Protein Design · Baker Lab
ProteinMPNN
Generates amino acid sequences predicted to fold into specified target protein backbone structures.
What ProteinMPNN is
ProteinMPNN is a deep learning model developed by the Baker Lab for inverse protein folding and sequence design. Given a target protein backbone structure, it generates amino acid sequences predicted to adopt that specific three-dimensional conformation.
In computational protein engineering, sequence design typically follows de novo backbone generation. ProteinMPNN achieves substantially higher experimental success rates than physics-based design methods such as Rosetta, offering a rapid and consistent procedure for generating candidate sequences for laboratory evaluation.
How it works
The model operates on structural backbones represented either by full backbone atom coordinates or by alpha-carbon (C-alpha) coordinates alone. A dedicated model variant is trained specifically on soluble proteins.
ProteinMPNN can evaluate conditional or unconditional residue probabilities and score existing backbone-sequence pairs by negative log probability. For sequence generation, amino acids are sampled autoregressively according to a configurable temperature parameter. The model accommodates design constraints including fixed chains, fixed positions, tied positions for symmetric assemblies, amino acid exclusion lists, amino acid composition biases, and position-specific scoring matrix (PSSM) integration.
Key facts
- Version
- 1.0.1
- Modality
- Protein structure
- Inputs
- Protein backbone coordinates (PDB or JSONL), sequence count, sampling temperature
- Outputs
- Sequences, per-design negative log-probability scores, sequence recovery rates
- Memory requirement
- 8 GB GPU memory
- Maximum sequence length
- 200,000 residues
- Supported backbones
- Full-atom backbone or C-alpha-only traces
What you can use it for
- De novo sequence design
- Generating novel amino acid sequences tailored to fold into computationally generated or native protein backbone coordinates.
- Symmetric and homooligomeric design
- Designing multi-chain complexes by tying residue positions together to enforce identical sequence identities across symmetric subunits.
- Site-directed sequence redesign
- Redesigning designated chains or flexible loop regions while keeping critical functional motifs and binding interfaces fixed.
- Backbone-sequence scoring
- Evaluating the structural compatibility of existing sequences against a backbone target using conditional and unconditional probability scores.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on Managed GPU endpoint
Runs on managed compute. Availability and startup time depend on current capacity.Run this model
ProteinMPNN
Strengths and limits
Good at
- Achieves higher experimental validation rates compared to traditional physics-based methods like Rosetta.
- Operates on both full-atom backbone representations and C-alpha-only coordinates.
- Provides extensive sequence control, including residue fixing, position tying, amino acid omission, and compositional biasing.
- Supports large macromolecular structures up to 200,000 residues in length.
Watch out for
- Requires a predefined or generated backbone coordinate file to design sequences.
- Higher batch sizes can lead to out-of-memory errors on graphics hardware with limited memory.
Inputs and outputs
You provide
Backbone (PDB)
Typically the output of RFDiffusion, or an existing structure to redesign.
Sequencesoptional
Sampling temperatureoptional
Lower is more conservative; 0.1 is the usual default.
You get back
Sequences with per-design score and global recovery rate.
Licence
Commercial use permitted · MIT
The MIT licence allows use in commercial research and products.
Questions
Can I use ProteinMPNN commercially?
Commercial use permitted. The MIT licence allows use in commercial research and products.
How do I run ProteinMPNN without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 16GB, typically in 6.0s, and a run costs about 42 credits. Visitors without an account get 1,000 credits a month.
What input coordinate formats does the model accept?
The model accepts structural backbones in standard PDB format or pre-parsed JSONL files, supporting both full backbone coordinates and C-alpha-only coordinates.
How does the sampling temperature affect sequence generation?
Sampling temperature governs the stochasticity of amino acid selection, with recommended values between 0.1 and 0.3. Lower temperatures yield higher sequence recovery, whereas higher temperatures produce greater sequence diversity.
Can specific residues or partner chains be held fixed?
Yes. Configuration flags allow specific chains or residue indices to remain unmutated, enabling partial sequence redesign and interface preservation.
What are the soluble model weights?
The soluble model weights are trained exclusively on soluble protein structures rather than the full diverse structural training set.
How to cite it
Dauparas, J., Anishchenko, I., Bennett, N., Bai, H., Ragotte, R. J., Milles, L. F., Wicky, B. I., Courbet, A., de Haas, R. J., Bethel, N., et al. Robust deep learning--based protein sequence design using ProteinMPNN. Science 378, no. 6615 (2022): 49-56.
Written from ProteinMPNN's own documentation (code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.