The null cells of normal mode are selected for having no chromosome / arm call (select_null_cells()), the
thresholds are trained on their rows and the budget is verified on the same rows by fold cross-validation, so a
selection that truncates the null tail (a too thin parametric null, cells with a noisier than modelled genome)
goes unseen. This check splits the autosomes into two halves (odd / even chromosome number), selects the null
cells on the rows of one half only, trains the per-cell thresholds as usual on those rows (caller_normal() with
the selected cells as null cells and the half as null_chroms: depth t null, rho, fold-cross-validated budget
thresholds, the run's calib options), and counts the exceedances (s_comb above its threshold) on the null
regions of the OTHER half of the same cells: rows that neither the selection nor the thresholds saw. The two halves
are swapped and the counts combined.
Usage
null_selection_check(
feat,
params = caller_defaults(),
multiplets = NULL,
regions = NULL,
depth_model = c("t", "gauss"),
select_args = list(),
null_chroms = paste0("chr", 1:22),
false_calls = NULL,
conf = 0.95
)Arguments
- feat
features long table (as for
caller_normal())- params
- multiplets
cells never selected as null cells
- regions
regions table (default
regions_from_features())- depth_model
"t" or "gauss"
- select_args
extra arguments for
select_null_cells()(itschromsis set per half)- null_chroms
the chromosomes whose rows can be null rows in the run; their numbered autosomes are split
- false_calls
optional
null_false_calls()ratesof the run, added ascv_per_1000(the run's in-sample cross-validated rate on all null rows)- conf
confidence level of the intervals
Value
data.table class, stratum (all = over the strata), n_rows (measured rows), n_exceed, per_1000 (null and
real-event exceedances per 1,000 measured rows: an upper bound of the null rate),
lo_per_1000, hi_per_1000, nominal_per_1000, insample_rows, insample_exceed, insample_per_1000 (the
cross-validated rate on the selection half's null rows of the same passes), cv_per_1000 (with false_calls);
attribute halves (half, select_chroms, measure_chroms, n_cells, n_null_cells)
Details
Cells are selected on one half only, so a cell that carries a real event on the measured half stays in and the
rows of that event count as exceedances: per_1000 is the null exceedance rate plus the rate of real events on
the measured half, an upper bound of the null rate. A rate clearly above the nominal 1 - null_q (and above the
in-sample cross-validated rate on the selection half, insample_per_1000) has two possible causes: a selected
null that is cleaner than the cells it is applied to (the per-cell false load is then underestimated), or real
events in the cells. Read it next to the share of cells with a called event: in a sample without real events
the excess measures the selection, in a sample with them it is an upper bound. The halves carry different
chromosomes, so a class whose regions differ in their null spread (e.g. chromosomes of very different size) also
shows up.
Confidence intervals are exact binomial (Clopper-Pearson) over rows; rows of one cell overlap (a chromosome and its
arms), so they are somewhat too narrow. Costs about one more caller_normal() scoring pass (two half passes
without pools). With calib_rho = "chrom" the measured half has no null rows and takes the global rho.