Join
ProteinGym mutation effects
Predict the fitness effects of mutations across 217 deep mutational scans (placeholder)
| # | Spearman ρ ↑ | Step | What changed |
|---|---|---|---|
| 7 | 0.4789 | ↑ 0.0021 | A later commit on this route |
| 6 | 0.4768 | ↑ 0.0035 | Aggregate the models’ ranks, with the structure route merged in |
| 5 | 0.4733 | ↑ 0.0044 | Calibrate the scores separately for each functional category |
| 4 | 0.4689 | ↑ 0.0018 | Add EVmutation-style coevolution scores from the MSAs |
| 3 | 0.4671 | ↑ 0.0023 | Encode the local 3D contact map around each site |
| 2 | 0.4648 | ↑ 0.0036 | Add EVmutation-style coevolution scores from the MSAs |
| 1 | 0.4612 | ↑ 0.0056 | Average the log-likelihood ratios of three protein language models |
| 0 | 0.4556 | The baseline, S2F |
