Skip to contents

Sample-level differential-state test: population x marker x group

Source: R/state-heatmaps.R
stats_marker_state.Rd

WHAT: for every (population, marker) pair, compares each non-reference cohort against the reference using the per-sample values already in population_marker_mfi.csv. One row per pair per comparison.

Usage

stats_marker_state(
  mfi,
  group_of,
  reference = NULL,
  measures = c("median_asinh", "pct_positive"),
  min_cells = 20L,
  min_n = 3L
)

Arguments

mfi

population_marker_mfi table (sample_id, population, marker, n_cells, median_asinh, pct_positive, and – after the deliverable change documented in the README – is_control and qc_status)

group_of

named character vector sample_id -> group label

reference

group every other group is compared against

measures

which per-sample quantities to test

min_cells

per-sample floor on the population's cell count

min_n

minimum samples per group for a test to be attempted

Value

data.frame, one row per population x marker x comparison x measure, or NULL when nothing is testable

Details

WHY BOTH MEASURES ARE TESTED AND REPORTED SEPARATELY: median_asinh the population's median intensity for that marker. Moves when the whole population shifts up or down – the classic "activation marker is brighter in patients" result. pct_positive the fraction of that population above the sample's own gating threshold. Moves when a SUBSET of the population turns positive while the rest does not. A bimodal shift (30% of cells go bright, 70% unchanged) moves pct_positive sharply and the median barely at all; a uniform shift does the opposite. Reporting only one measure makes the other kind of change invisible, so both are computed and the measure is a column, not an assumption.

WHY min_cells: the per-sample median of a marker over 12 cells is noise shaped like a measurement. Rows below the floor are dropped BEFORE testing rather than down-weighted, because a rank test has no weighting mechanism and silently treats a 12-cell median as equal evidence to a 4,000-cell one.

WHY THE FDR FAMILY IS THE WHOLE TABLE: this tests every marker in every population, which at a 12-marker panel and 12 populations is ~144 pairs x the number of comparisons. That is a screen, and its multiplicity has to be accounted for across the screen, not within each population.