thisutils (version 0.4.4)

compute_lisi: Compute Local Inverse Simpson's Index (LISI)

Description

Compute per-cell Local Inverse Simpson's Index (LISI) scores for one or more categorical variables. This is a clean-room reimplementation of the immunogenomics/LISI.

Usage

compute_lisi(
  X,
  meta_data,
  label_colnames,
  perplexity = 30,
  nn_eps = 0,
  use_rann = TRUE,
  nn_method = c("auto", "rann", "fnn", "hnsw", "exact"),
  nn_threads = 0,
  hnsw_ef = 100,
  tol = 1e-05,
  max_iter = 50
)

Value

A data frame with one row per cell and one column per label.

Arguments

X

A matrix-like object with cells in rows and embedding/features in columns.

meta_data

A data frame with one row per cell.

label_colnames

Character vector of column names in meta_data to evaluate.

perplexity

Effective neighborhood size. Defaults to 30.

nn_eps

Approximation factor passed to RANN::nn2 when RANN is available and use_rann = TRUE. Defaults to 0.

use_rann

Whether to prefer RANN::nn2 over FNN::get.knn when nn_method = "auto" decides not to use the package's built-in exact C++ backend. Defaults to TRUE.

nn_method

Nearest-neighbor backend. Defaults to "auto", which uses a simple heuristic: low-dimensional inputs use the package's exact C++ search, while larger/higher-dimensional inputs fall back to RANN, then FNN, then the built-in exact C++ backend. Set to "hnsw" to use RcppHNSW for a faster, approximate search.

nn_threads

Number of threads used by the HNSW backend. Defaults to 0 which lets RcppHNSW choose automatically.

hnsw_ef

Search breadth used by the HNSW backend. Larger values are more accurate but slower. Defaults to 100.

tol

Tolerance used in the binary search for the target perplexity. Defaults to 1e-5.

max_iter

Maximum number of binary-search iterations. Defaults to 50.

References

Korsunsky I, Millard N, Fan J, et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nature Methods (2019). https://www.nature.com/articles/s41592-019-0619-0

LISI reference implementation: https://github.com/immunogenomics/LISI

Examples

set.seed(1)
X <- rbind(
  matrix(stats::rnorm(100, mean = -2), ncol = 5),
  matrix(stats::rnorm(100, mean = 2), ncol = 5)
)
meta_data <- data.frame(
  batch = rep(c("A", "B"), each = 20),
  group = sample(c("g1", "g2"), 40, replace = TRUE)
)

res <- compute_lisi(
  X, meta_data,
  c("batch", "group"),
  perplexity = 10
)
head(res)
boxplot(res)