Allelically overdispersed tiles (data-driven allelic mask)
Source:R/caller.R
allelic_noisy_tiles.RdPer tile of class size, the method-of-moments beta-binomial overdispersion of the null rows' allele votes,
rho = (sum (nA - n/2)^2 - sum n/4) / sum n (n - 1) / 4, then z_rho = (log rho - median) / MAD of log rho
over the tiles with at least min_votes votes per null row. Tiles with z_rho > k_mad are masked: a local
stretch whose votes are far more overdispersed than the genome (e.g. a 5 Mb tile with rho 0.040 vs a
genome median of 0.011 over 5 Mb tiles) and that inflates the chromosome-level allelic LLRs of the cells. The votes of the
atoms in masked tiles (atoms_in_tiles()) are then left out of the allele features
(update_allelic_features(), call_scn(params$mask_allelic = TRUE)).
Usage
allelic_noisy_tiles(
feat_long,
null = NULL,
size = "5Mb",
k_mad = params$mask_rho_mad,
min_votes = params$mask_min_votes,
params = caller_defaults(),
f0_mad = params$mask_f0_mad,
f0_min = params$mask_f0_min,
rules = "rho"
)Arguments
- feat_long
features long table (class, region, chrom, nA, n_hap) with a null definition
- null
optional logical null flag per row
- size
tile class
- k_mad
threshold in MADs of log rho (
params$mask_rho_mad)- min_votes
minimal mean votes per null row of a tile for it to be assessed (
params$mask_min_votes)- params
- f0_mad
bias threshold in MADs of d0 (
params$mask_f0_mad)- f0_min
minimal deviation d0 of a bias outlier (
params$mask_f0_min)- rules
the rules that define
masked: any of"rho","f0"
Value
data.table tile, chrom, start, end, n_rows, votes, rho, z_rho, assessed, masked, f0, d0, z_f0, masked_rho, masked_f0 (one row per tile)
Details
Shared allelic bias (params$mask_f0): per tile also the vote-weighted haplotype fraction of the null rows
f0 = sum nA / sum n_hap, its deviation d0 = |f0 - 0.5| and z_f0 = (d0 - median) / MAD of d0 over the
assessed tiles. A tile is a bias outlier (masked_f0) when it is assessed, z_f0 > f0_mad and
d0 > f0_min: a stretch whose allele fraction is off 1:1 in the null cells for reasons that are not somatic
copy number (germline CNV or segmental duplication, imprinted or allele-specific accessible regions, mapping
bias, false heterozygous SNPs), against which the 1:1 allele null of every cell is wrong. The rule sees only
the null rows, but it cannot tell such a shared bias from a real clonal event whose carriers remain among the
null cells: an event frequent enough for that looks the same and can be masked (in a simulation with a focal
event in 35 % of cells, 60 of 96 carriers stayed in the null set and all four event tiles were masked). The
mask is a diagnostic to be read with the list of masked tiles, not a filter to apply blindly to samples that
may carry clonal events.
|f0 - 0.5| covers two orientations of the allele features: under loo (block) orientation the fraction is
major-oriented (a biased tile sits above 0.5), under chromosome-anchored orientation it sits at f0 or 1 - f0.
Under mirrored orientation each block counts its major allele, so f0 is above 0.5 in every tile by an amount
set by the votes per block: the statistic measures coverage, not shared bias, and call_scn() does not apply
the bias rule (it says so in the run notes and warns).
allelic_bias_tiles() gives the same table over all cells before calling.
rules selects which flags make up masked (the tiles whose atoms are dropped): "rho" (the default and
the behaviour of earlier versions: masked = masked_rho), "f0", or both (union).