Significance analysis and statistical dissection of variably methylated regions

Abstract

It has recently been proposed that variation in DNA methylation at specific genomic locations may play an important role in the development of complex diseases such as cancer. Here, we develop 1- and 2-group multiple testing procedures for identifying and quantifying regions of DNA methylation variability. Our method is the first genome-wide statistical significance calculation for increased or differential variability, as opposed to the traditional approach of testing for mean changes. We apply these procedures to genome-wide methylation data obtained from biological and technical replicates and provide the first statistical proof that variably methylated regions exist and are due to interindividual variation. We also show that differentially variable regions in colon tumor and normal tissue show enrichment of genes regulating gene expression, cell morphogenesis, and development, supporting a biological role for DNA methylation variability in cancer.

Fig. 1.

Examples of 1- and 2-group DNA methlyation patterns at VMRs and randomly chosen regions. The top panel of each plot is the DNA methylation for each array, and the bottom panel is the smoothed adjusted MAD or absolute difference in MADs statistic with shading representing statistically significant regions (a) One-group VMR on chromosome 12. (b) Two-group dVMR on chromosome 11 (black lines = cancer and gray lines = normal). (c) One-group randomly chosen region representing background variability in the genome. (d) Two-group randomly chosen region with similar variability across both groups (black lines = cancer and gray lines = normal).

Fig. 2.

VMRs in replicates. (a) Sample of biological replicates in peripheral blood: identical colors come from the same individual, solid lines are from the earlier study visit, and dashed lines come from the later study visit. The solid and dashed lines of the same color are closer together than solid lines of different colors in VMRs, indicating the VMRs are due to interpersonal rather than intrapersonal variation (b) Technical replicates from spleen tissue: identical colors come from the same spleen. The solid and dashed lines of the same color are closer than other pairwise line pairs (solid or dashed), indicated the VMRs are due to biological, rather than technical variation.

Fig. 3.

The distribution of euclidean distances between replicates and nonreplicates in VMRs and random regions. Each boxplot is the average distances between replicates minus the average distance between nonreplicates in all VMRs and an equal number of randomly selection genomic regions. (a) Biological peripheral blood replicates: the average distances are much smaller between replicates than nonreplicates within VMRs than random regions. (b) Technical spleen tissue replicates: the average distances are also smaller between replicates than nonreplicates in VMRs.