Feature Extraction Based On Multi Network Representation Learning And Its Application In Gene Function Prediction

Predicting the function of genes is one of the fundamental tasks in the postgenomic era.Although the functions of a large number of genes have been revealed,the functions of a large proportion remain unknown.In recent years,with the development of highthroughput technologies and the progressive discovery of the mechanism of interaction between genes,the massive amount of gene network data has been born.Alternative biological wet experiments integrating existing massive gene network data to predict the function of genes has become a new trend.In studies integrating gene network data for gene function prediction,many directly define models that run on the original graph adjacency matrix,often with insufficient representation of structural information and high computational overhead.Moreover,due to the flaws of high-throughput experimental methods themselves and the limitation of experimental conditions,the resulting network data usually contains a large number of false positives and false negatives.Therefore,to reduce the noise information in gene networks,it is important to integrate multi network data for gene function prediction.Based on this,in-depth research and analysis of network data feature extraction and integrated multi network algorithm in gene function prediction was carried out in this study,and the main work is summarized as follows:1.Aiming at the problem that the adjacency matrix is insufficient to represent the characteristic information of gene network topology,a gene network embedding method is proposed.This method uses a semi-supervised self-encoder to capture the topology information of the network,a supervised part learning the direct distance between nodes preserves the local structural information of the network,and an unsupervised part learning the indirect distance between nodes preserves the global structural information.This makes the extracted network structure information more complete,and the embedded vectors are also better able to retain the information of the original network.It was experimentally demonstrated that this algorithm can obtain a richer representation of gene nodes and improve the effect of functional prediction compared with previous algorithms.2.Aiming at the problem of multiple gene network integration,this paper proposes to embed multichannel attention mechanisms in graph convolutional neural networks to integrate the feature information of multiple networks,learning the most important low dimensional vector representation features of multiple networks.Harnessing the complementarity between multi-network data can effectively slow down the noise problem of single network data.Multi-network feature fusion effectively improved prediction efficacy compared to single gene network functional prediction through experimental validation.