scib_metrics.kbet_per_label

scib_metrics.kbet_per_label(X, batches, labels, alpha=0.05, diffusion_n_comps=100, return_df=False)

Compute kBET score per cell type label as in [Luecken et al., 2022].

This approximates the method used in the original scib package. Notably, the underlying kbet might have some inconsistencies with the R implementation. Furthermore, to equalize the neighbor graphs of cell type subsets we use diffusion distance approximated with diffusion maps. Increasing diffusion_n_comps will increase the accuracy of the approximation.

Parameters:

X NeighborsResults

A NeighborsResults object.

batches ndarray

Array of shape (n_cells,) representing batch values for each cell.

alpha float (default: 0.05)

Significance level for the statistical test.

diffusion_n_comps int (default: 100)

Number of diffusion components to use for diffusion distance approximation.

return_df bool (default: False)

Return dataframe of results in addition to score.

Return type:

float | tuple[float, DataFrame]

Returns:

kbet_scorefloat

Kbet score over all cells. Higher means more integrated, as in the kBET acceptance rate.

dfDataFrame

Dataframe with kBET score per cell type label.

Notes

This function requires X to be cell-cell connectivities, not distances.