Protein design on fixed backbone: ProteinMPNN & SolubleMPNN

This tool is powered by the high-performance graph neural network algorithm ProteinMPNN and is built for structure-guided amino acid sequence design from protein backbones. It combines strong native-structure recovery with fast computation, enabling efficient de novo sequence redesign, native-backbone reconstruction, and other molecular design tasks across enzyme engineering, antigen optimization, scaffold protein redesign, and protein complex interface design.

Upload a protein structure file in PDB or mmCIF format and the system will automatically parse the chain information. After selecting the chains to design and setting the parameters, the model will generate multiple optimized amino acid sequences together with a score for each sequence (score = -log_prob, lower is better) and sequence recovery.

The model is based on ProteinMPNN (Dauparas et al., Science 2022) and supports the standard model, the CA-only model, and the soluble-protein specialized variant (SolubleMPNN).

1. Upload a protein structure file (PDB / mmCIF):

3. Design parameters:

Model:
Number of sequences:		Generate 1-10 sequences per target
Sampling temperature:		Enter a single temperature value. Higher values increase sequence diversity.
Backbone noise:		Add Gaussian noise to backbone coordinates (Å) to increase generation diversity
Random seed:		Use 0 for a random seed each run; use a non-zero integer for a fixed seed
Excluded amino acids:		Checked amino acids will be excluded from generated sequences

4. Advanced options:

CA-only mode	Use cases: 1) structures with only C-alpha coordinates available, such as low-resolution cryo-EM, sparse NMR restraints, or de novo design backbones; 2) structures with missing or incomplete side-chain or backbone atom coordinates. This mode reads only C-alpha atoms and ignores other backbone atoms such as N/C/O. Note: it is not currently supported together with the soluble model.
Soluble model	Use cases: 1) designing soluble proteins (non-transmembrane proteins); 2) reducing the likelihood of membrane-protein-like sequence features; 3) applications with higher requirements for final solubility and expression. Membrane protein structures were excluded during training, so generated sequences are biased toward the amino acid distribution of soluble proteins. Model: SolubleMPNN
Maximum sequence length:		Residues beyond this length will be truncated

5. Constraints and biases (JSONL, optional):

The JSONL inputs below provide finer control than the UI fields above. If provided, they override the corresponding chain selection / excluded AA settings above.

Chain design specification - overrides the chain selection above (chain_id_jsonl):	Chain-specific excluded AAs - overrides the exclusion list above (omit_AA_jsonl):
Fixed positions (fixed_positions_jsonl):	AA composition bias (bias_AA_jsonl):
Residue-position bias (bias_by_res_jsonl):	Symmetric / tied positions (tied_positions_jsonl):

6. PSSM options (Position-Specific Scoring Matrix, optional):

PSSM JSONL (pssm_jsonl):
PSSM weight (pssm_multi):		0.0 = do not use PSSM, 1.0 = ignore MPNN predictions
PSSM threshold (pssm_threshold):		Restrict the selectable AA range at each position
Use log-odds (pssm_log_odds_flag)	PSSM bias mode (pssm_bias_flag)

About ProteinMPNN

ProteinMPNN (Dauparas et al., Science 2022) is a protein sequence design method based on a message-passing neural network.

Key features:
  - Uses protein backbone coordinates (N, CA, C, O, and related atoms) as input
  - Uses an encoder-decoder architecture to pass information over a 3D structural graph
  - Autoregressively samples amino acid sequences residue by residue
  - Supports multiple noise levels and model variants
  - Allows explicit design chains, context-only chains, and fixed positions
  - Supports sequence bias constraints, excluded amino acids, and related controls

Available models:
  v_48_002 - training noise 0.02A, 48 edges (most conservative; recommended for high-precision structures)
  v_48_010 - training noise 0.10A, 48 edges
  v_48_020 - training noise 0.20A, 48 edges (recommended default)
  v_48_030 - training noise 0.30A, 48 edges (more tolerant of low-resolution / NMR structures)

Reference: Dauparas, J., et al. "Robust deep learning-based protein sequence design using ProteinMPNN." Science 378.6615 (2022): 49-56.