Prediction of protein contact map by fully convolutional dilated residual network

Abstract

Residue contact maps contain important information for understanding the structure and function of proteins, thus contact map prediction is an important problem in the bioinformatics field. In recent years, deep learning has become a very popular tool in many research fields. However, the studies of using deep models to predict residue contact maps are very few. This work attempts to identify contact maps based on a recent breakthrough in deep learning, the residual network. The residual network distinguishes itself from other deep convolutional networks in that it incorporates a structure improvement called identity mapping to enable the neural network to go much deeper without consequent training difficulty. Moreover, dilated convolution is employed into this network to obtain a better performance by enlarging the receptive field of the network. A prediction is made based on the input features of a protein and all the features are input into the network at the same time. The experiments demonstrate that the dilated residual network outperforms the original residual network in contact map prediction. We test the networks on 3 test sets: CAMEO, CASP11 and a self-built independent test set. On top L/5 long-range contacts of the three test sets, the accuracy of the dilated network is higher than the non-dilated one by 5.2 %, 4.6% and 2.9%, respectively. Furthermore, it is confirmed that applying different networks on different features is a worse idea than taking them in together. The accuracy on top L/5 long-range contacts of the latter network is higher than the former one by 9.8% on the CAMEO set, 7.0% on the CASP11 set and 5.9% on the self-built independent test set.