Protein Design · Sequence prediction

ProteinMPNN

Protein sequence prediction, to conform with backbone coordinates. Does not take external molecules into account. A useful step after RFDiffusion in a protein design pipeline, and before validation with structure prediction.

A graph neural network designed for protein inverse folding, meaning it predicts the amino acid sequences most likely to fold into a specific 3D protein backbone structure. By interpreting the spatial coordinates and geometric features of a target structure, the model generates sequence candidates. Researchers use ProteinMPNN for applications such as optimizing enzymes, designing novel therapeutics, and improving the stability or solubility of synthetic proteins.

Documentation ↗ Input parameters ↗ Examples ↗ Paper ↗ Source code ↗ License ↗

No jobs before this; it will run immediately.

Not logged in. Register an account to keep track of your jobs and results, and add presets and bookmarks. Email results to this address when complete:

Ready

Example preset

Choose an example to fill in its settings and input structures.

Inputs

Official documentation ↗

Job name A label for this run and its results.

Backbone structure A single protein backbone to design or score. Leave chains blank to use every chain. In this form: Upload a PDB or mmCIF, or choose a preset with an included structure. The first model is used for multi-model structures. mmCIF is converted to PDB and requires single-character chain IDs.

Chains to design or score Space-separated chain IDs, e.g. A C. Other chains supply fixed sequence and structural context. Blank selects all chains.

Sequences to score (FASTA) Optional sequences for score-only mode. The native sequence is also scored. In this form: Supply one sequence per FASTA record, with designed chains in alphabetical order separated by /. Each chain must match its parsed backbone length.

Model and sampling

Official documentation ↗

Model checkpoint 48-neighbor models trained with backbone noise of 0.02, 0.10, 0.20 or 0.30 angstroms. CA-only weights are available through v_48_020.

Use soluble-protein weights Use weights trained on soluble proteins. Cannot be combined with CA-only mode.

Use CA-only structures and weights Use only alpha-carbon coordinates. Requires a CA-only checkpoint; v_48_030 is unavailable.

Samples per target Sequences per temperature in design mode; repeated evaluations in score/probability modes. Must be a multiple of batch size. In this form: This service accepts up to 1,000 samples per target.

Batch size Samples processed together. Reduce if GPU memory runs out. Must divide the sample count exactly.

Sampling temperatures One or more positive temperatures separated by spaces, e.g. 0.1 0.2 0.3. Higher values give more sequence diversity.

Random seed Zero chooses a random seed. Set a nonzero seed for reproducible sampling in the same runtime.

Inference backbone noise Standard deviation in angstroms of Gaussian noise added to backbone coordinates during inference; separate from checkpoint training noise.

Maximum total sequence length Maximum parsed sequence length across all chains. Longer inputs are rejected before running.

Position constraints

Official documentation ↗

Fixed positions (JSON) Positions whose native amino acids are retained, e.g. {"A": [1, 2, 3], "C": [10]}. Omitted chains are designed in full. In this form: Positions are 1-based indices in each parsed chain, including gaps and insertion-code residues, not PDB residue numbers. Enter the chain dictionary or the upstream dictionary keyed by structure name.

Tied positions (JSON) Groups sampled as the same amino acid, e.g. [{"A": [1], "C": [1]}, {"A": [2], "C": [2]}]. In this form: Uses the same 1-based parsed positions. Weighted groups also accept {"A": [[1], [0.5]], "C": [[1], [0.5]]}.

Tie all designed chains as a homo-oligomer Tie matching positions across the designed chains. Requires at least two chains of equal parsed length. In this form: Convenience equivalent of the official homooligomer helper. Use either this option or custom tied positions.

Amino-acid preferences

Official documentation ↗

Omit amino acids One-letter amino-acid codes excluded from sampling everywhere, e.g. AC excludes alanine and cysteine. X excludes the unknown residue.

Global amino-acid bias (JSON) Bias for each amino acid, e.g. {"A": -1.1, "F": 0.7}. Positive values favor an amino acid; negative values discourage it.

Per-position omitted amino acids (JSON) Chain-specific exclusions, e.g. {"A": [[[1, 2], "CP"], [[3], "W"]]}. Uses 1-based parsed positions. In this form: Enter a chain dictionary or an upstream dictionary keyed by structure name. Missing chains have no local exclusions.

Per-position amino-acid bias (JSON) Chain dictionary of L × 21 numeric bias matrices in alphabet order ACDEFGHIKLMNPQRSTVWYX. L is the parsed chain length. In this form: Upload JSON or provide an object through the API. An upstream dictionary keyed by structure name is also accepted. Missing chains receive zero bias.

PSSM guidance

Official documentation ↗

PSSM guidance (JSON) Chain dictionaries containing pssm_coef (L), pssm_bias (L × 21 probability distributions), and pssm_log_odds (L × 21). Alphabet: ACDEFGHIKLMNPQRSTVWYX. In this form: Use the output of the official make_pssm_input_dict.py helper, or a chain dictionary. The PSSM presets include converted official inputs.

PSSM mixing weight Global PSSM weight from 0 to 1, multiplied by each position’s pssm_coef. Zero uses ProteinMPNN alone; one with coefficient one uses PSSM alone.

Mix in PSSM probabilities Enable probability mixing using pssm_multi and per-position pssm_coef.

Restrict amino acids using PSSM log odds Apply the PSSM log-odds mask to sampling.

PSSM log-odds threshold Keep amino acids whose PSSM log odds are greater than this threshold when the log-odds option is enabled.

Outputs

Official documentation ↗

Condition only on the backbone For conditional probabilities, compute p(amino acid | backbone) instead of also conditioning on the other residues.

Save sampled sequence scores Save negative log probabilities for designed residues and global scores as NumPy .npz files.

Save sampled sequence probabilities Save per-position probabilities, log probabilities, sequences, masks and chain order as NumPy .npz files.

Suppress progress output Suppress ProteinMPNN’s progress messages. Result files are still saved.

These local CLI options are not applicable to Bio Web and are omitted from its inputs:

out_folder: The service creates a separate output directory for each job and archives the results.
path_to_model_weights: The model and soluble/CA-only controls select an installed official checkpoint.
jsonl_path: This page accepts one structure per job; folder-based official examples are split into presets per PDB.
chain_id_jsonl: pdb_path_chains supplies the designed chains; the other parsed chains are fixed context.

Ready