An Efficient Inference Schema for Gene Regulatory Networks using Directed Graph Neural Networks

Inferring gene regulatory networks (GRNs) from gene expression data has remained the computational challenge due to the large-scale number of genes and the complexity of expression data in systems biology, which may often formulate a reconstruction problem among nodes for a topology graph. Graph neural network (GNN) is one of the most promising approaches to reconstructing a GRN by integrating topological neighbor propagation throughout a gene network. This paper proposes an end-to-end gene regulatory directed graph neural network (GRDGNN) schema to infer GRN from scratch using gene expression data in a supervised framework. Specifically, the regulatory relationship of the GRN can be first described as a graph multi-classification problem to distinguish the connection kinds between two nodes for subgraphs. Then, using gene dominant expression features and graph embedding node features, subgraphs consisting of gene node pairs and their neighbor are classified into four classes by a directed GNN model. In addition, a starting network structure constructed with noise from partial information can guide GRN inference through an appropriate integration approach. Finally, we demonstrate the ability of this method using the test data from DREAM5 challenge and the human embryonic stem cells (hESC) and human mature hepatocytes (hHep) datasets. Computational results show that the proposed method displays greater inference performance.