A meta-analysis on age-associated changes in blood DNA methylation: results from an original analysis pipeline for Infinium 450k data

Abstract

Aging is characterized by a profound remodeling of the epigenetic architecture in terms of DNA methylation patterns. To date the most effective tool to study genome wide DNA methylation changes is Infinium HumanMethylation450 BeadChip (Infinium 450k). Despite the wealth of tools for Infinium 450k analysis, the identification of the most biologically relevant DNA methylation changes is still challenging. Here we propose an analytical pipeline to select differentially methylated regions (DMRs), tailored on microarray architecture, which is highly effective in highlighting biologically relevant results. The pipeline groups microarray probes on the basis of their localization respect to CpG islands and genic sequences and, depending on probes density, identifies DMRs through a single-probe or a region-centric approach that considers the concomitant variation of multiple adjacent CpG probes. We successfully applied this analytical pipeline on 3 independent Infinium 450k datasets that investigated age-associated changes in blood DNA methylation. We provide a consensus list of genes that systematically vary in DNA methylation levels from 0 to 100 years and that have a potentially relevant role in the aging process.

Figure 1. Infinium 450k probes classification and BOPs definition

(A) The 485577 probes included in the Illumina HumanMethylation450 BeadChip were divided in 4 classes on the basis of their genomic localization. (B) Graphic representation of how probes were grouped in BOPs. Probes mapping in the island and in the surrounding regions of the same CpG island were grouped in 5 functional units: probes in the N-Shelf of the island, probes in the N-Shore of the island, probes in the island, probes in the S-Shore of the island, probes in the S-Shelf of the island. Probes mapping in gene bodies were grouped on the basis of the gene in which they are located.

Figure 2. Characteristics of the BOPs belonging to different probe classes

(A) Numbers of BOPs and CpG probes in Class A, Class B and Class C. For Class A and Class B, subdivision in CpG islands, N-Shores, S-Shores, N-Shelves and S-Shelves is reported. In the lower part of the tables, descriptive statistics for the distribution of number of probes/BOP in the 3 Classes are reported. (B) Descriptive statistics for the distribution of mean bp distance between probes /BOP in the 3 Classes are reported.(C) Density distributions (upper panel) and frequency histograms (lower panels) of the mean bp distance between the probes/BOP.

Figure 3. Proposed analytical pipeline for Infinium 450k data

(A) Workflow for the use of single-probe or region-centric approaches on Infinium 450k data. (B) Graphical representation of the sliding window MANOVA used to normalize for BOPs lengths. CpG probes are represented as circles. The CpG probes considered in each round of MANOVA are highlighted in yellow. (C) Example of methylation values of CpG probes within a BOP. The BOP includes 22 CpG probes, 5 of which define a “bubble” of differential methylation between Group A and Group B. The p-value derived from MANOVA on this BOP is 2.70*10e-11. We hypothesized to have a shorter BOP including only the 5 CpG probes differentially methylated between Group A and Group B, plus a probe on both the sites whose methylation level is comparable between the two samples. In this case, although the extent of the bubble of differentially methylation is the same of the longer BOP, the p-value derived from MANOVA is lower, equal to 8.64*10e-14. This simple example shows that if we do not normalize for the length of the BOP, short BOPs tend to rank at higher positions than long BOPs.

Figure 4. Number of significant CpG probes per significant BOP

For each dataset, the boxplot reports the –log₁₀(q-value) of each significant Class A BOP (MANOVA analysis) against the number of significant CpG probes (q-value < 0.05, ANOVA analysis) included in each BOP.

Figure 5. The region-centric approach increases the common findings between the 3 datasets

(A) Intersection between the results provided by Hannum et al. and Heyn et al. (left panel) and between the results of the region-centric approach on the two datasets. (B) Intersection between a progressively increasing number of top ranking features (BOPs for the region-centric analysis, CpG probes for the single-probe analysis) in the three datasets.

Figure 6. Examples of DNA methylation profiles of selected age-associated BOPs

6 of the 42 selected BOPs are reported as an example. Mean methylation values in 10 age classes are reported for each CpG probe within the selected BOPs. For each BOP, beta-values from D1, D2 and D3 were joined together.