Skip to contents

Wrapper function for parameters for AUCell

params_sc_aucell.Rd

The three statistics consume the same within-cell ranking, but weight it very differently. "wilcox" is a pure function of the gene set's rank sum, so a gene at rank 2 and one at rank 200 count for almost the same thing. "recovery" and "ap" are top-heavy. For SCENIC use "recovery", the default. "wilcox" is a bixverse addition and its flatter score does not separate into on/off populations, so it binarises badly. Ranking ties are averaged (midranks) rather than broken at random the way AUCell does it, which is why there is no need for SCENIC's trick of setting the cutoff to the 1st percentile of genes detected per cell. Undetected genes all collapse onto one rank well outside any sensible max_rank. Note the recovery AUC is normalised by max_rank * length(gene_set), matching pySCENIC and AUCell's normAUC = FALSE. Modern AUCell divides by the attainable maximum instead, so absolute values differ by a per-gene-set constant. Cell ordering within a regulon is unaffected.

Usage

params_sc_aucell(
  auc_type = c("recovery", "wilcox", "ap"),
  max_rank = NULL,
  standardise = FALSE
)

Arguments

auc_type

String. Which statistic to calculate. "wilcox" is the AUC derived from the Mann-Whitney U statistic over the full ranking, with the null at 0.5 for any gene set size. "recovery" is the recovery-curve AUC under a rank cutoff, i.e. the actual AUCell statistic of Aibar, et al. "ap" is average precision, the most top-heavy of the three, but its null tracks the gene set prevalence so raw values are not comparable across gene sets of different size unless standardise is on. One of c("recovery", "wilcox", "ap"). Defaults to "recovery".

max_rank

Numeric or NULL. Rank cutoff for "recovery", counted from the top of the within-cell ranking. If NULL, resolves to the top 5% of the gene universe, following Aibar, et al. Ignored by the other two statistics. Defaults to NULL.

standardise

Boolean. Shall each gene set's scores be z-scored across the cells. This is what makes "ap" comparable across gene sets of different size. Defaults to FALSE.

Value

A named list with the following elements:

  • auc_type - String. Which statistic to calculate. "wilcox" is the AUC derived from the Mann-Whitney U statistic over the full ranking, with the null at 0.5 for any gene set size. "recovery" is the recovery-curve AUC under a rank cutoff, i.e. the actual AUCell statistic of Aibar, et al. "ap" is average precision, the most top-heavy of the three, but its null tracks the gene set prevalence so raw values are not comparable across gene sets of different size unless standardise is on. One of c("recovery", "wilcox", "ap"). Defaults to "recovery".

  • max_rank - Numeric or NULL. Rank cutoff for "recovery", counted from the top of the within-cell ranking. If NULL, resolves to the top 5% of the gene universe, following Aibar, et al. Ignored by the other two statistics. Defaults to NULL.

  • standardise - Boolean. Shall each gene set's scores be z-scored across the cells. This is what makes "ap" comparable across gene sets of different size. Defaults to FALSE.

References

Aibar, et al., Nat Methods, 2017