocket structure (PDB)optional
Optional. With a pocket the linker is generated inside it; without one the fragments are linked alone.
Linkersoptional
You get back
Complete molecules incorporating every input fragment, as an SDF file.
Licence
Commercial use permitted · MIT code / CC BY 4.0 weights
Verified 2026-10-24: the repository declares MIT, and both Zenodo records holding the released checkpoints declare CC BY 4.0 — commercial use permitted, attribution required.
Questions
Can I use DiffLinker commercially?
Commercial use permitted. Verified 2026-10-24: the repository declares MIT, and both Zenodo records holding the released checkpoints declare CC BY 4.0 — commercial use permitted, attribution required.