Kun:LabNotes/MONOD/2014-5-13

From ZhangLabWiki
Jump to navigation Jump to search

MONOD Data Analysis (continued)[edit]

  • While analyzing the MONOD V3 data in the past few weeks, I discovered a few issues.
    • First, by down-sampling with Picard, only one of a paired-end reads is randomly selected. As the result, most haplotypes from the down-sampled reads have missing alleles. And because I always down-sampled cancer reads and mixed them with blood reads, there was a systematic difference in the valid haplotype length between the reads from two different sources. It appeared that such a difference inflated the detection sensitivity. Solution: I wrote another script to perform down-sampling on the hapInfo files. The random selection was done on the haplotype level, not on individual reads: make_synthetic_mix_hapInfo.pl
  Bam down-sampling vs. hapInfo down-sampling
  File:Two down-sampling approaches 121227 MiSeq.png
    • On the grouping of haplotypes into clusters, I picked the dominant haplotype representing a cluster based on the abundance of the haplotype. When examining the haplotype clusters carefully, I discovered that sometime a dominant haplotype might have too many missing alleles, such that it can recruit very different haplotypes into a same cluster. I modified the mixMethHapAnalysis.pl script to pick the one that has the fewest missing alleles as the dominate haplotype. I also revised the remove_non_informative_sites subroutine to make it more robust.
    • For tumor DNA detection in the mixture, I used two strategies, one based on linear regression, and the other by assigning individual haplotypes to three categories: Blood, Tumor or Unassigned. Previously I did some filtering on the capture targets and the haplotypes prior to these two different analyses. Several cutoff values were used in those filters, and the sensitivity of tumor DNA detection heavily depends on the choice of those cutoffs. To facilitate the optimization of parameters, I reduced the filters into two, one for regression and a second of assigning haplotypes. This allowed me to adjust the two parameters independently. This is the latest version of mixMethHapAnalysis.pl.
    • For all probe sets tested so far, only a fraction of the probes contributed to the detection of cancers. To identify an informative subset of probes, I wrote another script (hapGroupDiffFilter.pl) to rank probes based on the difference of haplotype clusters between blood and tumor. For each capture region, I grouped all haplotypes in blood into a number of clusters, and then the haplotypes in cancer into another set of clusters. Then I defined a metric to capture the difference between the two groups of haplotype clusters. Basically, I calculated the pair-wise hemming distance for all pairs of dominant haplotypes in each group, then take the sum weighted by the frequencies of the dominant haplotpes.
   Dt=Sum(pHai*pHbj*(D+0.01)))
   pHai: the frequency of Hai haplotype in sample A
   pHbj: the frequency of Hbj haplotype in sample B
   D: the hemming distance between Hai and Hbj
   Note: I added 0.01 to D so that even the two perfectly matched haplotypes are counted, at a much lower weight.
   I pooled all whole blood data in one group and all cancer data in another group for the analysis.
                MONOD V3 (140414_MiSeq)              MONOD V1 (131227_MiSeq) 
   File:140414 MiSeq WB Cancer hapGroupDiff.pngFile:131227 MiSeq WB Cancer hapGroupDiff.png
   Based on the distribution, I used the criteria (D>=2; abs(mC diff) >=0.5) to pick the informative probes. 
   I ended up getting 788 probes for MONOD V1 set, and 2804 for the MONOD V3 set.
  File:MONO1 vs MONO3 self deconvolution.png
  File:MONO1 vs MONO3 cross deconvolution.png