Predict impact of single amino acid change upon protein structure

Abstract

Background: Amino acid point mutations (nsSNPs) may change protein structure and function. However, no method directly predicts the impact of mutations on structure. Here, we compare pairs of pentamers (five consecutive residues) that locally change protein three-dimensional structure (3D, RMSD>0.4Å) to those that do not alter structure (RMSD<0.2Å). Mutations that alter structure locally can be distinguished from those that do not through a machine-learning (logistic regression) method.

Results: The method achieved a rather high overall performance (AUC>0.79, two-state accuracy >72%). This discriminative power was particularly unexpected given the enormous structural variability of pentamers. Mutants for which our method predicted a change of structure were also enriched in terms of disrupting stability and function. Although distinguishing change and no change in structure, the new method overall failed to distinguish between mutants with and without effect on stability or function.

Conclusions: Local structural change can be predicted. Future work will have to establish how useful this new perspective on predicting the effect of nsSNPs will be in combination with other methods.

Figure 1

Structural and evolutionary features most predictive. Input features according to their cumulative contribution to performance measured by AUC, i.e. the area under the ROC curve (AUC* indicates that these values refer to results for a subset of the full cross-validation set). Our forward feature selection scheme suggested that three features raised performance above 0.8: evolutionary information (PSIC [31] diff), predicted secondary structure (from PROFsec [32,33]) around mutant (mutant position ± 8, i.e. 17 input units), and the PSI-BLAST information per residue for 21 consecutive residues. Additional six features only marginally increase performance up to mean AUC* ~0.84: predicted flexibility (PROFbval, w=21), difference in both PSI-BLAST PSSM (PSSM diff) and predicted secondary structure scores (PFOFsec diff), the fit of change position into a PFam domain (PFam fit, w=13), scores for predicted protein-protein interaction hotspots (ISIS, w=13) and residue volumes (VOLUME, w=5). High variability in AUC* distributions (long box plots, strong overlap between box plots) indicates instability in selected features.

Figure 2

Good discrimination between pentamers with and without effect. All values refer to full-cross validation averages of data not used for feature optimization. Left panel (A): our best model (solid line) reached an AUC of 0.8, compared to random predictions (dashed line) with AUC=0.5. Right panel (B): predictions for effect on structure (change) and predictions for no effect (neutral) reached similar levels on the accuracy vs. coverage plot equation (2).

Figure 3

Mutations predicted to affect structure often impact function. We investigated the subset of mutants predicted by our method to change structure from uncertain predictions (>0.5) to very strong predictions (>0.9). Residues in this subset, we found more often than expected to also have an observed effect in other data sets than used for our method, namely on protein stability (solid line) and protein function (dashed line). This suggests that strong structural change upon amino acid change results in increased likelihood to alter function or stability.

Figure 4

Correlation between structure and function not picked up by other methods. We applied three prediction methods to our dataset of structural effect: (A, D) the new method introduced here, (B, E) SNAP [17] predicting impact on function, and (C, F) I-Mutant3 [21] predicting the impact on stability. In lack of a better alternative, we chose the default threshold for each method (horizontal dashed lines) to distinguish neutral from effect. The method introduced here that is specialized to separate structural effect from neutral performs best at this task (A: little overlap between boxes; note: data in cross-validation mode of our method). The distributions from SNAP (functional effect prediction) and I-Mutant3 (stability prediction) both do not capture the structure signal.