Pitfalls and opportunities for applying latent variables in single-cell eQTL analyses

Area of Science:

• Genomics
• Computational Biology
• Single-cell analysis

Background:

• Latent variables correct confounders and boost statistical power in gene expression studies.
• Probabilistic Estimation of Expression Residuals (PEER) and Principal Component Analysis (PCA) are common for bulk RNA-seq, but their single-cell application is less understood.
• Single-cell RNA-seq data presents unique challenges like sparsity and skewness, potentially impacting latent variable performance.

Purpose of the Study:

• To evaluate the effectiveness of PEER and PCA for single-cell expression quantitative trait loci (eQTL) detection.
• To identify optimal data processing steps for latent variable generation in single-cell eQTL analysis.
• To assess the impact of cell type heterogeneity on latent variable performance and eGene discovery.

Main Methods:

• Analysis of pseudo-bulk matrices derived from single-cell RNA-seq data.
• Application and validation of PEER and PCA for latent variable identification.
• Integration of latent variables into eQTL association models.
• Sensitivity analyses across different cell types and gene subsets.

Main Results:

• PEER and PCA require specific quality control and transformation steps on pseudo-bulk data to yield valid latent variables, avoiding highly correlated factors.
• Incorporating validated latent variables increased expression-associated gene (eGene) detection by 1.7% to 13.3%.
• Using highly variable genes for latent variable generation achieved comparable eGene discovery to using all genes, with a ~6.2-fold speedup in computation.

Conclusions:

• Properly processed latent variables significantly enhance eQTL discovery in single-cell data.
• Cell type-specific adjustments are crucial for optimizing latent variable-based eQTL analysis.
• Leveraging highly variable genes offers an efficient strategy for latent variable computation without compromising discovery power.