Learn/Redesign a sequence
Design new sequences
Phi has four inverse folding models, each a job at the same price. The request has one shape for all of them: the structure, the chain to design, the positions to keep, how many sequences and how warm to sample. The model is one word in it.
The models
| Model | What it is | Takes positions |
|---|---|---|
| ProteinMPNN | Phi's production inverse folding model (Dauparas et al. 2022), soluble or vanilla weights; designs, scores sequences and scores every residue. | Yes |
| PottsMPNN | KeatingLab's ProteinMPNN with a Potts head: the same request and rows, plus the Potts energy of each sequence; soluble or vanilla, with or without MSA weights; scores residues too. | Yes |
| ESM-IF1 | Meta's inverse folding model (Hsu et al. 2022): designs one chain from its backbone, the partner chains as context, and scores a chain's sequence by its log-likelihood. | No, the whole chain |
| Proton-PottsMPNN | PottsMPNN with protonation states (Jacobsen et al. 2026): designs a chain for a pH switch, sweeping stability against selectivity; the whole chain, no positions. | No, the whole chain |
Four sequences at a sampling temperature of 0.1 is a check on a backbone, not an exploration; tens explore, and 0.2 to 0.3 wanders further from what the backbone prefers. The knowledge behind Phi records runs from four to hundreds for different purposes and no synthesis settles the count; say what the count is for.
Read more: Phi API: PottsMPNN · Phi API: ESM-IF1 · Hsu et al. 2022, ESM-IF1