Prediction Of Arabidopsis Regulatory Network Based On Graph Nerual Network

Gene regulatory networks(GRN)are composed of regulatory relationships between genes and biology objects such as proteins and RNA.There are many aspects to describe the elements in GRN,including gene functions.Here we describe a flexible framework based on embedding methods that can integrate different types of data for studying their joint effects on gene expression,which is named the Gene Information Aggregation Network(GIAN).GIAN has four steps.Firstly,GIAN determines the gene's neighborhood based on random walk with restart(RWR).And then bidirectional long-and short-term memory(Bi-LSTM)is applied to aggregate the different features of one gene.And then neigh-borhood gene features are aggregated by Bi-LSTM based on the type of genes.After that,an attention layer is set to aggregate all the features.We tested the ability of GIAN to ensemble gene features using k-means clustering.The results showed that GIAN can well aggregate gene expression data and ontology annotations.And then we demonstrate the application of GIAN in integrating several diverse types of data including gene expression,gene ontology annotations,transcript factors'protein sequences,and their DNA binding motifs,and finally,a GRN would be created by GIAN.Compared to EXPLICIT,the algorithm proposed by us before,GIAN can predict 36%more genes than EXPLICIT.GIAN can be used to explore unknown genes.