Classifying protein structures into folds by convolutional neural networks, distance maps, and persistent homology

The fold classification of a protein reveals valuable information about its shape and function. It is important to find a mapping between protein structures and their folds. There are numerous machine learning techniques to predict protein folds from 1-dimensional (1D) protein sequences, but there are few machine learning methods to directly class protein 3D (tertiary) structures into predefined folds (e.g. folds defined in the SCOP database). We develop a 2D-convolutional neural network to classify any protein structure into one of 1232 folds. We extract two classes of input features for each protein: residue-residue distance matrix and persistent homology images derived from 3D protein structures. Due to restrictions in computing resources, we sample every other point in the carbon alpha chain to generate a reduced distance map representation. We find that it does not lead to significant loss in accuracy. Using the distance matrix, we achieve an accuracy of 95.2% on the SCOP dataset. With persistence homology images of 100 x 100 resolution, we achieve an accuracy of 56% on SCOPe 2.07 dataset. Combining the two kinds of features further improves classification accuracy. The source code of our method (PRO3DCNN) is available at https://github.com/jianlin-cheng/PRO3DCNN.