Predicting sulfotyrosine sites using the random forest algorithm with significantly improved prediction accuracy

Abstract

Background: Tyrosine sulfation is one of the most important posttranslational modifications. Due to its relevance to various disease developments, tyrosine sulfation has become the target for drug design. In order to facilitate efficient drug design, accurate prediction of sulfotyrosine sites is desirable. A predictor published seven years ago has been very successful with claimed prediction accuracy of 98%. However, it has a particularly low sensitivity when predicting sulfotyrosine sites in some newly sequenced proteins.

Results: A new approach has been developed for predicting sulfotyrosine sites using the random forest algorithm after a careful evaluation of seven machine learning algorithms. Peptides are formed by consecutive residues symmetrically flanking tyrosine sites. They are then encoded using an amino acid hydrophobicity scale. This new approach has increased the sensitivity by 22%, the specificity by 3%, and the total prediction accuracy by 10% compared with the previous predictor using the same blind data. Meanwhile, both negative and positive predictive powers have been increased by 9%. In addition, the random forest model has an excellent feature for ranking the residues flanking tyrosine sites, hence providing more information for further investigating the tyrosine sulfation mechanism. A web tool has been implemented at http://ecsb.ex.ac.uk/sulfotyrosine for public use.

Conclusion: The random forest algorithm is able to deliver a better model compared with the Hidden Markov Model, the support vector machine, artificial neural networks, and others for predicting sulfotyrosine sites. The success shows that the random forest algorithm together with an amino acid hydrophobicity scale encoding can be a good candidate for peptide classification.

Figure 1

RF ROC curves for the 10-mer, 20-mer and 30-mer data sets. The horizontal axes are the false alarm rates (1 - specificity) and vertical axes are the sensitivity. For specific threshold for discriminating between positive (true sulfotyrosine sites) and negative (unconfirmed sulfotyrosine sites) data points, there will be a pair of these two values, i.e., 1 - specificity and sensitivity. A pair of values is then represented by a point in this two-dimensional space. Each curve is made by connecting all these points. A model is said to be robust whether its ROC curve is close to the top left corner. The area under a ROC curve is a quantitative indicator of this robustness.

Figure 2

SVM ROC curves for the 10-mer, 20-mer and 30-mer data sets. The horizontal axes are the false alarm rates (1 - specificity) and vertical axes are the sensitivity. For specific threshold for discriminating between positive (true sulfotyrosine sites) and negative (unconfirmed sulfotyrosine sites) data points, there will be a pair of these two values, i.e., 1 - specificity and sensitivity. A pair of values is then represented by a point in this two-dimensional space. Each curve is made by connecting all these points. A model is said to be robust whether its ROC curve is close to the top left corner. The area under a ROC curve is a quantitative indicator of this robustness.

Figure 3

The correlation of the predictions between 10-mer model predictions (horizontal axis) and the 20-mer model predictions (vertical axis) for the blind data set.

Figure 4

The correlation of the predictions between 10-mer model predictions (horizontal axis) and the 30-mer model predictions (vertical axis) for the blind data set.

Figure 5

The correlation of the predictions between 20-mer model predictions (horizontal axis) and the 30-mer model predictions (vertical axis) for the blind data set.

Figure 6

The ranking results of residues in three RF models. The horizontal axis represents residue positions in peptides. The upper panel is for the 10-mer data, hence having residue positions ranging from N₅ to C₅. The middle panel is for the 20-mer data, hence 20 bars. The lower panel is for the 30-mer data, hence 30 bars. The vertical axis indicates the mean decrease Gini measures.

Figure 7

The conditional density functions drawn at N₁ and C₁ residues, respectively. The horizontal axes represent the Cornette scale values while the vertical axes represent the density values. The density functions are estimated using the kernel approach using the R stats package with default parameter setting. The graph shows that the density functions drawn at N₁ demonstrate a larger separation between two classes while this difference is getting smaller for the residue C₁, which does not have a high rank value from RF models. Note that negative means unconfirmed sulfotyrosine whilst positive means experimentally verified sulfotyrosine.