Skip to contents

Per pool (x by group, e.g. region): the per-state depth LLRs and the per-row allelic LLRs are summed over the pool's rows, then each state takes its best row (depth + alleles); scores as in score_cells() (s_depth = best pooled depth LLR over the states with a depth shift).

Usage

pool_scores(
  d,
  pools = NULL,
  states = caller_states("normal"),
  by = intersect(c("region", "class"), names(d)),
  loo = FALSE,
  scale = NULL
)

Arguments

d

rows scored with keep_rows = TRUE (score_cells()); with a pool column when pools is NULL

pools

NULL or a table pool, cell (a cell may be in several pools)

states

the state set used for scoring

by

grouping within a pool (default region and class when present)

loo

add the leave-one-out scores s_loo, s_rob, state_rob, lo_<state>

scale

NULL or a pool_scale() table (per-region null median / MAD of the pooled best LLR)

Value

data.table pool, by, n, cl_, s_depth, state_depth, s_allelic, state_allelic, s_comb, state_comb (+ s_loo, s_rob, state_rob, lo_; + s_comb_llr with scale)

Details

Leave-one-out (loo = TRUE): for every state row (depth + one allelic row) the largest single-cell contribution is removed from the pooled sum; s_loo is the best such row of the called state (state_comb) and s_rob = pmin(s_comb, s_loo): a pool passes on s_rob only if it still passes after dropping its most influential cell, so one strong carrier cannot make ~20 overlapping kNN pools call. state_rob = state_comb; lo_<state> holds every state's leave-one-out score.

Region scale (scale, a pool_scale() table of null pools of the same construction): the best state is chosen on the LLRs, then s_comb and s_loo are standardised per region, (s - med) / mad of the region's null pools; s_comb_llr keeps the raw best LLR.