Topology Structure Prediction Of Outer Membrane Protein Based On Deep Learning

Outer Membrane Protein(OMP)is an important class of transmembrane proteins,which has prominent significance for antibiotic resistance,vaccine design and cancer diagnosis and treatment.Because the outer membrane proteins are located in a special environment of biological outer membrane,leading to the inability of biological experimental techniques to analyze the protein structures on a large scale and explore the biological functions determined by structure.Based on the current conditions,it has become an important research approach in the field to explore the prediction method of structural characteristics of outer membrane protein by computational means to overcome the problem of small sample quantity.Topological structure is a representative structural feature of the outer membrane protein family and its function.The objective of topological prediction is to predict the transmembrane and non-transmembrane sequence segments alternating from the sequence of outer membrane protein residues by computational methods.Previous studies had tried statistical analysis,machine learning and other methods to try topological structure of outer membrane protein,but it still failed to achieve ideal prediction accuracy under the condition of limited sample number,and fell into a bottleneck in recent years.However,our preliminary in-depth study showed that the use of significant structural patterns of outer membrane proteins,the gradual increase in the number of samples,and the enhancement of end-to-end learning ability of deep learning methods would become a favorable breakthrough to improve the prediction accuracy of outer membrane protein topological structure.Therefore,based on the above favorable factors,this paper explored an effective deep learning method to improve the prediction accuracy of outer membrane protein topological structure,and compared and analyzed the deep neural network model applicable to sequence mode in the research work experiment.Firstly,TopOMPcapsnet,the prediction model of outer membrane protein topology based on capsule network was established,and the improvement performance of capsule vector on small sample data set was utilized to train the prediction model,and the comparative experiments were conducted on the effectiveness of multiple outer membrane protein a priori sequence features.The model achieved good recognition effect in the cross validation of outer membrane protein,and the prediction accuracy of three-state accuracy(Q3)reached 87.7%,significantly better than the traditional machine learning prediction method.The prediction accuracy of outer membrane protein topological structure was further improved by establishing TopOMP-deep HMM,a deep hidden Markov prediction model integrating deep learning and traditional machine learning.By choosing the bidirectional long and short-term memory network,which is good at handling temporal sequence problems,and combining with the hidden Markov model,deep learning can better abstract the features of outer membrane proteins,and transform the output of the bidirectional long and short-term memory network into the observation probability matrix into the hidden Markov model.The Q3 index was 89.6%,and the segment overlap fraction(SOV)was 92.6%,which achieved the optimal prediction performance compared with the current methods in the same field.This paper shows that the topological prediction of outer membrane proteins is not a small sample problem in the absolute sense because of the clear sequence pattern.Deep learning technology can recognize corresponding sequence patterns under the condition of limited samples,which has certain reference value for the research of this kind of problem.The research results can promote the large-scale structural and functional classification and screening based on outer membrane protein sequences,and contribute to the development of outer membrane protein related research.