Kun:LabNotes/CpgSeq/2011-9-27

From ZhangLabWiki
Jump to navigation Jump to search

Analysis of the BSPP data from UCLA RO 96 blood samples[edit]

Mid-parent offspring (MPO) analysis[edit]

  • The purpose of this analysis is to quickly identify the CpG sites that exhibit heritable patterns on DNA methylation.
  • I first combined all the BED files Noi generated with allBED2Matrix.pl. I filtered out sites that has less than 40 valid methylation values or STD<0.1.
  • I then wrote a script for MPO analysis: mpo.pl, which yielded 2921 sites that could potentially be heritable. Note that the criteria are relatively loose. I only require r>=0.2 and chi-square >1.
  • Then I wrote a second script to extract the methylation values for individual CpG sites, so that I can carefully examine how well does MPO work.extract_trios_methyl_values.pl.
  • I picked a few sites that have the highest r values, and drew the scatter plots with MS-Excel. It seems to me that the correlation is modest at best even with these sites. Also when I looked up these sites on UCSC genome browser, I found that these sites are all CpG-SNPs. This sort of validated the MPO analysis. However, to identify heritable CpG sites that are not on SNPs, we will have to filter out the SNP sites prior to MPO analysis.
  • After filtering out CpG-SNPs based on snp130, the list was reduced to 351. Two examples are shown below.
  File:Chr17 4559596 MPO.png File:Chr1 145016206 MPO.png
  • I also plotted all the 351 sites along the chromosomes. There are 46 clusters of CpG sites (>1 sites in a <2kb window).
  File:Heritable Cpg sites ideogram.png

Update on 11/23/11[edit]

  • Noi found errors in the sample information sheet received from UCLA. For some individuals, the parent IDs were incorrectly assigned. A revised sample information sheet was received from UCLA.
  • I switched to a different Perl module (Statistics::OLS) for regression analysis. This module provides t-test statistics, which makes more sense to me than the chi-square statistics provided in the Statistics::LSNoHistory. I changed the definition of heritable CpG sites to: R2 > 0.1 and abs(t-stat)>2. mpo_v2.pl
  • With the UCLA_RO_BSPP_20110921_methylMatrix_minSTD_0.1 matrix, 103,178 sites were tested, and 14,364 (14%) were found heritable.
  • With the UCLA_RO_BSPP_20110921_methylMatrix_noSNP matrix, 100,781 sites were tested and 14,008 (13.9%) were found heritable.
  • These numbers are very different from the previous analysis. They suggest that the methylation of >10% CpG sites has a genetic component. Also the majority of these sites are not CpG-SNPs.
  File:Heritable Cpg sites ideogram Nov23 2011.png

Hierarchical clustering[edit]

   A=read.table("UCLA_RO_BSPP_20110921_methylMatrix_minSTD_0.1.txt",header=TRUE,row.names=1)
   B=na.omit(A)
   B$STDEV=NULL
   B$min_RD=NULL
   B$mean_RD=NULL
   B.cor=cor(B,use="pairwise.complete.obs",method="pearson")
   B.d = as.dist(1-B.cor)
   B.h = hclust(B.d,"complete")
   plot(B.h)
   A quick observation is that most data sets cluster based on the genetic relatedness, not the disease status.
   This implies that doing subtraction within the families could be effective to identify disease-specific methylation sites/patterns.
   File:UCLA RO SZ 94 blood samples BSPP Hierarchical clustering.png

PCA analysis[edit]

  A=read.table("UCLA_RO_BSPP_20110921_methylMatrix_minSTD_0.1.txt",header=TRUE,row.names=1)
  B=na.omit(A)
  B$STDEV=NULL
  B$min_RD=NULL
  B$mean_RD=NULL
  pca1 <- prcomp(t(B), scale=TRUE)
  plot(pca1$x, type="n") 
  text(pca1$x, rownames(pca1$x), cex=0.7)
  
  There are clearly two groups separated by PC2. Why is that? Is that due to data quality, gender, age, disease or other factors?
  File:UCLA RO SZ 94 blood samples BSPP PCA plot.png