Highly Accurate Fragment Library for Protein Fold Recognition

Proteins play a crucial role in living organisms as they perform many vital tasks in every living cell. Knowledge of protein folding has a deep impact on understanding the heterogeneity and molecular functions of proteins. Such information leads to crucial advances in drug design and disease understanding. Fold recognition is a key step in the protein structure discovery process, especially when traditional computational methods fail to yield convincing structural homologies. In this work, we present a new protein fold recognition approach using machine learning and data mining methodologies. First, we identify a protein structural fragment library (Frag-K) composed of a set of backbone fragments ranging from 4 to 20 residues as the structural “keywords” that can effectively distinguish between major protein folds. We firstly apply randomized spectral clustering and random forest algorithms to construct representative and sensitive protein fragment libraries from a large-scale of high-quality, non-homologous protein structures available in PDB. We analyze the impacts of clustering cut-offs on the performance of the fragment libraries. Then, the Frag-K fragments are employed as structural features to classify protein structures in major protein folds defined by SCOP (Structural Classification of Proteins). Our results show that a structural dictionary with ~400 4- to 20-residue Frag-K fragments is capable of classifying major SCOP folds with high accuracy. Then, based on Frag-k, we design a novel deep learning architecture, so-called DeepFrag-k, which identifies fold discriminative features to improve the accuracy of protein fold recognition. DeepFrag-k is composed of two stages: the first stage employs a multimodal Deep Belief Network (DBN) to predict the potential structural fragments given a sequence, represented as a fragment vector, and then the second stage uses a deep convolution neural network (CNN) to classify the fragment vectors into the corresponding folds. Our results show that DeepFrag-k yields 92.98% accuracy in predicting the top-100 most popular fragments, which can be used to generate discriminative fragment feature vectors to improve protein fold recognition