ProteinMPNN

Offers full backbone and CA-only models, plus soluble-protein-trained weights, enabling adaptation to different structural inputs and design scenarios like membrane or soluble proteins.

Supports fixing residues, tying positions for symmetry, adding amino acid biases, and incorporating PSSM profiles via JSONL files, allowing precise sequence customization as shown in the helper scripts.

Provides scores, probabilities, and uncertainty metrics like global_score and seq_recovery in outputs, enabling researchers to assess design quality and reliability.

Can generate multiple sequences per target with configurable sampling temperatures, facilitating exploration of diverse sequence variants, as indicated by the --num_seq_per_target flag.