Dinh 2011/NOTES/2011-8-29

From ZhangLabWiki
Jump to navigation Jump to search

48 African Methylomes (UPenn)[edit]

  • PCA plots: on matrix with ~50K Cpgs (no missing values) and on matrix with ~1.2K SNPs (also no missing values allowed)
File:UPenn African PCA.pdf
  • Note: qc factor is very strongly correlated with genetic factors within these samples
  • Sarah mentioned that PC2 might be related to metabolism
  • Histogram of the Cpg rotations on PC2:
File:UPenn African PCA rotations.pdf
  • Used >=0.012 cut off to obtain the dominating Cpgs on PC2 to get 805 Cpg sites (791 are on chromosome X, could be related to X chromosome inactivation or imprinting?)
  • Uploaded 805 sites and the ~50K sites as background to GREAT. Enrichment at GO Biological Process terms:
File:UPenn Great PC2 GO Biological Process.txt
  • Dr.Zhang: we should expand to include sites that have missing values
  • New methylation matrix:
 -QC cut off of 0.95 for every sample (only 47 samples reached 0.95 QC value)
 -Starts minimum depth cut off of 5
 -Requires at least 40 values at each Cpg
 -Requires minimum standard deviation at each site of 0.1
 -Total number of sites = 93,924
  • K-means or H-clustering can handle missing values but can't calculate the correlation on a large matrix (all paired Cpgs correlation test is computationally intensive). In addition, the best way to handle missing values is by imputation (which takes a long time), if the missing values are ignored then the correlation might be inaccurate.
  • So I decided to perform linear regression of PC2 scaled values versus all Cpgs
-Each sample (n=47) has a value on PC2
-Each Cpg has at least 40 methylation values.
-Imputation is 1000 times (not a lot but took over 20 hours to complete)
-Used 0.001 p(empirical) cut off => 2399 significant sites (FDR = 0.001*93924/2399 = 0.04)
  • Uploaded 2399 sites and the ~94K sites as background to GREAT.
File:UPenn Great PC2 regressedSites.txt