Research On Link Prediction Algorithms Of Biological Network Based On Network Representation Learning

The construction of biomolecular networks has become more feasible as well as the advent of the big-data era,and traditional data-mining tasks are gradually applied to those networks.The storage of a sizeable network means expensive costs,and it is difficult to do complex computing directly.Therefore,network representation learning has become an indispensable step in related data mining tasks.Restricted by biological experiments,genome-wide networks are always not completed.This paper aims to predict gene interactions and optimize gene networks,which mainly includes the following two works:(1)We propose a GAN-based network representation learning algorithm,AWGI,which combines Generative Adversarial Networks and random walk to reduce the impact of missing interactions on network embedding.AWGI aims to generate node sequences with high similarity to the real random walks,and then learns node embeddings from those generated sequences.The experimental results prove that AWGI is superior to other algorithms on network embedding,and improves the accuracy of the link-prediction task on multiple gene networks by 2%to 4%.(2)Gene networks are diverse in size and interaction types as well as vary greatly in different researches.Therefore,we propose a framework for the optimization of gene networks,NIHO,which integrates the neighbor information of multiple networks to infer gene interactions and achieves mutual optimization between networks.NIHO applies Convolutional Neural Networks(CNN)to extracting and fusing the network structural features and reconstructs networks in the low-dimensional space by matrix strategy,which forms an end-to-end way to learning gene representations.After a compact feature learning,NIHO calculates the weight of the association between genes,which set as the basis of inferring gene interactions.Finally,NIHO fills the gene network by adding those potential interactions.We accurately evaluate the performance of the gene network through"A Benchmark for Evaluating Networks Based on Gene Sets",and design different experiments to verify the effectiveness and scalability of NIHO.Finally,we apply AWGI to protein-protein interaction predictions.Experimental result shows that AWGI can independently predict 229 pairs of interacting-protein that have been verified compared to other methods.