Inference after latent variable estimation for single-cell RNA sequencing data.

Area of Science:

• Computational biology
• Genomics
• Statistical genetics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables cell state characterization via latent variable estimation (e.g., cell type, pseudotime).
• Testing gene association with latent variables using the same data compromises statistical validity, failing Type 1 error control.
• Traditional methods like sample splitting are unsuitable for scRNA-seq data analysis.

Purpose of the Study:

• Introduce count splitting, a novel statistical framework for valid inference in scRNA-seq.
• Address the challenge of statistical error control when latent variables and gene associations are derived from the same dataset.
• Provide a flexible method applicable to diverse latent variable estimation techniques and inference approaches.

Main Methods:

• Developed the count splitting framework, assuming a Poisson distribution for gene counts.
• Count splitting enables valid statistical inference by separating latent variable estimation from gene association testing.
• The method is designed to be compatible with various latent variable estimation algorithms.

Main Results:

• Demonstrated robust Type 1 error control through simulation studies.
• Showcased the statistical power of count splitting in detecting gene associations.
• Successfully applied count splitting to analyze a dataset of differentiating pluripotent stem cells.

Conclusions:

• Count splitting offers a statistically sound and flexible approach for analyzing scRNA-seq data.
• The method effectively addresses the limitations of standard statistical procedures in this context.
• Count splitting facilitates reliable gene association studies in complex biological systems, such as cell differentiation.