GRNFormer: accurate gene regulatory network inference using graph transformer.

Area of Science:

• Computational biology
• Network biology
• Genomics

Background:

• Deciphering gene regulatory networks (GRNs) from single-cell transcriptomics data is challenging due to data sparsity and high dimensionality.
• Existing inference models often lack scalability and generalizability across different species, cell types, and platforms.

Purpose of the Study:

• To develop a generalizable graph transformer framework, GRNFormer, for accurate and scalable GRN inference.
• To overcome limitations of existing methods by not requiring cell-type annotations or prior regulatory information.

Main Methods:

• GRNFormer integrates a transformer-based gene expression encoder (Gene-Transcoder) with a variational graph autoencoder (GraViTAE) using pairwise attention.
• TF-Walker, a transcription factor-anchored subgraph sampling strategy, is employed to capture gene regulatory interactions from single-cell or bulk RNA-seq data.

Main Results:

• GRNFormer outperforms state-of-the-art methods in blind evaluations, achieving high Sampled AUROC and AUPRC values (0.90-0.98).
• The model successfully recovers known and novel regulatory networks, including pluripotency circuits in hESCs and immune cell modules in PBMCs.
• Achieved an average Sampled F1 score of 0.87-0.98, demonstrating robust performance.

Conclusions:

• GRNFormer provides a scalable, biologically interpretable, and transferable tool for GRN inference.
• The framework enables accurate GRN discovery across diverse datasets, cell types, and species, advancing network biology research.