GRouNdGAN: GRN-guided simulation of single-cell RNA-seq data using causal generative adversarial networks

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-dimensional data crucial for understanding cellular heterogeneity.
• Accurate simulation of scRNA-seq data is essential for developing and benchmarking computational methods, especially for gene regulatory network (GRN) inference.
• Existing simulation methods often struggle to capture the complex, causal relationships within biological systems.

Purpose of the Study:

• To introduce GRouNdGAN, a novel generative model for simulating scRNA-seq data.
• To enable in silico perturbation experiments, such as transcription factor (TF) knockouts.
• To provide a robust platform for benchmarking GRN inference algorithms using biologically realistic simulated data.

Main Methods:

• GRouNdGAN employs a gene regulatory network (GRN)-guided, reference-based causal implicit generative model.
• The model architecture incorporates a user-defined GRN to guide the simulation process.
• Training was performed on six experimental scRNA-seq reference datasets.

Main Results:

• GRouNdGAN successfully simulates both steady-state and transient-state scRNA-seq data, reflecting causal gene expression under TF control.
• The model captures non-linear TF-gene dependencies and preserves key biological features like cell trajectories and pseudo-time ordering.
• Simulated data accurately reflects technical and biological noise without explicit parameterization.
• In silico TF knockout experiments were successfully performed, demonstrating the model's predictive capabilities.

Conclusions:

• GRouNdGAN provides a powerful tool for generating high-fidelity, biologically relevant scRNA-seq data.
• The model effectively bridges the gap between simulated and real biological data for benchmarking GRN inference.
• GRouNdGAN offers gold-standard ground truth data for advancing the field of GRN inference and computational biology.