Table 3.
CALIBER (above line) and other models (below lines) Linear epitope prediction performance on the external test set. CALIBER is applied with different embedding methods: AA random embedding, ESM-2, Kidera embedding, Kidera embedding + biochemical properties, to other algorithms: BepiPred-3.0 and EpiDope. The run times of CALIBER in seconds per protein (averaged over 50 proteins from the test.
| AUC | BAC | MCC | PR-AUC | Time | |
|---|---|---|---|---|---|
| Random | 0.789 | 0.71 | 0.13 | 0.12 | 0.08 |
| Random up to 1023 | 0.764 | 0.69 | 0.14 | 0.13 | - |
| Kidera | 0.787 | 0.71 | 0.13 | 0.12 | 0.08 |
| Kidera+bio | 0.783 | 0.71 | 0.13 | 0.12 | 0.15 |
| ESM-2 | 0.768 | 0.7 | 0.14 | 0.13 | 14.34 |
| BepiPred-3.0 | 0.622 | 0.59 | 0.06 | 0.04 | - |
| EpiDope | 0.604 | 0.58 | 0.05 | 0.03 | - |