Returns the per-cell contamination loading rho_i and the fraction of a
cell's expected counts attributed to its local ambient field. Under the
per-cell contamination model the expected count is
mu_ig = mu_bio_ig + rho_i * a_ig, where a_ig is the cross-cell-type
ambient field at cell i and rho_i is a single empirical-Bayes shrunken
loading shared across genes. The contamination fraction summarises that
second term over genes,
contamFraction_i = sum_g mu_spill_ig / sum_g mu_ig, and is the quantity
reported as contam_frac in the solver's fitting trace.
Arguments
- object
A PACEFit fitted with
contamination = "percell_hc".- ...
Unused.
Value
A data frame with one row per cell, in fitting order: cell (the
working-frame cell identifier, or the row index if the frame is unnamed),
celltype (the label the cell was fitted under), rho (the contamination
loading) and contamFraction.
Details
A high contamination fraction marks a cell whose profile is substantially
explained by its neighbours rather than by its own cell type, which is the
expected signature of a segmentation or transcript-assignment error. It is
not a complete cell-quality score: because rho_i carries a fixed
gene-direction (the local ambient), it will under-report contamination whose
direction departs from the neighbourhood average, and a cell that is wholly a
segmentation artefact may fit some incorrect cell type with a low rho_i.
Pair it with ordinary per-cell quality control.
Read contamFraction rather than rho on its own. The ambient field is
cross-cell-type, so a cell with no differently-typed neighbour inside the
technical kernel has a_ig = 0 for every gene and therefore no contamination
at all. Its rho_i is then unidentified and shrinks to the empirical Bayes
prior mean, giving every such cell the same apparently middling loading. On
the shipped example fit this is 4,589 of 7,898 cells, 84% of the tumour cells
in a tumour-dominated crop. contamFraction reports them as zero, which is
correct; rho alone does not.
Examples
fit <- readRDS(system.file("extdata", "pace_fit_example.rds", package = "PACE"))
cc <- cellContamination(fit)
head(cc)
#> cell celltype rho contamFraction
#> 1 442 Tumour 0.115044202 0.000000000
#> 2 444 B_Cell 0.002827062 0.001557514
#> 3 446 Macrophage 0.587792638 0.123800294
#> 4 449 Tumour 0.115044202 0.000000000
#> 5 451 Macrophage 0.089269111 0.102356337
#> 6 455 Tumour 0.115044202 0.000000000
# Contamination is highest where cells sit against a different type.
tapply(cc$contamFraction, cc$celltype, median)
#> B_Cell Dendritic_Cell Endothelial Macrophage Myoepithelial
#> 0.09101894 0.08295640 0.01808932 0.07264364 0.03763919
#> Stromal T_Cell Tumour
#> 0.03767385 0.07218389 0.00000000