GRDGNN: A directed graph neural network framework for multi-relational inference of gene regulatory networks

Inferring gene regulatory networks (GRNs) from gene expression data is a central challenge in systems biology. Graph neural networks (GNNs) offer a promising approach due to their ability to process graph-structured data. However, existing GNN methods for GRN inference often treat the problem as binary classification, limiting their ability to capture comprehensive regulatory relationships. This paper introduces two learning algorithms that utilize an end-to-end gene regulatory directed graph neural network (GRDGNN) schema for efficient inference of causal relationships in large-scale networks. These algorithms incorporate a directed graph neural network (DGNN) and a graph multi-classification task to identify explicit interactions between transcription factors (TFs) and target genes. The proposed approach consists of four key steps: (1) constructing a directed initial network using regression Pearson correlation and mutual information analysis, (2) extracting subgraphs of observed TF-gene pairs and applying a DGNN for information aggregation, (3) projecting the aggregated information into a low-dimensional space using graph pooling to generate graph representations of TF-gene pairs, and (4) classifying subgraphs using a multilayer perceptron (MLP) for link prediction and inference of explicit regulatory relationships. Evaluation of the DREAM5 microarray and scRNA-seq datasets demonstrates that our transductive and inductive learning methods can accurately and effectively infer explicit regulatory relationships compared to benchmark methods. These results demonstrate that the proposed GRDGNN schema exhibits strong generalization across species, data types, and modalities in cross-species, cross-data type, and cross-modality learning.