grassp.tl.prune_markers

prune_markers(adata, gt_col, key_added=None, min_probability=0.9)

Remove “outliers” from marker proteins whose compartment label is not supported by their k-NN neighbourhood.

Runs competitive_diffusion() on the existing markers and retains only those whose neighbours confidently predict the same compartment label. Markers whose predicted label disagrees with their annotated label, or whose neighbourhood confidence falls below min_probability, are set to NaN in the output column. Non-marker proteins (NaN in gt_col) are always set to NaN.

This is useful for cleaning noisy or misannotated training sets before semi-supervised classification.

Parameters:

adata AnnData

anndata.AnnData with a populated neighbour graph (adata.obsp["connectivities"]).

gt_col str

.obs column containing the ground-truth compartment labels. Proteins with NaN values are treated as unannotated.

key_added str | None (default: None)

.obs column to write the pruned labels to. Defaults to "{gt_col}_pruned".

min_probability float (default: 0.9)

Minimum k-NN probability for a marker to be considered neighbourhood-consistent. Markers below this threshold are removed. Increasing this value will remove more markers, leading to more consistent but sparser annotations. Default is 0.9.

Return type:

None

Returns:

Modifies adata.obs[key_added] in place: retained markers keep their original label; removed markers are set to NaN.