Search

Using latent variables in gene expression data can help correct unobserved confounders and increase statistical power for expression quantitative trait Loci (eQTL) detection. The probabilistic estimation of expression residuals (PEER) and principal component analysis (PCA) are widely used methods that can remove unwanted variation and improve eQTL discovery power in bulk RNA-seq analysis. However, their performance has not been evaluated extensively in single-cell eQTL analysis, especially for different cell types. Potential challenges arise due to the structure of single-cell RNA-seq data, including sparsity, skewness, and mean-variance relationship. Here, we show by a series of analyses that PEER and PCA require additional quality control and data transformation steps on the pseudo-bulk matrix to obtain valid latent variables; otherwise, it can result in highly correlated factors (Pearson's correlation r = 0.63 ~ 0.99). Incorporating valid PFs/PCs in the eQTL association model would identify 1.7 ~ 13.3% more eGenes. Sensitivity analysis showed that the pattern of change between the number of eGenes detected and fitted PFs/PCs varied significantly in different cell types. In addition, using highly variable genes to generate latent variables could achieve similar eGenes discovery power as using all genes but save considerable computational resources (~ 6.2-fold faster).

Selective serotonin reuptake inhibitors (SSRIs) are standard of care for major depressive disorder (MDD) pharmacotherapy, but only approximately half of these patients remit on SSRI therapy. Our previous genome-wide association study identified a single-nucleotide polymorphism (SNP) signal across the glutamate-rich 3 (ERICH3) gene that was nearly genome-wide significantly associated with plasma serotonin (5-HT) concentrations, which were themselves associated with SSRI response for MDD patients enrolled in the Mayo Clinic PGRN-AMPS SSRI trial. In this study, we performed a meta-analysis which demonstrated that those SNPs were significantly associated with SSRI treatment outcomes in four independent MDD trials. However, the function of ERICH3 and molecular mechanism(s) by which it might be associated with plasma 5-HT concentrations and SSRI clinical response remained unclear. Therefore, we characterized the human ERICH3 gene functionally and identified ERICH3 mRNA transcripts and protein isoforms that are highly expressed in central nervous system cells. Coimmunoprecipitation identified a series of ERICH3 interacting proteins including clathrin heavy chain which are known to play a role in vesicular function. Immunofluorescence showed ERICH3 colocalization with 5-HT in vesicle-like structures, and ERICH3 knock-out dramatically decreased 5-HT staining in SK-N-SH cells as well as 5-HT concentrations in the culture media and cell lysates without changing the expression of 5-HT synthesizing or metabolizing enzymes. Finally, immunofluorescence also showed ERICH3 colocalization with dopamine in human iPSC-derived neurons. These results suggest that ERICH3 may play a significant role in vesicular function in serotonergic and other neuronal cell types, which might help explain its association with antidepressant treatment response.

Millions of patients suffer from major depressive disorder (MDD), but many do not respond to selective serotonin reuptake inhibitor (SSRI) therapy. We used a pharmacometabolomics-informed pharmacogenomics research strategy to identify genes associated with metabolites that were related to SSRI response. Specifically, 306 MDD patients were treated with citalopram or escitalopram and blood was drawn at baseline, 4 and 8 weeks for blood drug levels, genome-wide single nucleotide polymorphism (SNP) genotyping and metabolomic analyses. SSRI treatment decreased plasma serotonin concentrations (P