Protein fragment selection using machine learning

Abstract

Protein fragment selection is an important step in predicting the three-dimensional (3D) structure of proteins. Selecting the right fragments contributes significantly to accurate prediction of 3D structure. In this thesis, a machine learning approach is employed to predict whether a pair of protein fragments have similar 3D structures or not, which can be used to select fragment structures for a target protein with unknown structure. To design input features, a concepy hierarchy is implemented, which considers sequence profile matrices, predicted secondary structure, solvent accessibility and torsion angle classes as features in various combinations and projections. Several machine learning classifiers and regressors are trained and optimized for predicting the structural similarity of 3-mer and 9-mer fragments including logistic regression, AdaBoost, decision tree, k-nearest neighbor, naive Bayes, random forest, SVM and multi-layer perceptron. The results demonstrate that combining different feature sets through concept hierarcy and model optimization improves the prediction accuracy substantially. Furthermore it is possible to predict the structural similarity of fragment pairs with high accuracy, which is assessed by perforing cross-validation experiments on fragment datasets. When the structural similarity of fragments is defined as a classification problem, the accuracy of different classifiers are obtained as similar to each other. Among the regression methods, random forest provided the best accuracy metrics.