Null pools are built with the same construction as the test pools
(pools_fun, e.g. random pools within cell type of null cells, or kNN
pools of null cells), scored per pool x region, and the type-7 q
quantile is taken per by (class). Random constructions are redrawn
(seed + draw) until every class has at least n_draws null pool-regions;
a deterministic construction (same pools twice, e.g. kNN) is used once.
With cv one threshold per fold from the null cells of the other folds,
plus fold 0 from all null cells. With loo the leave-one-out scores (s_rob) are computed and can be
thresholded, exactly as call_pools() scores the test pools.
Usage
pooled_thresholds(
d_null,
pools_fun,
n_draws = 20000,
by = "class",
q = params$null_q,
cv = TRUE,
states = caller_states("normal", params),
scores = c("s_depth", "s_allelic", "s_comb", if (loo) "s_rob"),
params = caller_defaults(),
max_draws = 200,
loo = FALSE,
standardize = FALSE,
min_region_pools = 1000
)Arguments
- d_null
null rows scored with
keep_rows = TRUE(cell, region, class, fold)- pools_fun
function(cells, seed) returning a pools table (pool, cell)
- n_draws
minimal null pool-regions per class
- by
grouping of the thresholds; with
sexgrp(autosome / X / Y,caller_normal()for males withpool_sex_thr), a group with fewer thanmin_region_poolsnull pool-regions takes the class threshold (pooled_fallback), and redraws are counted per class- q
null quantile
- cv
per-fold thresholds
- states
state set used for scoring
- scores
pooled scores to threshold
- params
- max_draws
maximal redraws
- loo
leave-one-out scores (adds
s_robto the defaultscores)- standardize
standardise the pooled best LLR per region (
pool_scale()of the fold's null pools) before thresholding; the fold-0 scale is returned as attributepool_scale(pass it tocall_pools()). Withparams$pool_tail_q > 0the scale carries the per-region tail factor (tail_k = params$pool_tail_k,n_eff= the fold's null cells /params$n_pool). Withparams$pool_tail_xfit(default) the tail is cross-fitted: the fold's null cells are split in two halves (seeded), pools are built within each half (n_draws/ 2 andmin_region_pools/ 2 each), each half's pools are standardised by the other half's scale before the quantile, and the returned scale is the mean of the two half scales (geometric mean of the tails)- min_region_pools
with
standardize, random constructions are redrawn until every region also has this many null pools (a stable median / MAD per region; fewer pools make the in-sample standardised threshold anti-conservative)