Evaluation of commonly used analysis strategies for epigenome- and transcriptome-wide association studies through replication of large-scale population studies

Abstract

BackgroundA large number of analysis strategies are available for DNA methylation (DNAm) array and RNA-seq datasets, but it is unclear which strategies are best to use. We compare commonly used strategies and report how they influence results in large cohort studies.ResultsWe tested the associations of DNAm and RNA expression with age, BMI, and smoking in four different cohorts (n = ~ 2900). By comparing strategies against the base model on the number and percentage of replicated CpGs for DNAm analyses or genes for RNA-seq analyses in a leave-one-out cohort replication approach, we find the choice of the normalization method and statistical test does not strongly influence the results for DNAm array data. However, adjusting for cell counts or hidden confounders substantially decreases the number of replicated CpGs for age and increases the number of replicated CpGs for BMI and smoking. For RNA-seq data, the choice of the normalization method, gene expression inclusion threshold, and statistical test does not strongly influence the results. Including five principal components or excluding correction of technical covariates or cell counts decreases the number of replicated genes.ConclusionsResults were not influenced by the normalization method or statistical test. However, the correction method for cell counts, technical covariates, principal components, and/or hidden confounders does influence the results.

Genome-Wide DNA Methylation in Peripheral Blood and Long-Term Exposure to Source-Specific Transportation Noise and Air Pollution: The SAPALDIA Study

“…The relevance of confounding by leukocyte composition on the association between exposures and peripheral blood DNA methylation has been of interest (Heiss and Brenner 2017;van Rooij et al 2019). In post hoc analyses, we demonstrated the general weakening of top CpGs (with some signals not attaining nominal significance) when leukocyte composition was not considered.…”

VTRNA2-1: Genetic Variation, Heritable Methylation and Disease Association

“…We used methylation M-values as their distribution is usually closer to Gaussian than methylation beta-values [ 54 ]. As a sensitivity analysis, we performed a more direct normalization of beta-values using rank-based inverse normal transformation (RINT) which was applied previously in the context of DNA methylation analyses and provides a Gaussian methylation distribution, which is not the case for the M-values [ 55 , 56 ].…”

Benchmarking Association Analyses of Continuous Exposures with RNA-seq in Observational Studies

“…Recently, van Rooij et al (30) also performed a benchmarking study comparing analysis approaches for transcriptome-wide analysis of RNA-seq in population-based studies, including when using continuous phenotypes in association testing. While we used similar statistical methods to theirs, we took a different analytical approach.…”

“…Here, we did not study their effect because it was beyond the scope of our investigation. van Rooij et al (30) in their benchmarking study focusing on replication across cohorts, compared a few approaches for adjusting for technical covariates, including estimating and adjusting for latent confounders (38). They concluded that inclusion of more technical adjusting covariates, including hidden confounders, increases the rate of replication between studies.…”