ai4protein ProteinMPNN Design Workflow
Overview
Orchestrates VenusFactory denovo tools for backbone-conditioned sequence design and scoring. Residue indices in fixed_residues_json are 1-indexed.
Project Tools (VenusFactory2)
| Tool | Args | When |
|---|---|---|
| proteinmpnn_sequence_design_from_structure | pdb_path; optional designed_chains, fixed_chains, fixed_residues_json, homomer, num_sequences, temperatures, out_dir |
Design sequences |
| proteinmpnn_sequence_scoring_from_structure | pdb_path; optional fasta_path, designed_chains, out_dir |
Score designs / native |
| pdb_chain_sequences / check_pdb_apo | PDB path | Prep / inspect |
| zero_shot_mutation_structure_prediction | designed PDB context | Optional post-design ranking |
Workflow patterns
| Goal | Key args |
|---|---|
| Single-chain redesign | pdb_path only |
| Design chain B against fixed A | designed_chains=["B"], fixed_chains=["A"] |
| Keep catalytic residues | fixed_residues_json='{"A":[57,102,195]}' |
| Homomer | designed_chains=["A","B","C"], homomer=true |
- Validate PDB (
check_pdb_apoif ligand-free backbone expected). - Design with explicit
out_dir. - Score designs with
proteinmpnn_sequence_scoring_from_structure. - Optional: structure prediction / property checks on top sequences.
Common mistakes
- Passing nearly all residues as fixed when intending to redesign the whole chain (or the inverse) — check tool log warnings.
- 0-based residue indices in
fixed_residues_json. - Expecting affinity/ΔG; ProteinMPNN gives sequence likelihoods, not binding free energy.