Mapping the stabilome: a novel computational method for classifying metabolic protein stability

Abstract

Background: The half-life of a protein is regulated by a range of system properties, including the abundance of components of the degradative machinery and protein modifiers. It is also influenced by protein-specific properties, such as a protein's structural make-up and interaction partners. New experimental techniques coupled with powerful data integration methods now enable us to not only investigate what features govern protein stability in general, but also to build models that identify what properties determine each protein's metabolic stability.

Results: In this work we present five groups of features useful for predicting protein stability: (1) post-translational modifications, (2) domain types, (3) structural disorder, (4) the identity of a protein's N-terminal residue and (5) amino acid sequence. We incorporate these features into a predictive model with promising accuracy. At a 20% false positive rate, the model exhibits an 80% true positive rate, outperforming the only previously proposed stability predictor. We also investigate the impact of N-terminal protein tagging as used to generate the data set, in particular the impact it may have on the measurements for secreted and transmembrane proteins; we train and test our model on a subset of the data with those proteins removed, and show that the model sustains high accuracy. Finally, we estimate system-wide metabolic stability by surveying the whole human proteome.

Conclusions: We describe a variety of protein features that are significantly over- or under-represented in stable and unstable proteins, including phosphorylation, acetylation and destabilizing N-terminal residues. Bayesian networks are ideal for combining these features into a predictive model with superior accuracy and transparency compared to the only other proposed stability predictor. Furthermore, our stability predictions of the human proteome will find application in the analysis of functionally related proteins, shedding new light on regulation by protein synthesis and degradation.

Figure 1

Heat-map and cluster analysis generated from the output of the Yen experiment [[1]]. A Euclidean distance metric was used for the cluster analysis. Bright (yellow) colouring in the heat-map represents a value approaching 1, with values approaching 0 having a dull (red) colouring. Labels on the right show how the data can naturally be broken up into groups of stable and unstable proteins, with the remainder being classed as “non-assigned”.

Figure 2

Graph representing BN + SVM model. The type of model parameters are indicated by conditional probability tables (CPT), noisy-OR (conditional probability table with a noisy-OR assumption) and Gaussian density tables (GDT) representing continuous values. For the sake of clarity, two continuous nodes that are children to the tyrosine and serine/threonine phosphorylation PTM nodes are not included in the graph. These continuous nodes contain the PWM scores for the sequence. The BN model is identical, but with “Sequence SVM” removed. The SVM model contained only the “Stability” node with the “Sequence SVM” node as a child.

Figure 3

Comparison of true positive and false positive rates for protein stability prediction models. Receiver operating characteristic (ROC) curves for the BN, SVM and BN+SVM models were calculated using 10-fold cross validation. We evaluated the three models on a total of 743 unstable proteins and 794 stable proteins.

Figure 4

Comparison with the IFS and NN classifier. The performance of the BN, SVM and BN+SVM models were compared against Huang and Colleagues’ [7] IFS plus NN method through calculation of ROC. The models were evaluated on a subset of 250 genes overlapping between our stable/unstable classes and the extra long/short classes as defined by Huang and colleagues [7].

Figure 5

Comparison of protein stability prediction models trained on a trimmed data set. The performance of the BN, SVM and BN+SVM model on a trimmed data set with secreted and transmembrane proteins removed. Due to the set containing a smaller number of samples (300 stable proteins and 227 unstable proteins), 25-fold cross validation was used to calculate the ROC curves.