Adjustment for gene expression PCA scores may induce reproducible false positive associations in eQTL analysis partly due to endogenous selection bias

Abstract

Motivation: Expression quantitative trait loci (eQTL) analysis has been widely applied to map cis and trans regulatory elements of gene expression. Gene expression analysis is plagued with observed and latent biases such as batch effects and cell composition.
Principal component analysis (PCA) has the potenti al to capture these latent variables associated with global gene expression levels.
Adjustment of the eQTL model for principal components (PC) of the gene expression increases the yield of statistically significant eQTLs and consequently has been widely adopted in the field. The explanation that accompanies the large increase in reproducible eQTLs is that adjustment for PCs reduce variation induced by technical and biological latent factors affecting global gene expression.

Results: We here report that such practice may induce reproducible false positive results partly due to endogenous selection bias. That is to say, adjusting for PCs may open a path between a SNP not associated with expression but correlated with a PC that in turn is associated with the expression of a gene. Our simulation shows that false positive results induced by PCA adjustment can be reproducible. Real dataset analysis suggests that regression models with a SNP not associated to expression but correlated to a PC may result in both variables mutually acting as suppressor variables thereby inflating SNP effect sizes. Similarly, adjustment for multiple PCs increases R2 of the regression model and thereby reduces the standard errors of the beta coefficients. These two effects taken together increase significance of p-values and may induce false positives. We propose a simple procedure to detect whether SNPs and PCs are acting as suppressor variables.

Conclusions: We recommend a few techniques to deal with false positive associations in eQTLs and gene expression associations that arise from adjusting models for gene expression PCs

More information

Identifiers

ORCID for Alexessander Couto Alves: orcid.org/0000-0001-8519-7356

Catalogue record

Export record

Altmetrics