Assigns each cell to a cellular neighbourhood ("niche") by clustering cells
on the phenotype composition of their local spatial neighbourhood. This is
the windowed-neighbourhood approach popularised for multiplexed imaging:
for every cell the phenotype frequencies among its k nearest
neighbours form a composition vector, and those vectors are clustered.
Usage
CellularNeighbourhoods(
object,
n_neighbourhoods = 8L,
k = 20L,
phenotype_col = "phenotype",
seed = NULL
)Arguments
- object
A
SpatialCellData-classobject with a phenotype column.- n_neighbourhoods
Integer. Number of neighbourhoods to find. Default
8.- k
Integer. Neighbourhood size (nearest neighbours per cell). Default
20.- phenotype_col
Character. Metadata column of phenotype labels. Default
"phenotype".- seed
Integer or
NULL. Random seed for k-means.
Value
The object with a neighbourhood column in meta_data;
the neighbourhood-by-phenotype composition matrix is stored in
object@spatial$neighbourhood_composition. Availability is recorded
in object@spatial$neighbourhood_availability: per_cell retains cell/sample
IDs, input row_index and status (defined or no_neighbors); by_sample
reports n_total, n_defined and n_unavailable. The weighting field
records cell. Existing metadata other than the neighbourhood label is
preserved.
Details
Sample identifiers and phenotype labels must be complete, nonblank
vectors. Counts and k must be positive whole numbers; malformed inputs are
rejected before clustering or changing the random state. Cells in singleton
samples have no neighbour composition: they retain their row and receive
an NA neighbourhood label, and are excluded from clustering. Centres are
cell-weighted means of defined composition vectors, not equal-sample means.
The requested number of centres cannot exceed the distinct defined vectors.
With no defined vectors (including empty input), no clustering or random
draws occur and the centre matrix has zero rows.