Inferring gene regulatory networks via adversarially regularized directed graph autoencoder

Revealing complete gene regulatory networks (GRNs) is important for a deeper understanding of biological processes. Although numerous GRN inference methods have been proposed, they are not only struggle to cope with the complex topology of GRNs, but also do not take into account the non-identical distribution property of gene expression data. To address the above challenges, We propose to infer GRNs via Adversarial Regularized Directed Graph Autoencoder (ARDGA). First, two structure matrices are computed based on adjacency matrix, which contain the first-order and second-order proximity to capture the complex topology of GRNs. Second, a novel message-passing module is developed by leveraging structure matrices. Based on this module, source and target encoders are deployed to learn the source and target vectors of each node in different neighborhoods, thereby aggregating enriched neighborhood information using first and second-order proximity. Third, in order to maintain the biostatistical property of the gene expression data, the target vectors are regularized to prior distribution of raw data by an adversarial training strategy. Finally, the source and target vectors are fed into the decoder to infer GRNs. Extensive experiments on the DREAM5 dataset and seven single-cell RNA sequencing (scRNA-seq) datasets with four types of ground-truth networks demonstrated that ARDGA outperforms recent strong baselines, achieving competitive results. The code is publicly available at: https://github.com/longkf/ARDGA.