Deep Learning Based Prediction Of Protein Residue Contact

Residue contact describes the spatial distance between a pair of residues in protein three-dimensional structure, the contacted residue-pairs play an important role on the stability of protein structure. The contact relationships of all residue-pairs in protein determine the two-dimensional topology of protein three-dimensional structure, so it is very important to obtain accurate contact relationships for the prediction of protein three-dimensional structure.The accuracy of the residue contact prediction, especially long range contact, has been very low. The main reason is that the high nonlinearity between residue features and residue contact, except that the serious imbalance of the proportion of positive over negative samples also decreases the generalization ability of model. This thesis studies a deep sequence model based on bidirectional recurrent neural network and a training algorithm to reduce the influence of the sample imbalance. Bidirectional recurrent neural network can not only receives variable-length protein sequence features, but also does not need to be specified the sliding window size when process residue features,which was required by regular shallow learning approaches. Our training algorithm can adjusts the proportion of positive over negative samples, as well as dynamically selects the samples as the input of classification layers.The deep neural network transforms the original features into high-level features through a large number of nonlinear transformations, which is suitable for residue contact prediction. However, the hyper-parameter selection will become a problem because of the depth of deep neural network. This thesis implements the parallel hyperparameter optimization for our deep sequence model based on Hyperopt framework.Through the rapid parallel search, we found out a model which achieves comparable results with the models that spent a lot of time on setting hyper-parameter manually.This model obtained the accuracy of middle range residue contact prediction better than other methods 10% on multiple benchmarks, and the results of long range residue contact prediction are competitive with current popular approaches.