GRANet: a graph residual attention network for gene regulatory network inference

Area of Science:

• Genomics
• Systems Biology
• Computational Biology

Background:

• Gene regulatory network (GRN) reconstruction is vital for understanding gene expression mechanisms.
• Single-cell RNA sequencing (scRNA-seq) enables GRN inference at the single-cell level.
• Existing global GRN models face accuracy limitations due to network scale, noise, and data sparsity.

Purpose of the Study:

• To develop a novel deep learning framework, GRANet (Graph Residual Attention Network), for accurate GRN inference.
• To leverage residual attention mechanisms and multi-dimensional biological features for enhanced GRN reconstruction.
• To overcome the limitations of existing global GRN inference methods.

Main Methods:

• Developed GRANet, a deep learning framework utilizing residual attention mechanisms.
• Integrated multi-dimensional biological features into the GRN inference process.
• Benchmarked GRANet against state-of-the-art methods using multiple scRNA-seq datasets.

Main Results:

• GRANet consistently outperformed existing methods in GRN inference tasks across various datasets.
• The framework demonstrated high prediction accuracy in identifying regulatory interactions.
• A case study on EGR1, CBFB, and ELF1 successfully identified known and novel regulatory relationships.

Conclusions:

• GRANet offers a significant advancement in the accuracy and comprehensiveness of GRN inference.
• The proposed method has the potential to accelerate research into gene regulation and the underlying mechanisms of diseases.
• GRANet provides a powerful tool for analyzing complex gene interactions from scRNA-seq data.