PATH - Prediction of Amyloidogenicity by Threading and Machine Learning

Abstract

Amyloids are protein aggregates observed in several diseases, for example in Alzheimer's and Parkinson's diseases. An aggregate has a very regular beta structure with a tightly packed core, which spontaneously assumes a steric zipper form. Experimental methods enable studying such peptides, however they are tedious and costly, therefore inappropriate for genomewide studies. Several bioinformatic methods have been proposed to evaluate protein propensity to form an amyloid. However, the knowledge of aggregate structures is usually not taken into account. We propose PATH (Prediction of Amyloidogenicity by THreading) - a novel structure-based method for predicting amyloidogenicity and show that involving available structures of amyloidogenic fragments enhances classification performance. Experimental aggregate structures were used in templatebased modeling to recognize the most stable representative structural class of a query peptide. Several machine learning methods were then applied on the structural models, using their energy terms. Finally, we identified the most important terms in classification of amyloidogenic peptides. The proposed method outperforms most of the currently available methods for predicting amyloidogenicity, with its area under ROC curve equal to 0.876. Furthermore, the method gave insight into significance of selected structural features and the potentially most stable structural class of a peptide fragment if subjected to crystallization.

Figure 1

Prediction procedure. Using comparative modeling, query sequence was threaded into seven templates representing different structural classes (class numbering in accordance with). For each of them ten models were proposed and the model with the lowest DOPE score was selected. For this model REF15 and selected PyRosetta energy terms were calculated and used as an input for machine learning classifiers.

Figure 2

DOPE score of amyloidogenic and non-amyloidogenic sequences threaded onto steric zipper structures for the benchmark set.

Figure 3

ROC curves for compared machine learning methods, for two test sets.

Figure 4

Coefficients of logistic regression model for energy terms.

Figure 5

Mean z-scores for energy terms obtained from Boruta feature selection method.