scSGL

scSGL infers signed gene regulatory networks (GRNs) from single-cell RNA-seq (scRNAseq) data by learning graph topology that captures activating and inhibitory interactions while accounting for dropouts and cell-cycle heterogeneity.

Key Features:

• Graph Signal Processing (GSP): Leverages graph signal processing to infer GRN topology from signals defined on graphs.
• Signed Graph Learning: Learns signed graphs by modeling smoothness over activating edges and non-smoothness over inhibitory edges in gene expression data.
• Kernel Extension: Incorporates kernel methods to capture nonlinear co-expression relationships and to model smoothness/non-smoothness in a higher-dimensional feature space.
• Optimization Framework: Formulates the problem as a non-convex optimization solved using an Alternating Direction Method of Multipliers (ADMM) framework.
• Dropout and Heterogeneity Handling: Explicitly accounts for high proportions of zero values (dropouts) and cell-cycle heterogeneity common in large scRNAseq datasets.

Scientific Applications:

• GRN Reconstruction: Reconstruction of gene regulatory networks from single-cell RNA-seq data to resolve activating and inhibitory interactions.
• Regulatory Mechanism Analysis: Investigation of cellular functions and regulatory mechanisms at single-cell resolution.
• Method Validation and Benchmarking: Validation and benchmarking on simulated and real single-cell datasets demonstrating superior performance relative to existing algorithms.

Methodology:

Applies graph signal processing and signed graph learning with assumptions of smoothness on activating edges and non-smoothness on inhibitory edges, extends the model via kernel methods to handle nonlinear co-expression and dropouts, and solves the resulting non-convex optimization problem using an ADMM algorithm.