AnomalGRN: deciphering single-cell gene regulation network with graph anomaly detection.

Abstract

Background: Single-cell RNA sequencing (scRNA-seq) is now essential for cellular-level gene expression studies and deciphering complex gene regulatory mechanisms. Deep learning methods, when combined with scRNA-seq technology, transform gene regulation research into graph link prediction tasks. However, these methods struggle to mitigate the impact of noisy data in gene regulatory networks (GRNs) and address the significant imbalance between positive and negative links.

Results: Consequently, we introduce the AnomalGRN model, focusing on heterogeneity and sparsification to elucidate complex regulatory mechanisms within GRNs. Initially, we consider gene pairs as nodes to construct new networks, thereby converting gene regulation prediction into a node prediction task. Considering the imbalance between positive and negative links in GRNs, we further adapt this issue into a graph anomaly detection (GAD) task, marking the first application of anomaly detection to GRN analysis. Introducing the cosine metric rule enables the AnomalGRN model to differentiate between homogeneity and heterogeneity among nodes in the reconstructed GRNs. The adoption of graph structure sparsification technology reduces noisy data impact and optimizes node representation.

Conclusions: