SGMHA: semantic graph reconstruction with multi-head attention for gene regulatory network inference

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) data is crucial for understanding cellular heterogeneity.
• Data sparsity and dropout events in scRNA-seq pose significant challenges for inferring gene regulatory networks (GRNs).
• Existing supervised methods struggle with noise, limiting the reliability of GRN inference.

Purpose of the Study:

• To develop a robust and accurate method for inferring GRNs from sparse scRNA-seq data.
• To address the limitations of current methods in handling data noise and sparsity.
• To facilitate biomarker discovery in complex diseases like cancer metastasis.

Main Methods:

• Proposed SGMHA, a two-stage framework decoupling representation learning and link prediction.
• Employed a self-supervised graph masked autoencoder (GraphMAE) for learning robust gene representations.
• Utilized a multi-head attention (MHA)-based fine-tuning module for accurate regulatory link inference.

Main Results:

• SGMHA consistently outperformed eight state-of-the-art methods on seven scRNA-seq datasets (AUROC and AUPRC).
• Identified context-specific GRNs in breast cancer metastasis, revealing 26 high-confidence candidate drivers.
• Validated six candidate drivers as prognostic biomarkers with detailed mechanistic roles.

Conclusions:

• SGMHA offers an accurate, scalable, and biologically interpretable tool for GRN inference.
• The method effectively mitigates sparsity-induced distortions in scRNA-seq data.
• SGMHA shows strong promise for advancing biomarker discovery in complex diseases.