scSGL: kernelized signed graph learning for single-cell gene regulatory network inference.

Area of Science:

• Computational Systems Biology
• Bioinformatics
• Genomics

Background:

• Inferring gene regulatory network (GRN) topology from single-cell RNA sequencing (scRNA-seq) data is challenging due to cell-cycle heterogeneity and data sparsity (dropouts).
• Existing graph learning (GL) methods struggle with signed graphs, which are crucial for representing activating and inhibitory gene interactions, and cannot effectively handle the high proportion of zero values in scRNA-seq datasets.

Purpose of the Study:

• To propose a novel signed graph learning (GL) approach, scSGL, for accurate GRN inference from scRNA-seq data.
• To address the limitations of existing GL methods in handling signed relationships and zero-inflated data characteristic of GRNs.

Main Methods:

• Developed scSGL, a signed GL method assuming smoothness over activating edges and non-smoothness over inhibitory edges in gene expression.
• Extended scSGL with kernels to capture non-linear gene co-expression and handle excessive zero values.
• Formulated the approach as a non-convex optimization problem solved via an efficient Alternating Direction Method of Multipliers (ADMM) framework.

Main Results:

• Kernelized scSGL demonstrated superior performance in GRN recovery compared to state-of-the-art methods on simulated datasets.
• The effectiveness of scSGL was further validated on real-world scRNA-seq datasets from human and mouse embryos.

Conclusions:

• scSGL provides a robust and effective framework for inferring signed gene regulatory networks from challenging single-cell data.
• The method's ability to handle non-linearity and data sparsity offers significant advantages for understanding gene regulation.