Deep learning-based cell-specific gene regulatory networks inferred from single-cell multiome data.

Area of Science:

• Genomics and Bioinformatics
• Systems Biology
• Computational Biology

Background:

• Gene regulatory networks (GRNs) are crucial for understanding genome regulation and genetic information transfer.
• Single-cell multi-omics data offer high-resolution insights into GRNs but are limited by data loss in single-cell sequencing.
• Accurate GRN inference is essential for elucidating cellular mechanisms in health and disease.

Purpose of the Study:

• To develop a deep learning framework, scMultiomeGRN, for inferring transcription factor (TF) regulatory networks.
• To integrate single-cell RNA sequencing (scRNA-seq) and single-cell ATAC sequencing (scATAC-seq) data for robust GRN reconstruction.
• To improve the accuracy and resolution of GRN inference from challenging single-cell multi-omics data.

Main Methods:

• Developed scMultiomeGRN, a deep learning framework utilizing unique integration of scRNA-seq and scATAC-seq data.
• Implemented modality-specific neighbor aggregators and cross-modal attention modules to learn TF representations.
• Conceptualized TF network graph structures for enhanced network elucidation.

Main Results:

• scMultiomeGRN demonstrated superior performance compared to state-of-the-art models on benchmark datasets.
• The framework successfully identified disease-relevant regulatory networks, including SPI1 and RUNX1 in microglia for Alzheimer's disease.
• Achieved accurate cell type-specific GRN inference from single-cell multi-omics data.

Conclusions:

• scMultiomeGRN provides a powerful deep learning approach for reconstructing gene regulatory networks.
• The framework effectively overcomes limitations of single-cell data loss for improved GRN inference.
• Enables discovery of cell type-specific regulatory mechanisms relevant to human diseases.