proteinmpnn-sequence-design

name	proteinmpnn-sequence-design
description	Design or score protein sequences for fixed backbones with ProteinMPNN, LigandMPNN, or SolubleMPNN. Use for backbone-conditioned design, ligand-context design, soluble-protein design, residue constraints, or side-chain packing.
license	MIT

ProteinMPNN Sequence Design

Gate

Ask once before cloning a pinned repository commit, downloading model weights, or using GPU compute. Follow $cx-compute-environment; do not automatically submit generated sequences for synthesis or remote screening.

Workflow

1. Audit backbone source, assembly, chains, missing residues, ligand/cofactor atoms, fixed/designable positions, symmetry/tied residues, and target context.
2. Select ProteinMPNN, LigandMPNN, or SolubleMPNN for the declared objective. Pin commit, checkpoint checksum, noise level, temperature, seed, batch size, residue constraints, biases, and ligand/side-chain context flags.
3. Smoke-test an upstream example, then generate and score into artifacts/<run-id>/proteinmpnn/. Preserve structures, configs, FASTA/PDB, per-position probabilities, scores, atom/residue maps, and all candidates.
4. Report recovery only as a diagnostic. Evaluate diversity, novelty to training and known sequences, structural consistency, clashes, ligand-context dependence, and developability proxies with orthogonal methods.
5. Record the candidate funnel with $science-provenance; review with $science-review before selecting candidates.

Boundaries

• Model likelihood, recovery, or confidence is not function, binding, solubility, safety, or experimental validation.
• Prevent template/parent leakage and report similarity to the input sequence.