A generic method for assignment of reliability scores applied to solvent accessibility predictions

Abstract

Background: Estimation of the reliability of specific real value predictions is nontrivial and the efficacy of this is often questionable. It is important to know if you can trust a given prediction and therefore the best methods associate a prediction with a reliability score or index. For discrete qualitative predictions, the reliability is conventionally estimated as the difference between output scores of selected classes. Such an approach is not feasible for methods that predict a biological feature as a single real value rather than a classification. As a solution to this challenge, we have implemented a method that predicts the relative surface accessibility of an amino acid and simultaneously predicts the reliability for each prediction, in the form of a Z-score.

Results: An ensemble of artificial neural networks has been trained on a set of experimentally solved protein structures to predict the relative exposure of the amino acids. The method assigns a reliability score to each surface accessibility prediction as an inherent part of the training process. This is in contrast to the most commonly used procedures where reliabilities are obtained by post-processing the output.

Conclusion: The performance of the neural networks was evaluated on a commonly used set of sequences known as the CB513 set. An overall Pearson's correlation coefficient of 0.72 was obtained, which is comparable to the performance of the currently best public available method, Real-SPINE. Both methods associate a reliability score with the individual predictions. However, our implementation of reliability scores in the form of a Z-score is shown to be the more informative measure for discriminating good predictions from bad ones in the entire range from completely buried to fully exposed amino acids. This is evident when comparing the Pearson's correlation coefficient for the upper 20% of predictions sorted according to reliability. For this subset, values of 0.79 and 0.74 are obtained using our and the compared method, respectively. This tendency is true for any selected subset.

Figure 1

Graphical overview of the method. Graphic overview of the method used in training of the primary and secondary neural networks. 'PSSM' is a Position-Specific Scoring Matrix. 'Sec. Structure' is the raw output from secondary structure predictions. 'Primary Networks' are an ensemble of artificial neural networks (ANN) and 'B/E Classification' is the raw buried/exposed output from these ANNs. 'Secondary Networks' are also an ensemble of ANNs, trained to predict the relative surface exposure of an amino acid. The last box shows output from the web server.

Figure 2

The average error as a function of the predicted reliability. The left panel shows NetSurfP Z-score versus mean error, and the right panel shows the consistency reliability score versus mean error.

Figure 3

Histogram of mean error as a function of predicted exposure values. The bars show the histogram for four groups of predictions with high and low reliabilities: "High R" and "low R" for the consistency method and "high Z" and "low Z" for the NetSurfP method, where "high" is the 50% most reliable predictions according to the chosen reliability score, and "low" is the 50% least reliable predictions.

Figure 4

Histogram of the number of predicted residues (A: Real-Spine and B: NetSurfP) as a function of the predicted relative exposure value for all residues in the CB511 data set at different cut-offs. The full line shows the calculated (measured) exposure distribution of the full set. The distribution of the 25%, 50%, 75% and 80% most reliably A: Real-Spine predicted residues according to consistency score, and B: NetSurfP predicted residues according to the Z-score, are also shown. Insert shows the number of predicted residues/all predictions in a given threshold as a function of the predicted RSA.

Figure 5

Reliability baseline and standard deviation fitting. The reliability is shown as a function of the predicted exposure for the Cull-1764 data set. In grey is shown the fitted reliability baseline and standard deviation. The insert shows the baseline corrected Z-scores as a function of the predicted surface exposure.