June 27, 2011

This directory contains genotype calls and likelihoods for a set of structural variations on chr20.

The deletion sites list was called by Genome STRiP using 929 Illumina samples with paired-end data,
from the June 1 call set (N=19,086 genome wide, of which 464 are on chr20).
These sites were then genotyped in 946 samples with Illumina data, using read depth and aberrant read pairs only
(i.e. not using split reads which requires precise breakpoints).

The sites list is further subsetted into a set of sites with reliable genotype likelihoods that should be used
for integration with SNPs and indels.  Both the complete site list (N=464) and the reliable site list (N=314)
are provided as separate VCF files.  In the complete site list, the FILTER field indicates which sites have
reliable genotype likelihoods (those with a value of "PASS").  Ignoring redundant sites (N=14), 69.8% of the
sites were deemed genotypable (314/450).

The complete set of sites are in:
ALL.chr20.BI_genome_strip.all.20101123.dels.genotypes.vcf.gz
The passing sites are in:
ALL.chr20.BI_genome_strip.pass.20101123.dels.genotypes.vcf.gz

The filtering process was performed as follows:

After initial genotyping, redundant calls were removed using genotype-aware duplicate detection,
which flagged 14 sites as duplicates (i.e. redundant calls of the same underlying event).

The remaining sites were then further filtered to determine those with well-calibrated read depth.
The test for well-calibrated read depth was based mostly on the degree of separation between the copy number
clusters assigned during genotyping.  This is indicated by the GSCLUSTERSEP metric in the info field, which
is the mean per-sample Mahalanobis distance between the CN1 and CN2 clusters.  It is possible to experiment
with using more separated or less separated clusters by adjusting the GSCLUSTERSEP filter, which was set at 2.0
(i.e. roughly two SDs of separation between the CN1 and CN2 clusters).

The total set of filters were applied to select the set of sites that should be used for integrating the
genotype likelihoods:

  CLSEP: GLCLUSTERSEP < 2.0 (cluster separation less than roughly two standard deviations)
  CLMEAN: GSM1 <= 0.5 || GSM1 >= 1.5 (overall read depth is too high or too low at this site)
  ELENGTH: GSELENGTH <= 100 (amount of reliably alignable sequence is less than 100bp)
  DUP: GSDUPLICATESCORE <= 2.0 (genotype-aware duplicate detection flagged this site as a duplicate)
  IBC: GSINBREEDINGCOEFF < -0.9 (remove one site with excess het calls)

For more information, contact Bob Handsaker (handsake at broadinstitute.org).