Study On Protein Interaction Prediction Method Based On Graph Neural Network

Protein-protein Interactions are an important part of the cellular biochemical reaction network and the basis for understanding life activities.Computational methods are one of the popular methods for predicting protein interactions.But most computational methods rely too much on protein sequence information and do not consider the impact of false-negative and false-positive data.To overcome these problems,the following research work is carried out in this paper.(1)We propose a novel model for PPI prediction based on graph neural network,namely SLFVGA.First,SLFVGA introduces a symmetric non-negative latent feature model to extract the latent features from the PPI network,as this can reduce the negative impact of noisy data.To obtain high-quality feature information,SLFVGA uses principal component analysis to integrate the latent and sequence features of proteins.Then,from the PPI network structure and protein feature information,an efficient representation of the network is learned through variational graph autoencoding.Finally,a classifier is trained to predict protein interactions.The experimental results show that,compared with mainstream algorithms,SLFVGA has achieved superior performance in multiple evaluation indicators.(2)In order to make full use of the information in the PPI network,this paper proposes an adjacency matrix calculation method that combines protein sequence information,low-order and high-order neighbor node information.The method combines the idea of random walks and protein interaction theory in the computational process.On this basis,we propose the PASNVGA model for protein interaction prediction.PASNVGA further learns the PPI network representation from the obtained high-order adjacency matrix,and then applies a neural network to complete the prediction.Experimental results show that the high-order adjacency matrix obtained by our proposed method can further improve the prediction performance.