The data is here: /work/DAT_002__TCGA_Ovarian/01_2011/vita_work/ExprMeth

/work/DAT_002__TCGA_Ovarian/01_2011/vita_work/ExprMeth
The basic question we asked here is how much DNA methylation is correlated with gene expression. We have all heard that DNA methylation is a repressor but there are plenty of articles that suggest a dual role for DNA methylation (Jones, 1999, Yasui, 2007 and myself, 2008). Additional recent body of work has been devoted to the study of CpG island shores, "regions of comparatively low CpG density that are located near traditional CpG islands, than at the CpG islands themselves". It has been shown that "differential DNA methylation correlated with gene expression more strongly at CpG island shores than CpG islands". Illumina's 27k DNA methylation platform is designed around the CpG islands and it is a little worrisome how well it might be suited for the analysis of DNA methylation changes in cancer (+ we don't have controls!!!). However, 25% of all CpG loci on the platform are not within the CpG islands (source: a table from Bioconductor's IlluminaHumanMethylation27k.db package), so we have some hope.

To correlate DNA methylation with gene expression I did the following:

Extracted chromosome and chromosomal position of each CpG from IlluminaHumanMethylation27k.db package (hg18). TCGA doesn't provide any annotation information for the platform.
Downloaded chromosomal position for each gene on the Affymetrix platform from TCGA here.
Imported all coordinates for the genes and CpG loci to a mysql database and mapped CpGs based on the location within +/-2kb of the TSS of each gene. Total number of mappings is 20029, which results from the fact that there are on average 1.9 CpGs per Gene (need to double check this number)

Once again: all correlation analyses here is for cis regulation, not trans.

Correlations for the mb and Affy expression:

Correlation for the mbc and Affy expression:

It is nice to see a strong signal. FDR calculation (The analysis was done on June 29th. Number of permutations was 20. Is it sufficient?).

mb:

P value	FDR, %	Number of pairs
0.05	11.33	8931
0.01	3	6813
0.005	1.7	6146
0.001	0.4	4957

mbc:

P value	FDR, %	Number of pairs
0.05	11.67	8488
0.01	3.12	6315
0.005	1.7	5661
0.001	0.4	4500

I selected P value cutoff of 0.005 because of its very acceptable FDR. for this cutoff mb vs Affy max Rho value is 0.7, min is -0.79 which appears to be a very strong correlation.

Number of positively correlated pairs: 4675, negatively correlated pairs: 1471. Interesting to see a big number of positive correlations of the level of DNA methylation with gene expression. I expected it to be either 50/50 or more negative than positive correlations.

For the gene ontology analysis I focused on the correlation coefficient rho rather than the P values. Rho >= 0.4 and rho <= -0.4 gave me a total of 581 targets. Did DAVID analysis (GO BP, MF, CC - level 3, KEGG pathways, PIR keywords and INTERPRO. Parameters: min count 10, EASE threshold 0.001; background: HT Human Genome U133A). Positive correlations were represented by intracellular part/organelle, acetylation, nuclear part, nucleobase, nucleotide, nucleoside, RNA binding, phosphoprotein, gene expression. Negative correlations were mostly represented by regulation of immune response.

mbc:

FDR: 1.7%, P value: 0.005, total number of targets: 5661

max R value: 0.73, minimum: -0.78. Number of positively correlated pairs: 4464, number of negatively correlated pairs: 1197. Rho >= 0.4 - 339, Rho <= -0.4 - 141.

DAVID analysis with the same parameters as for mbc didn't return any significant categories for the negatively correlated pairs (no immune response), However GO analysis for the positively correlated pairs appeared to be very similar to the analysis with mb.

Overlaps between correlated CpGs of mb and mbc sets:

Note: it will be interesting to investigate what kind of TD binding site will be enriched in the promoter regions of the gene correlated positively with DNA methylation vs negatively.

To do: from the analysis of comethylation network I found that many CpGs that map to a nearby gene (based on IlluminaHumanMethylation27k.db) actually end up in different modules. This is not an artifact, they do in fact have different methylation status. Given 2 CpGs that map to the same nearby gene, how well do they correlate with the expression of that gene? Is it similar for both, completely different? How many cases?

Browser not supported