An evaluation of methods correcting for cell-type heterogeneity in DNA methylation studies

Abstract

Background: Many different methods exist to adjust for variability in cell-type mixture proportions when analyzing DNA methylation studies. Here we present the result of an extensive simulation study, built on cell-separated DNA methylation profiles from Illumina Infinium 450K methylation data, to compare the performance of eight methods including the most commonly used approaches.

Results: We designed a rich multi-layered simulation containing a set of probes with true associations with either binary or continuous phenotypes, confounding by cell type, variability in means and standard deviations for population parameters, additional variability at the level of an individual cell-type-specific sample, and variability in the mixture proportions across samples. Performance varied quite substantially across methods and simulations. In particular, the number of false positives was sometimes unrealistically high, indicating limited ability to discriminate the true signals from those appearing significant through confounding. Methods that filtered probes had consequently poor power. QQ plots of p values across all tested probes showed that adjustments did not always improve the distribution. The same methods were used to examine associations between smoking and methylation data from a case-control study of colorectal cancer, and we also explored the effect of cell-type adjustments on associations between rheumatoid arthritis cases and controls.

Conclusions: We recommend surrogate variable analysis for cell-type mixture adjustment since performance was stable under all our simulated scenarios.

Keywords: Cell-type mixture; DNA methylation; Deconvolution; Matrix decomposition.

Fig. 1

Clustered heat map showing patterns of methylation in 46 SARD samples (columns) and 200 CpG sites (rows). The sites were selected to highlight the methylation differences between cell types. Consequently, the samples cluster by cell type: monocytes, B cells, and then T cells

Fig. 2

QQ plots showing distributions of p values in simulation scenario 1 where the true effects in the different cell types have very distinct distributions. Results are shown with no adjustment for cell-type mixture as well as with eight other methods; these are split across two panels, (a) and (b), to clarify the display

Fig. 3

Comparison of performance metrics over ten replications in Scenario 7, few associated DMS. a The number of non-differentially methylated sites with a raw p value less than 10⁻⁴. b Power at significance level 10⁻⁴. c Kolmogorov–Smirnov statistic. d Genomic inflation factor

Fig. 4

Comparison of performance metrics over ten replications in the many associated DMSs scenario. a Number of non-differentially methylated sites with a raw p value less than 10⁻⁴. b Power at significance level 10⁻⁴. c Kolmogorov–Smirnov statistic. d Genomic inflation factor

Fig. 5

QQ plots of −log₁₀ p values from the ARCTIC study with different adjustment methods. These are split across two panels, (a) and (b), to clarify the display

Fig. 6

Agreement of the proportion of top d CpGs declared significant in the different cell-type adjustment methods with the top d CpGs in the originally reported list of significant CpGs from the rheumatoid arthritis study

Fig. 7

Computational time comparison. a Sample size. The latent dimension was estimated by the algorithms as needed. b Latent dimension. The sample size is fixed at 50