Inferring gene regulatory networks via adversarially regularized directed graph autoencoder

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Understanding gene regulatory networks (GRNs) is crucial for deciphering biological processes.
• Existing GRN inference methods often fail to capture complex network topologies and the non-identical distribution of gene expression data.

Purpose of the Study:

• To propose a novel method, Adversarial Regularized Directed Graph Autoencoder (ARDGA), for accurate GRN inference.
• To address limitations of current methods in handling complex GRN structures and data properties.

Main Methods:

• Computed two structure matrices capturing first-order and second-order proximity for complex topology.
• Developed a message-passing module with source and target encoders to learn node representations.
• Employed adversarial training to regularize target vectors to the prior distribution of raw gene expression data.

Main Results:

• ARDGA demonstrated superior performance compared to strong baselines on the DREAM5 dataset.
• Evaluated on seven single-cell RNA sequencing (scRNA-seq) datasets, ARDGA achieved competitive results across different ground-truth networks.
• The method effectively integrates topological and statistical properties of gene expression data.

Conclusions:

• ARDGA offers a robust and effective approach for inferring gene regulatory networks.
• The method advances the field of GRN inference by addressing key challenges in network complexity and data distribution.
• Publicly available code facilitates further research and application.