Correcting for cell-type effects in DNA methylation studies: reference-based method outperforms latent variable approaches in empirical studies

Abstract

Based on an extensive simulation study, McGregor and colleagues recently recommended the use of surrogate variable analysis (SVA) to control for the confounding effects of cell-type heterogeneity in DNA methylation association studies in scenarios where no cell-type proportions are available. As their recommendation was mainly based on simulated data, we sought to replicate findings in two large-scale empirical studies. In our empirical data, SVA did not fully correct for cell-type effects, its performance was somewhat unstable, and it carried a risk of missing true signals caused by removing variation that might be linked to actual disease processes. By contrast, a reference-based correction method performed well and did not show these limitations. A disadvantage of this approach is that if reference methylomes are not (publicly) available, they will need to be generated once for a small set of samples. However, given the notable risk we observed for cell-type confounding, we argue that, to avoid introducing false-positive findings into the literature, it could be well worth making this investment.Please see related Correspondence article: https://genomebiology.biomedcentral.com/articles/10/1186/s13059-017-1149-7 and related Research article: https://genomebiology.biomedcentral.com/articles/10.1186/s13059-016-0935-y

Correcting for cell-type heterogeneity in epigenome-wide association studies: premature analyses and conclusions

“…In scenarios where cell-type composition drives most of the variation and where the underlying main cell subtypes are known, a reference-based method is, in our view, the current state-of-theart since this approach should yield fairly reliable cell-type composition estimates, as well as a reliable set of DMCs. This recommendation is further supported by a recent letter by Hattab et al [30]. We further note that a reference-based method can be easily combined with batch-correction methods such as COMBAT if evidence for batch effects is present.…”

Maximizing ecological and evolutionary insight in bisulfite sequencing data sets

“…Second, if no measures of cell type heterogeneity are available for the samples of interest, another option is to use epigenomic profiles from sorted cells 76 , 77 (‘reference epigenomes’) to predict the composition of mixed samples (a process known as ‘deconvolution’ 71 , 78 ). Even if obtained from different individuals or populations (and likely even if obtained from a closely related species), this approach can provide reasonable control for cell type heterogeneity 79 . Reference epigenome-based deconvolution is an active area of research, and several software packages exist to execute it 79 , 80 .…”

“…Even if obtained from different individuals or populations (and likely even if obtained from a closely related species), this approach can provide reasonable control for cell type heterogeneity 79 . Reference epigenome-based deconvolution is an active area of research, and several software packages exist to execute it 79 , 80 . Data from reference epigenomes can also be used to test if sites that are differentially methylated with respect to the predictor of interest are also differentially methylated by cell type, which would suggest the two are confounded 8 , 37 .…”

“…In comparison, FaST-LMM-EWASher detected a single differentially methylated site in the same data set. Recent comparisons of reference-based and reference-free methods suggest that reference-based approaches are consistently better powered 79 , 80 , and reference-free methods should only be used when sorted cell profiles are not available (SVA and RefFreeEWAS are recommended in these cases 79 , 80 ). Such results suggest that researchers should consider investing in the generation of a small set of reference epigenomes, if possible for their system.…”

Convergence of evidence from a methylome-wide CpG-SNP association study and GWAS of major depressive disorder

“…Second, we regressed out the demographic variables age and sex. Third, to avoid confounding due to cell-type heterogeneity, we regressed out blood cell type proportions as estimated by the methylation data 29 using MBD-Seq “reference methylomes” we generated after isolating all common cell types in blood 30 . Fourth, principle component analysis was used to capture any remaining unmeasured source of variation.…”