Classifies modules as outliers, redundant, or valid based on pseudotime spread, UMAP distance and coverage heuristics. A module is classified as outlier if its IQR and UMAP distance have a MAD z-score greater than 3.5 in absolute value. Modules that fit in the 85% of the pseudotime IQR or the average UMAP distance of the baseline (the entire dataset) are considered outliers as well. Redundancy is based on the percentage of new cells covered by the module compared to the already covered cells (the modules are ordered based on their median UMAP distance and IQR pseudotime, so the best modules are evaluated first). A second round allows redundant modules to be labeled as non-redundant if they provide a significant percentage of new cells compared to the already covered population (by default, the threshold is the median F1 score of the non-redundant modules). The third run removes redundancy for modules whose population either match 1 on 1, or are a subset of modules that are already non-redundant. By default, the overlap should be pretty high (95%) for the redundancy removal to happen.
Usage
detect_outlier(
modules_stats,
cell_masks,
psd_value,
thresh_psd_good = NULL,
thresh_psd_bad = NULL,
umap_dist_threshold = NULL,
overlap_threshold = 0.95
)Arguments
- modules_stats
A data frame of module summary statistics.
- cell_masks
A logical matrix of module-to-cell memberships.
- psd_value
A numeric pseudotime vector.
- thresh_psd_good
Threshold for accepting a module as non-redundant. If NULL, it will be set to the 25% quantile of the IQR pseudotime of the non-outlier modules.
- thresh_psd_bad
Threshold for flagging a module as an outlier. If NULL, it will be set to the 85% quantile of the IQR pseudotime of the entire dataset.
- umap_dist_threshold
Optional threshold for median UMAP distance.
- overlap_threshold
Overlap threshold that allows redundant modules to become non-redundant if most of their cells already belong to the population of an already non-redundant module.