| Enzymes are closely related to the biochemical,metabolic,and disease processes of organisms.Studying the functional classification of enzymes is of great significance in disease diagnosis,production and daily life.Faced with the rapid increase in enzyme data,traditional techniques for identifying enzyme functions in biological experiments are difficult to meet practical needs due to their low efficiency and high cost.With the deep development of machine learning,more and more enzyme function classification methods based on computational technology have emerged one after another.Although most enzyme classification algorithms have achieved good results,there is still significant room for improvement,mainly manifested in: 1)Current computational methods are not yet able to fully explore the sequence,structure,and other information of enzymes;2)Further research is needed on how to efficiently combine features from different perspectives for classification.In response to the above challenges,this article proposes three new enzyme function prediction methods from the perspectives of multi perspective features,convolution,and self attention,and conducts system design and implementation.At the same time,we also explore future improvements.The main research content and contributions of this article can be summarized in the following four aspects:(1)In order to mine the sequence and structural characteristics of enzymes for new enzyme classification methods,a new method BP-TSK(BBCNet,Pointnet++and TSK)for enzyme function prediction is proposed,which takes both molecular structure and sequence information into account,extracts in-depth features,and uses multi view classification.Firstly,in order to effectively extract discriminative information from enzyme sequence data,we propose a sequence learning deep neural network that integrates BBA(Binding Bi-LSTM and selfAttention)residual module with Bio-CS(Biological Channel and Spatial Attention)attention mechanism to extract the sequence depth features of enzymes;Secondly,we innovatively introduce point clouds to express the three-dimensional structural information of enzymes,and use point cloud deep learning networks to extract the deep structural features of enzymes;Finally,the deep features from both sequence and structure perspectives are input into the multi view TSK fuzzy system to construct a multi-view classifier based on fuzzy rules and fuzzy reasoning,thereby achieving comprehensive enzyme function decision-making from multiple perspectives.The proposed model has been compared with various representative methods and achieved the highest accuracy and precision on the Protein Data Bank(PDB)protein dataset.(2)On the basis of the previous work,BP-TSK further improved its feature extraction ability in sequence,and proposed a new enzyme functional classification model CCANet(Convolutional Combine Attention Network)that combines convolution and self attention mechanisms in multiple layers.This model uses the one hot sequence of enzymes as input,and uses a multi-layer combination of convolution and self attention for feature extraction and functional prediction.Convolutional and self attention are usually considered two different types of representation learning.In this paper,we investigate the intrinsic relationship between them and propose a new method of combining convolution and self attention,SESA,using this connection.Through experimental verification,our method has achieved high-quality enzyme function prediction results.(3)In order to combine the advantages of the above two methods,we further propose a decision level method for enzyme function classification,BCP-Net(BBCNet,CCANet and Pointnet++),based on the sequence and structural information of enzymes.This method consists of three different networks in parallel.We conducted weighted fusion in the decisionmaking section and used ensemble learning classifiers for prediction,resulting in a more significant performance improvement compared to the two previously proposed methods.(4)In order to better apply our algorithm,we have designed and implemented two algorithms,CCANet and BCP-Net,and passed corresponding tests. |
