Kun:LabNotes/CpgSeq/2011-10-14

From ZhangLabWiki
Jump to navigation Jump to search

Analysis of BSPP data from Geographic Astrophy (GA) patients[edit]

  • 96 blood samples from cases and controls were provided by Kang Zhang's group.
  • Noi performed the BSPP capture and the sequencing was done in BioGem (110929_SN1001)

Summary statistics on sequencing, mapping and CpG methylation calling[edit]

Noi should fill in this part.

Regression analysis[edit]

  • The working directory is on genome-miner
  /home/kunzhang/mQTL/KangZhang_GA
  • One data set (2200001) was removed due to low capture efficiency and low sequencing depth.
  • Methylation levels from the rest of 95 samples (turned out to be 94) were combined into a matrix (up to 20% missing values was allowed):
  allBED2Matrix.pl BED 76 1 0.1 > KangZhang_GA_BSPP_95_methylationMatrix.txt &
  ./GA_methylation_regression.pl  KangZhang_GA_BSPP_95_methylationMatrix.txt > KangZhang_GA_BSPP_95_methylation_regression.txt &
  • Sites that might be significantly contributed by one of these factors were reported.
    • I used a stringent cutoff of 3.5 for the t-statistics on reporting the data in individual files (allowed for more detailed analysis in R).
    • The test statistics for all the tested sites were also reported in a separate file (KangZhang_GA_BSPP_95_methylation_regression.txt).
    • If I simply use a cutoff for |t-value| of 2.5 (equivalent to p=0.014), then I can count how many Cpg sites are significantly correlated with each factor.
     Factor         # sites
     GA             1515 
     Sex            2718 
     Age            2310
     BMI            1861
     Smoking        2002
     GA+Age         249  (somewhat expected as age is a major contributor of GA)
     GA+Sex         53
     GA+BMI         28
     GA+Smoking     34
     
     It's important to note that no correction for multiple testing was done here. 
     Even with the least stringent FDR correction, no site is significant. An independent estimation 
     of the false positive rate is to repeat the analysis on randomly permutated data, and ask how 
     many sites appear significant based on the same criteria. I did this by randomly shuffling the
     methylation levels for a CpG sites among the samples. 
     Factor         # sites
     GA             1979
     Sex            1855
     Age            1913
     BMI            2201
     Smoking        1816
     Therefore we got close to 2000 significant sites even with randomly permutated data, which suggests 
     that the majority of the sites found in this regression analysis are false positive. An exception is
     probably Age, which is correlated with more Cpg sites in non-permutated data.
     To further test the possibility that there are some true signals among the false positives, I performed
     pathway enrichment analysis with GREAT on the sites significantly correlated with GA. For the sites
     identified in the non-permutated data set, I found 45 enrichment pathways or categories. In contrast, 
     with the sites identified in the randomly permutated data set, no enrichment was found.

Methylation marker selection[edit]

  • In this analysis, I tried to select a minimal subset of CpG sites that can classify the patients from the healthy controls at the greatest power.
  • Since there are too many sites in the data set, I decided to filter out most of the non-informative sites first, then use a feature selection method to identify a small subset of most informative markers. I did this with a Perl script and mRMR.
    • Step I: perform linear regression as previously, and only keep the sites with |t value|>2.5.
    • Step II: reformat the data in the right format, and perform feature selection with mRMR.
    • The output of mRMR is a list of 50 sites ranked based on a score (mRMR output file). I then took the methylation data on these 50 sites, and did PCA analysis with R. Clearly patients and controls can be classified quite well based on the top 50 CpG sites. Even with the top 20 sites, the classification was still quite accurate.
    • Gene associated with the top 50 markers.
   File:Classification top 10 sites.png
   File:Classification top 20 sites.png
   File:Classification top 50 sites.png