Readcount correction for GC and mappability bias
Corrects readcounts for GC and mappability bias using the binning/loess method optimized for speed.
correctReadcount(x, mappability = 0.9, samplesize = 50000, verbose = TRUE)
Mappability threshold [0, 1] below which points are ignored during creating the correction curve.
The number of points sampled during LOESS fitting, decreasing reduces runtime and memory usage, at the expense of robustness to data randomness.
Set to FALSE it messages are not desired.
Input read counts are contained in the IRanges object, where number of reads within bins (or sometimes called windows) of defined genomic size are specified. GC content should also be computed using the exact same boundaries for each bin.
Ensure that the GC content and mappability files have been mapped to the same genome build (e.g. hg18) as the tumour and normal read libraries.
Here is a summary of the correction procedure.
Filter out bins with 0 reads and 0 GC content
Filter out bins with reads within the top and bottom 1% quantile (assumed to be outliers)
Filter out bins with GC content within the top and bottom 1% quantile
Filter out bins with a mappability score of greater than 0.9 ('mappability' argument).
Randomly sample up to 50000 ('samplesize' argument) of the remaining high-quality bins for the purposes keeping a good runtime.
The first loess (on the reads-by-GC curve) with a small span (smoothing window) is performed, obtaining typically a highly sensitive curve (follows low density tails of distribution, but gets jagged in high density center).
A second loess (on the first loess results) with a larger span is performed, recapitulating the curve in the low density tails and smoothing out the jagged regions in the high density center.
'cor.gc' is obtained by correcting each bin for GC content. The number of observed reads is divided by the number of reads predicted by the loess curve given an observed GC proportion.
Filter out just the top 1% quantile of the cor.gc bins, then _randomly_ sample up to 50000 ('samplesize' argument) bins.
A separate lowess curve is computed for mappability-by-GC (cor.gc).
'cor.map' is obtained by correcting each bin for mappability. The cor.gc value is divided by the cor.gc value predicted by the mappability lowess curve generated in the previous step.
'copy' is obtained by setting all cor.map values <= to NA (i.e. NaN), then apply log2
The original A
appended with 5 new columns:
Valid bins, which have valid read, gc, and mappability values
Ideal bins of high mappability and no outliers
Mappability corrected GC-corrected readcounts
cor.map values after log base 2
Yuval Benjamini and Terence P Speed. Summarizing and correcting the gc content bias in high-throughput sequencing. Nucleic Acids Res, 40(10):e72, May 2012.
wigsToRangedData to easily generate the proper input.
data(tumour) # Load tumour_reads tumour_copy <- correctReadcount(tumour_reads)