Per-cell evidence on the MEDICC2 pieces: depth and oriented molecule votes
Source:R/medicc2_cells.R
cell_piece_data.RdDepth (the per-cell analogue of clone_profiles()): the cell's 1 Mb window counts summed over the piece (windows
fully inside it; region_features()), log2ratio = log2((x + 0.5) / (e + 0.5)) relative to the reference
pool's pseudobulk of the piece (ref_center = "pseudobulk", log2((sum x + 0.5) / (sum e + 0.5)) over the
reference cells, as the clone profiles; "mean" = the mean of the reference cells' log2 ratios, as
assign_cells(), which a few reference cells with a large negative offset pull down), then (center) minus the
cell's own baseline: the window-weighted median of its log2 ratios over the autosomal pieces (the clone profiles are
centred on the clone's autosomal median in the same way, so states are relative to the baseline). SD: the cell's
null SD of the piece's size class (sd_null of the piece's sd_class); with sd_scale, a piece shorter than its
tile class (e.g. a 2 Mb piece scored with the 5 Mb SD) gets sd_null * sqrt(class size / covered window length)
(the independent-window limit); pieces on haploid chromosomes get haploid_sd x that; s_sys (systematic depth
bias of a region between the scored cells and the reference pool, the clone model's s_sys) is added in
quadrature: sd = sqrt(sd_scaled^2 + s_sys^2).
z = log2ratio / pmax(sd, sd_floor); pieces without a window have NA depth (no evidence).
Alleles: the cell's molecule votes on the piece (atom start in [start, end)) per phase set, oriented by the
block table: oA = votes of the cn_a haplotype (nA when sign > 0, nB when sign < 0), oN = oriented votes.
With loo, the reference clone's pooled votes are first reduced by the cell's own votes when the cell was one of
the cells that oriented the block (loo maps cells to the clone whose pool they joined; only blocks whose ref
is that clone are affected): the orientation is then the one the other cells give, and the cell's own votes cannot
confirm themselves. This is the only use of a cell's clone membership, and it only removes the cell's own data
from the orientation input.
Usage
cell_piece_data(
features,
pieces,
blocks,
cells = NULL,
ref_cells = NULL,
loo = NULL,
params = caller_defaults(),
sd_scale = TRUE,
ref_center = c("pseudobulk", "mean"),
center = TRUE,
s_sys = 0,
haploid_sd = sqrt(2)
)Arguments
- features
scn_features list (
win,long,regions,votes,atoms,cells; as the tumour mode reads it)- pieces
piece table (
piece, chrom, start, end, one row per piece or per clone x piece)- blocks
block orientation (
piece, ps, signand, forloo,ref, nA_ref, nB_ref)- cells
cells to score (default all cells of
features)- ref_cells
reference cells of the depth centring (>= 10 used; NULL = no centring)
- loo
NULL, or data.table
cell, clone: the cells whose votes formed each reference clone's pool- params
caller_defaults()(sd_floor)- sd_scale
scale the null SD of pieces shorter than their tile class
- ref_center
reference centring:
pseudobulkormean- center
subtract the cell's autosomal median (window-weighted over pieces)
- s_sys
systematic SD (log2) added in quadrature to the per-cell SD
- haploid_sd
factor on the null SD of pieces on haploid chromosomes (male chrX / chrY):
sd_nullcomes from diploid regions, and one copy gives half the fragments, so z on haploid pieces is overdispersed