Align each batch towards a goal that respects the design
Source:R/batch-correct.R
align_within_groups.RdWHY NOT ONE POOLED GOAL. Mapping each batch's distribution onto the pooled distribution of every sample assumes the batches hold the same mixture of biology. Where one batch holds mostly patients and another mostly controls, the pooled goal is a blend of the two groups, each batch is pulled towards it, and a real difference between the groups is reduced along with the batch effect. That homogeneity is exactly what anchor samples exist to avoid assuming (Schuyler et al. 2019). Without anchors, CytoNorm 2.0 tailors its goal distribution to the experimental design (Quintelier et al. 2025).
Arguments
- x
numeric vector for one marker.
- batch
batch label per cell.
- group
study group per cell, or NULL for one goal over all cells.
- goal
passed to
align_quantiles().
Details
WHAT IS DONE. Each study group is aligned separately: a batch's cells of
group g are mapped onto group g's goal, the mean of its quantiles over the
batches EVERY group was measured in (see common_batches() and
align_quantiles()). A batch effect common to the groups is removed in each
of them, and the groups keep whatever separates them. Cells with no group
are aligned towards the same batches.
WHY THE GOAL USES ONLY THE SHARED BATCHES. Each group's goal carries the mean batch effect of the batches it is built from. Built from each group's own batches, two groups measured in different batches end on different scales: with batch 2 reading one unit high, a group measured in batch 1 alone stayed where it was while a group spread over both moved half a unit up, so a correction meant to remove a batch effect put a 0.5 difference between groups of identical cells. Built from the same batches for every group, every goal carries the same batch effect. Where no batch holds every group there is no shared scale to align to, and nothing is moved.