Protein Design · Baker Lab / UW IPD
LigandMPNN
Inverse folding that sees the ligand: sequence design conditioned on the atoms around the protein.
What LigandMPNN is
Standard inverse folding reads a protein backbone and nothing else, so it designs a binding site as though the thing being bound were not there. LigandMPNN conditions on the non-protein atomic context as well — small molecules, nucleotides, metals — which is the difference that matters when the residues being designed line a pocket. It is published as inference code with several model variants sharing one interface, including the original protein-only model and a variant trained for soluble expression, plus a side-chain packing model that returns coordinates rather than sequence alone. Per-sequence confidence values come back with the designs. Code and parameters are MIT, which makes it one of the few design models with no licence caveat at all.
Run it here
We list this model but cannot run it yet: its weights are published on the authors' own server, so a GPU image has to fetch and pin them — 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
Structure (PDB)
The context is read from the same file as the backbone, so keep the non-protein atoms in it.
Sequencesoptional
Sampling temperatureoptional
Lower is more conservative. 0.1 is the published default.
You get back
Sequences with the model's confidence and recovery against the input.
Coordinates for the designed residues, when side-chain packing is requested.
Licence
Commercial use permitted · MIT
Verified 2026-10-24: the repository declares MIT and its README states that the code and the model parameters are both available under it. The parameters are distributed from the authors' institutional server rather than a model hub.
Questions
Can I use LigandMPNN commercially?
Commercial use permitted. Verified 2026-10-24: the repository declares MIT and its README states that the code and the model parameters are both available under it. The parameters are distributed from the authors' institutional server rather than a model hub.