Fig 1.
Pearson correlation coefficients between experimentally-determined and calculated values of changes in protein stability (ΔΔG) for PremPS performing three types of cross-validation (A) and leave-one-protein-out validation on S5296 (B), and tested on the dataset of Ssym (C) and S921 (D), respectively. PremPSM: the model was retrained after removing the overlapped mutations including their corresponding reverse mutations with the dataset of Ssym from the training dataset. See also S2 Fig and S1 and S4 Tables.
Table 1.
The performance of PremPSM tested on eight datasets.
Here, the PremPS model was retrained after removing the overlapped mutations including their corresponding reverse mutations with each test set from the training set. The number of overlapped mutations is provided in the S1C Table.
Table 2.
Comparison of methods’ performance on the independent test set of S921.
Fig 2.
Comparative performance of PremPS and four other methods of INPS3D, PoPMuSiC, FoldX, and mCSM on the independent test set of S921.
(A) ROC curves for predicting highly destabilizing (ΔΔGexp ≥ 1 kcal mol-1) and highly stabilizing mutations (ΔΔGexp ≤ -1 kcal mol-1). PremPS has substantially higher AUC-ROC than other methods (p-value < 0.01, DeLong test, S3B Fig). (B) Pearson correlation coefficients between predicted and experimental ΔΔG for mutations occurring in protein core and surface. The difference in R between PremPS and other methods is significant (p-value < 0.01, Hittner2003 test). More details are shown in S3 and S4 Figs.
Table 3.
PremPS’ performance on four datasets.
Leave-one-protein-out validation (CV4) results are shown for S2297 and RS2297. More details are provided in S9 Table.
Table 4.
Pearson correlation coefficient between experimental and predicted ΔΔG values calculated using experimental and modeled structures in different ranges of sequence identity.
Selected models: when several templates were available in a given range of sequence identity, the one whose sequence identity with the target was closest to the middle of the range and deviation from the experimental structure was the lowest was selected. Leave-one-protein-out validation (CV4) results are shown for S2297 and RS2297. More details are provided in S13 Table.
Fig 3.
Left corner: the entry page of PremPS server; right corner: the third step for selecting mutations and three options are provided: “Specify One or More Mutations Manually”, “Upload Mutation List” and “Alanine Scanning for Each Chain”, see also S9 Fig; and bottom: final results, see also S10 Fig. “Processing time” refers to the running time of a job without counting the waiting time in the queue. The contribution of each feature is provided in the download file.