Evaluating the Reproducibility of Single-Cell Gene Regulatory Network Inference Algorithms

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Biological networks are essential for understanding biological systems.
• Single-cell RNA sequencing (scRNA-seq) has enabled new approaches for inferring gene regulatory networks.
• Existing benchmarks for scRNA-seq network inference often rely on simulated data or limited gene sets.

Purpose of the Study:

• To benchmark six single-cell network inference methods using real biological data.
• To evaluate method performance based on reproducibility across independent datasets.
• To compare inferred networks against known biological interactions.

Main Methods:

• Applied six network inference algorithms to independent scRNA-seq datasets from human retina, T-cells in colorectal cancer, and human hematopoiesis.
• Assessed reproducibility by comparing inferred networks between datasets for the same biological condition.
• Evaluated intersection with ground-truth biological interactions for up to 100,000 links.

Main Results:

• GENIE3 was the most reproducible algorithm for inferring gene regulatory networks from scRNA-seq data.
• GENIE3 and GRNBoost2 showed higher concordance with ground-truth interactions.
• Results were consistent across different sequencing platforms, cell annotation systems, and dataset sizes.
• GRNBoost2 and CLR exhibited improved reproducibility with stringent network thresholding (1,000-100 links).

Conclusions:

• GENIE3 and GRNBoost2 are robust and reliable methods for single-cell gene regulatory network inference.
• Reproducibility is a key metric for evaluating network inference methods on real scRNA-seq data.
• The scNET Jupyter notebook facilitates reproducible benchmarking and further research in this area.