A hybrid deep learning framework for gene regulatory network inference from single-cell transcriptomic data.

Area of Science:

• Genomics
• Systems Biology
• Bioinformatics

Background:

• Gene regulatory networks (GRNs) are crucial for understanding cellular phenotypes and biological processes.
• Single-cell transcriptomic data offers detailed insights into cell-to-cell variations but presents challenges like sparsity and noise.
• Accurate GRN inference is vital for studying cell functions and interactions.

Purpose of the Study:

• To develop a robust deep learning framework for inferring GRNs from challenging single-cell transcriptomic data.
• To enhance the accuracy and reliability of gene regulatory relationship identification.
• To discover novel gene interactions and regulatory pathways.

Main Methods:

• A hybrid deep learning framework, DGRNS, was developed, integrating recurrent neural networks (RNN) and convolutional neural networks (CNN).
• The model utilizes sliding windows to extract features while preserving regulatory directionality.
• Gated recurrent units capture temporal information, and CNNs learn spatial relationships.

Main Results:

• DGRNS demonstrated superior performance compared to state-of-the-art methods on mouse hematopoietic stem cell data.
• Achieved a 16% higher area under the receiver operating characteristic curve (AUROC) than unsupervised methods and a 10% higher area under the precision-recall curve (AUPRC) than supervised methods.
• Validated robustness and generalization on human datasets, identifying novel, high-confidence gene interactions.

Conclusions:

• DGRNS provides a powerful and accurate approach for GRN inference from single-cell transcriptomic data.
• The framework effectively overcomes data sparsity and noise challenges.
• Identified novel regulatory relationships with high confidence, warranting further experimental investigation.