Research On Function Prediction Method Of Protein Structural Domain Based On Deep Learning

Domain is the basic unit of protein function,and it has relatively stable spatial structure.A highly reliable method of functional annotation is the basis of understanding protein functions,so the function prediction of domains is quite critical.With the development of high-throughput modern molecular biology experiments,the number of protein chains with unknown functions is very large.Experimental functional identification of these sequences of unknown structural information and functional information is time-consuming and expensive,and experiments are usually conducted strictly on a few selected organisms.Therefore,the use of automatic functional computing(AFP)to predict the function of structural domains is a research focus of various research groups.In recent decades,more and more methods are mainly composed of machine learning methods,using similarity structure domain characteristics,the position specific scoring matrices used to predict the function of the protein structural domain.A new functional annotation method based on amino acid sequence of protein structural domain which is named as DeepDomGO was proposed by combining deep convolutional neural network model with deep cyclic neural network model and classifying the results predicted by deep learning model with hierarchical classifier.The main innovation of this method is that it only uses the amino acids sequence of structural domain of proteins and the features extracted from sequence as input of deep learning model,the target is to annotate the Gene Ontology term function of the structural domain,considering the local sequence chain amino acids and global interaction of amino acids.The results show that the accuracy of this function prediction method is high and at the same time,it also has universality.Research object of DeepDomGO is the structural domain of human protein,using the data of single structural domain data to train deep learning model parameters to get a best model,implements function prediction of the domain structure of human proteins(including single structure domain and structure domain),and use the same data set to evaluate multiple function prediction method,the Fmax values were 0.656,0.784,and 0.774,compared with DeepGO method whose predicted results are 0.435,0.503,0.639 and SDA method whose predicted results are 0.539,0.670,significantly improve performance of the function prediction of biological process,cell composition and molecular,and provide online service of querying the function of the single structural domain of proteins.