Research On Topological Prediction Of ?-helical Transmembrane Protein Based On Multi-scale Residual-BiLSTM Network

The research of protein structures prediction is of great significance for understanding protein function,promoting protein engineering and drug research and development.The transmembrane protein is a type of protein with a special structure.They penetrate the phospholipid bilayer through a special transmembrane structure and are stably fixed on the biofilm for a long time.They are the main undertakers of biofilm functions.Therefore,the research of transmembrane protein structures has very important biological and medical significance.According to the structure of the transmembrane region,transmembrane proteins can be divided into two categories: alpha helix and beta barrel.As an important class of transmembrane proteins,the research of alpha-helical transmembrane protein structures is of great significance to the functional research of transmembrane proteins.Because the number of transmembrane proteins with known structures is far behind the estimated number of these proteins in various genomes,it is particularly important to perform low-resolution topology prediction.Therefore,it is a new research hotspot to use the amino acid sequence information of transmembrane proteins and the computational biology method to predict the topological structure of transmembrane proteins.The accuracy of the existing methods for predicting alpha-helical transmembrane proteins begins to decline with the increase of the number of transmembrane proteins with known structures.Therefore,it is urgent to develop accurate prediction methods.With the development and maturity of deep learning algorithms,they have also made breakthroughs in the field of computational biology.Therefore,the use of deep learning algorithms to improve the performance of alpha-helical transmembrane protein topology prediction methods has very good research prospects.Although there are existing methods for predicting the topology of ?-helix transmembrane proteins using deep learning algorithms,there is still room for innovation and improvement in the stability of the method and the bio interpretability of the model.In this paper,a new method for predicting the topology of alpha-helical transmembrane proteins is proposed.This method uses the known protein sequence information to construct a prediction model of the alpha-helical transmembrane protein topology.In this method,multiscale residual network is used to extract local correlation features of transmembrane protein sequences at different scales.Based on the inherent advantages of recurrent neural network in processing and predicting sequence data,a bidirectional long short-term memory(BiLSTM)network is introduced in the field of transmembrane protein topology prediction to learn the global association information and potential long-term dependence in protein sequence,and finally the multi-scale residual BiLSTM network is constructed.At the same time,the attention mechanism and mask network layers are introduced into the network to strengthen key information and reduce the negative impact of padding,so as to enhance the classification learning ability of the model and improve the prediction performance of the model.The prediction method in this paper realizes the use of simple and effective HHblits profiles features to obtain the optimal experimental results.For the final prediction model,this paper will adopt strict and unified evaluation indicators as the measurement standard,and compare it with the other five newest alpha-helical transmembrane protein topology prediction methods,and the performance of the model is evaluated and analyzed from three levels: single residue,transmembrane helix fragment and overall topology.The comparison results show that the prediction method has better prediction performance than the previously mentioned methods,and can more accurately predict the topology of alpha-helical transmembrane proteins.Of course,there is still room for innovation and improvement in the aspects of biological logic in the prediction results of this method.