Table 2.
Performance on other data sets*
| Method | LacI repressor | Lysozyme | HIV-1 protease | Melanocortin-4 receptor |
|---|---|---|---|---|
| Standard deviation | 3.3 | 3.7 | 3.2 | 3.5 |
| SIFT | 69.4 | 67.6 | 78.3 | 57.8 |
| PolyPhen | 68.7 | 57.9 | *** | 51.1 |
| SNAP | 70.7 | 70.0 | 68.5 | 71.1 |
| SNAPannotated | 72.7 | 73.2 | 72.3 | 75.5 |
| SNPs3D | ** | ** | ** | 62.2 |
*Data sets: LacI repressor—4041 mutants, Lysozyme—2015 mutants, HIV-1 protease—336 mutants, Melanocortin-4 receptor—45 mutants. Methods: as for Table 1, except for: SNPs3D a method predicting human SNP effects from known 3D structures or profiles when structure modeling is not possible (Melanocortin-4 mutations predicted from profiles). Measure: Percentages are reported for the overall two-state accuracy [Equation (2), last rightmost column in Table 1]. Highest values in each column is highlighted in bold letters. Note that the differences in precision in the reported values reflected differences in the underlying standard deviation [Equation (5)].
**SNPs3D is optimized for human proteins; predictions for non-human mutants were therefore not included here.
***PolyPhen made no predictions for HIV-1 protease.