Learn/Redesign a sequence
Read the designs
Each design is shown against the input: the strip is chain B, one column per residue, the changed residues in blue, so where a model changed things is seen before any number. The numbers are the model's own, named as it names them, and they rank designs within one model only.
ProteinMPNN
4 designs| Design | Chain B | Identity | score | global score | recovery | pLDDT |
|---|---|---|---|---|---|---|
| design 1 | 85% | 0.796 | 1.457 | 46% | 77.3 | |
| design 2 | 84% | 0.851 | 1.447 | 43% | 76.9 | |
| design 3 | 84% | 0.808 | 1.457 | 43% | 77.4 | |
| design 4 | 84% | 0.846 | 1.453 | 43% | 76.9 |
PottsMPNN
4 designs| Design | Chain B | Identity | score | global score | recovery | Potts energy | pLDDT |
|---|---|---|---|---|---|---|---|
| design 2 | 86% | -0.049 | -0.092 | 98% | 797.278 | 79.3 | |
| design 3 | 86% | -0.046 | -0.091 | 98% | 810.220 | 78.1 | |
| design 4 | 86% | 0.015 | -0.081 | 98% | 810.820 | 78.3 | |
| design 1 | 85% | -0.039 | -0.090 | 98% | 815.704 | 77.6 |
ESM-IF1
4 designs| Design | Chain B | Identity | score | global score | recovery | pLDDT |
|---|---|---|---|---|---|---|
| design 1 | 54% | 0.592 | 0.592 | 54% | 77.2 | |
| design 2 | 56% | 0.617 | 0.617 | 56% | 75.7 | |
| design 3 | 58% | 0.607 | 0.607 | 58% | 77.6 | |
| design 4 | 56% | 0.609 | 0.609 | 56% | 79.4 |
Proton-PottsMPNN
4 designs| Design | Chain B | Identity | Potts energy | selective energy | λ | pLDDT |
|---|---|---|---|---|---|---|
| design 1 | 74% | -171300.328 | -13.792 | 0.000 | 77.5 | |
| design 2 | 64% | -171197.563 | -26.130 | 0.333 | 76.6 | |
| design 3 | 26% | -170422.938 | -106.574 | 0.667 | 51.4 | |
| design 4 | 12% | -169815.328 | -120.578 | 1.000 | 54.8 |
Two sequences scored by ProteinMPNN
the input, and the input with three alanines in its third loop| Sequence | Chain B | Identity | score | global score |
|---|---|---|---|---|
| native | 100% | 1.284 | 1.492 | |
| cdr3_aaa | 98% | 1.348 | 1.507 |
Score mode asks the model how likely a sequence you give it is on this backbone, without designing anything. The native sequence and a variant are scored side by side; the variant is less likely to the model, as it should be, and that is all the number says.
Two sequences scored by ESM-IF1
the input, and the input with three alanines in its third loop| Sequence | Chain B | Identity | score | global score |
|---|---|---|---|---|
| native | 100% | 1.231 | 1.231 | |
| cdr3_aaa | 98% | 1.229 | 1.229 |
Score mode asks the model how likely a sequence you give it is on this backbone, without designing anything. The native sequence and a variant are scored side by side; the variant is less likely to the model, as it should be, and that is all the number says.
No source states a threshold on score, global score or recovery for choosing a sequence, and the lab draws none. Recovery is how much of the input a model kept, not how well the design works; a design that keeps less is not worse for it. What settles the question is the next step.
Read more: Phi docs: sequence scores · Dauparas et al. 2022, ProteinMPNN