alterlab-proteinmpnn

ProteinMPNN (fixed-backbone sequence design)

Overview

ProteinMPNN (Dauparas et al., Science 2022; dauparas/ProteinMPNN) solves the inverse-folding problem: given a protein backbone (a 3D structure with no or a placeholder sequence), it designs amino-acid sequences predicted to fold to that backbone. It is fast, robust, runs on CPU, and is the standard "sequence" step between backbone generation (alterlab-rfdiffusion) and structure validation (alterlab-alphafold).

When to Use This Skill

Use this skill when the user wants to:

• Inverse-fold a backbone PDB into one or more candidate sequences.
• Redesign only selected positions while fixing the rest (partial design).
• Enforce symmetry by tying residues/chains so homo-oligomers get identical sequences.
• Bias the amino-acid composition (e.g. avoid cysteines) or use the soluble model.
• Produce the sequence step of a design → fold → score loop.

Does NOT Trigger

Scenario	Use instead
Design a pocket/interface with a ligand, metal, or nucleic acid present	alterlab-ligandmpnn
Generate a new backbone (no starting structure)	alterlab-rfdiffusion
Refold a designed sequence to check it (validation)	alterlab-alphafold
Generative multimodal (sequence+structure+function) design	alterlab-esm

Core Capabilities

1. Basic inverse folding

# Parse the PDB(s), then design sequences (dauparas/ProteinMPNN), from the repo
# root. HS points at the repo's helper_scripts directory.
HS=./helper_scripts
python "$HS"/parse_multiple_chains.py --input_path=pdbs/ --output_path=parsed.jsonl
python protein_mpnn_run.py \
  --jsonl_path parsed.jsonl --out_folder out/ \
  --num_seq_per_target 8 --sampling_temp "0.1" --seed 37

# Single structure, no parsing step:
python protein_mpnn_run.py --pdb_path backbone.pdb --pdb_path_chains A \
  --out_folder out/ --num_seq_per_target 8 --sampling_temp "0.1"

Lower --sampling_temp (e.g. 0.1) gives conservative, high-confidence designs; higher temperatures increase diversity. Output FASTA headers carry the model score (lower = better) and sequence recovery.

2. Fixed positions and chains

Supply a fixed-positions spec (JSONL from make_fixed_positions_dict.py, in the repo's helper_scripts directory alongside the parser above) to keep catalytic/known residues while redesigning the rest, and assign_fixed_chains.py from the same directory to design only some chains. Add --use_soluble_model to load the soluble-only weights, and --ca_only for CA-only backbones (it switches to the CA model set).

3. Symmetry / tied positions

Tie positions across chains so a homo-oligomer receives one sequence applied symmetrically — essential for symmetric alterlab-rfdiffusion outputs.

4. Design → fold → score loop

The canonical de-novo pipeline:

1. Generate a backbone with alterlab-rfdiffusion.
2. Design sequences for it here (ProteinMPNN), sampling several per backbone.
3. Score by refolding each with alterlab-alphafold and accepting only self-consistent designs (returns to the target backbone with high pLDDT, low PAE).

Resources

• references/proteinmpnn_usage.md — install/pinning, helper-script inputs (fixed positions, tied chains, bias), the soluble model, temperature guidance, and loop integration. Loaded on demand.

Part of the AlterLab Academic Skills suite.