Learn/Redesign a sequence
Compare the models
Four models designed on the same backbone. Their scores are in their own units and are not compared; what is compared is how far each moved from the input, the best of each model's own number, and, once folded, how confidently each design came back.
The same backbone, four models
| Model | Designs | Identity to the input | Its own best | Refolds, mean pLDDT |
|---|---|---|---|---|
| ProteinMPNN | 4 | 84 to 85% | score 0.796 | 76.9 to 77.4 |
| PottsMPNN | 4 | 85 to 86% | score -0.049 | 77.6 to 79.3 |
| ESM-IF1 | 4 | 54 to 58% | score 0.592 | 75.7 to 79.4 |
| Proton-PottsMPNN | 4 | 12 to 74% | Potts energy -171300.328 | 51.4 to 77.5 |
The input's sequence, scored by two models
| Model | Mean native fit | Residues scored |
|---|---|---|
| ProteinMPNN | -1.37 | 761 |
| PottsMPNN | -1.30 | 761 |
Two models score the same sequence on the same backbone, and the per-residue views in the lab can be switched between them. Where both fit a residue poorly is a stronger signal than where one does.
ESM-IF1 designs the whole chain, not the marked loops, so its identity to the input runs lower by construction; read its row with that in mind. In the lab, Expert chooses the designer on the Design step and the residue scorer on the Residues step; Simple sees ProteinMPNN alone.
Read more: Phi API: inverse folding models · Phi docs: sequence scores