Inferring Gene Regulatory Networks via Directed Graph Contrastive Representation Learning

Inferring Gene Regulatory Networks (GRNs) has been a key focus in computational biology. Many methods have emerged that utilize existing graph structures and node vectors to infer potential edges, among which contrastive learning methods have received widespread attention. However, current graph contrastive learning methods are not specifically designed for directed graphs, which limits their effectiveness in inferring directed GRNs. In this paper, we propose to infer GRNs based on Directed Graph Contrastive Representation Learning (DGCRL). Considering the topological properties of directed graphs, we elaborately design an augmentation function, reversing edges, for transforming GRNs into source and target views. The augmentation function provides different node contexts for message passing in different views. Based on the two views, two attention- based encoders are deployed to learn the source and target vectors for each node, respectively. The dual vectors can capture richer semantic information and asymmetric node relationships in the directed graph. As the model parameters are continuously optimized, the source and target node of the regulatory edge are indirectly pulled closer. This enables the identification of such a specific pattern using a decoder, thereby inferring GRNs via a link prediction approach. Experiments on the DREAM5 and seven single-cell RNA sequencing (scRNA-seq) datasets with four types of ground-truth networks show that DGCRL can achieve better performance, demonstrating its great potential for large-scale GRN inference. The code is available at https://github.com/longkf/DGCRL.