Research On Intelligent Prediction Methods For Multi-Level Enzyme Function

As a class of extremely significant of biocatalysts,enzymes play an important role in the process of biological reproduction,metabolism,aging,and death.Therefore,the prediction of enzyme function is of great significance in biomedicine fields.However,traditional enzyme function prediction by biological experiments is extremely time-consuming and costly.In recent years,some computational methods using machine learning for predicting enzyme function have been proposed,which effectively reduces the cost of enzyme function prediction.Most of these methods have made significant progress by mining the multi-view features of enzymes and constructing corresponding multi-view intelligent models,but there are still the following deficiencies:(a)Features of different views have different effects on enzyme function prediction,while existing methods usually treat the information contained in each view equally and fail to identify the importance of different views;(b)The common information of different views,which is very important for the prediction of enzyme function,has not been fully mined and utilized in the existing methods;(c)Most current methods ignore correlation information between enzyme classes.This paper discusses the above problems,proposes two new methods,and develops a corresponding online service platform based on the proposed methods.Specifically,the work of this paper includes the following aspects:(1)Aiming at the shortcomings of treating views equally and failing to mine interaction information,a multi-level enzyme function prediction method based on multi-view deep interactive learning is proposed.The algorithm is a multi-view deep interactive network(MVDINET)based on parallel architecture to predict enzyme function.The algorithm first converts the enzyme sequence into a combined view of solvent accessibility and secondary structure,a view containing evolutionary information(PSSM),and a structural domain view,thereby constructing the initial multi-view feature data;Then the above three original views are input into various deep specific network modules to get the depth-specific information of each view;Further,a deep view interaction network is designed to extract the interaction information between views;finally,the specific information of each view and their interaction information are imported into a multi-view adaptively loss weighted classifier module for collaborative training,so as to train the final prediction model.Experimental research shows that the prediction accuracy of the proposed new method,MVDINET,has been greatly improved compared with the existing methods.(2)Aiming at the above shortcomings of treating views equally,failing to mine interaction information,and ignoring category-related information,a new enzyme function prediction method is proposed based on multi-view information bottleneck and dynamic category graph convolution network(GCN).The algorithm first converts the enzyme sequence into a PSSM view containing evolutionary information,a One Hot view of sequence position information,and a structural domain view.Then it uses a hybrid network to extract deep features for each initial view,including a deep convolutional network and BLi STM with an attention mechanism and a multi-layer perceptron network.Then it uses the multi-view information bottleneck to mine the common information among multiple deep feature views;Further,in order to obtain the correlation information between categories to assist classification,this method inputs multi-view information into the dynamic category GCN for category correlation information extraction;Finally,the multi-view collaborative classification module to fuse the classification results from multiple views.Experiments show that the prediction accuracy of the enzyme function model fused with multi-view information bottleneck and dynamic category GCN is significantly improved compared with previous methods.(3)Based on the above two works,an online service platform for multi-level enzyme function prediction was developed.Firstly,a detailed demand analysis of the platform is carried out to determine the functional modules of the system.Then it is developed based on the Django framework,providing functions such as sequence function prediction,prediction record,and famous sequences.Combining the algorithm with the platform provides certain data references and theoretical support for future enzyme function prediction.