Skip to contents

Calculate LISI scores on any label

rs_lisi.Rd

[Experimental] Computes the Local Inverse Simpson's Index on the kNN graph: the effective number of labels in each cell's neighbourhood. On batch labels this is iLISI (higher is better mixing), on cell type labels cLISI (lower is better separation). Both come back rescaled to [0, 1], higher is better, as in scIB.

Usage

rs_lisi(knn_mat, knn_dist, labels, perplexity, verbose)

Arguments

knn_mat

Integer matrix. The rows represent the cells and the columns the neighbour indices (0-indexed!).

knn_dist

Numeric matrix or NULL. The kNN distances, same shape as knn_mat. If provided, neighbours are weighted with a perplexity-calibrated Gaussian kernel as in Korsunsky et al.; if NULL, neighbours are weighted uniformly.

labels

Integer vector. The label (batch or cell type) per cell. The codes need not be 0-based or contiguous.

perplexity

Numeric or NULL. Perplexity for the weighted version. NULL defaults to 30; values above k are clamped to k. Ignored if knn_dist is NULL.

verbose

Boolean. Controls verbosity of the function.

Value

A list with the following items

  • per_cell - Per-cell LISI scores

  • mean_lisi - Mean LISI

  • median_lisi - Median LISI

  • n_labels - Number of distinct labels

  • ilisi_norm - Median LISI rescaled as iLISI, (median - 1) / (n - 1)

  • clisi_norm - Median LISI rescaled as cLISI, (n - median) / (n - 1)

References

Korsunsky, et al., Nat Methods, 2019; Luecken, et al., Nat Methods, 2022