Table 1.
Approaches for identifying differentially methylated loci in bisulfite sequencing data sets.
Fig 1.
MACAU appropriately controls for genetic covariance in simulated and real WGBS data and eliminates false positive identification of differentially methylated sites.
(A, B) The distribution of p-values for 4000 simulated true negative sites (n = 24 accessions; effect of secondary dormancy on DNA methylation levels = 0). For each simulation, h2 was set to 0.3 (A) or 0.6 (B). Simulated data were analyzed with a beta-binomial model or MACAU, and compared against the expected uniform distribution. (C) QQ-plots comparing the p-value distributions for (i) a model testing for effects of secondary dormancy on DNA methylation levels in real WGBS data, with quantiles plotted on the y-axis; and (ii) the same model when the secondary dormancy values were permuted across individuals, with quantiles plotted on the x-axis. The genomic control factor, λ, is shown for each set of results.
Fig 2.
MACAU controls for genetic covariance in data sets that span a range of sample sizes and levels of correlation between population structure and a predictor variable of interest.
Genomic control factor when simulated datasets (n = 5000 sites per dataset; h2 = 0.6) were analyzed with either (A) a beta-binomial model or (B) a BMM implemented in MACAU. The correlation between the simulated predictor variable and the first principal component of genome-wide genotype data is plotted on the x-axis. Genotype data are for Arabidopsis accessions, as reported in [87].
Fig 3.
MACAU exhibits increased power to detect differential methylation when DNA methylation levels are heritable.
Receiver operating characteristic (ROC) curves and true positive rates at a 10% false discovery rate threshold for simulated age effects on DNA methylation levels at (A-C) simulated sites with moderately heritable DNA methylation levels (h2 = 0.3) and (D-F) simulated sites with highly heritable DNA methylation levels (h2 = 0.6). Panels B and E are enlarged versions of panels A and D, respectively. They focus on false positive rates below 0.1, because the performance of alternative methods at low false positive rates tends to be most important to researchers in practice; that is, it is unlikely to matter if method performance is identical when accepting a 50% false positive rate, which would yield very poor inferential power. Each simulated dataset contained n = 80 individuals and 5000 simulated CpG sites, with 500 true positives and 4500 true negatives. Here, we show results where the simulated percent variance explained by age = 10%. A binomial model could not detect true positives at a false positive rate below 0.10 (when h2 = 0.3) or below 0.9 (when h2 = 0.6); the binomial is therefore removed from panel B, and only shown for large false positive rates in panel E.
Fig 4.
Age-associated CpG sites identified by MACAU in the baboon RRBS data.
(A) The number of age-associated CpG sites detected at a given empirical FDR. The binomial model cannot detect age-associated sites at a false discovery rate below 0.20 and is consequently removed from the panel. (B) For age-associated sites detected by MACAU (at a 10% FDR), the proportion of sites that gain or lose methylation with age is shown by genomic region. Positive = DNA methylation levels increase with age; Negative = DNA methylation levels decrease with age. (C) Age-associated CpG sites detected using MACAU (10% FDR) are more likely to fall near genes that are expressed in whole blood, compared to the background set of CpG sites near genes (**p < 10−10). Further, age-associated CpG sites are more likely to occur near genes that are differentially expressed (DE) with age, compared to CpG sites near genes that are not DE with age (*p = 0.032).