Recovering time-varying networks from single-cell data.

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Gene regulation is crucial for development and disease.
• Traditional methods struggle with large, time-series single-cell data.
• Novel approaches are needed for accurate temporal gene network reconstruction.

Purpose of the Study:

• To develop a deep learning model for inferring dynamic gene regulatory networks.
• To address the challenges posed by time-series single-cell gene expression data.
• To enable the identification of gene interactions in dynamic biological processes.

Main Methods:

• Developed Marlene, a deep neural network utilizing a self-attention mechanism.
• Incorporated recurrent units for time-evolving network weights.
• Employed meta-learning to enhance accuracy, especially for rare cell types.

Main Results:

• Marlene successfully infers dynamic gene networks from time-series single-cell data.
• The model accurately reconstructs temporal networks, even for rare cell populations.
• Identified gene interactions relevant to COVID-19 immune response, fibrosis, and aging.

Conclusions:

• Marlene offers a powerful new approach for reconstructing dynamic gene regulatory networks.
• The tool facilitates the discovery of gene interactions driving specific biological responses.
• This work paves the way for novel therapeutic strategies targeting complex diseases.