Study On Protein Contact Map Prediction Based On Deep Neural Network

The protein contact map describes the spatial distance relationship between different residues in the protein and provides abundant information about protein structure.Therefore,protein contact map prediction is an important direction in the field of protein structure research,which is of great significance for a deep understanding of protein function and promotion of rational design of protein structure.Existing methods for protein contact map prediction can be broadly divided into structure-based methods and sequence-based methods.The structure-based methods rely on the protein template structure to predict contact maps,which makes it limited in practical applications,while the sequence-based methods can directly use protein sequences to predict contact maps and do not depend on the information of protein structure,which receives extensive attention in recent years.Most sequence-based methods use classical machine learning algorithms to predict protein contact maps.Although great progress has been made,it is difficult to effectively learn residue contact patterns in large amounts of protein data.However,the deep learning methods can learn high-level features from a large amount of training data automatically,and achieve expression of the data,which is suitable for protein contact map prediction.Therefore,in this thesis we propose a deep convolutional neural network method,RUcon,based on the residual network and U-Net for protein contact map prediction.Furthermore,we proposed a method based on a generative adversarial network called GANcon.The main contributions are as follows:1.We propose a novel deep convolutional neural network method,RUcon,by combining the residual network with the encoder and decoder in U-Net,which can effectively extract and fuse features of different levels and avoid the performance degradation caused by the increased network depth,so that it can fully capture the important information about residue contact and accurately predict the protein contact map.According to the prediction performance on multiple metrics,RUcon achieves accurate protein contact map prediction results,and its prediction performance is better than that of the residual network and U-Net alone.2.To further improve the prediction performance,we propose a generative adversarial network method called GANcon for protein contact map prediction by introducing an adversarial learning strategy into RUcon.We use RUcon as the generator network in GANcon for generating protein contact maps,and design a discriminator network to distinguish between generated and real protein contact maps.Through adversarial learning between the two networks,the generator network can produce protein contact maps that are highly similar to real protein contact maps.At the same time,to cope with the imbalance problem and take into account the symmetry of contact maps,we also propose a novel symmetrical focal(SF)loss in this study,which can further enhance the effectiveness of adversarial learning for better prediction results.The evaluation results show that the prediction performance of GANcon is greatly better than RUcon,and is also comparable to most of the existing methods.