Skip to contents

Wrapper function for consensus NMF parameters

params_nmf_consensus.Rd

Controls the clustering step of consensus NMF: the components of every restart are pooled, outliers are dropped by local density, and the survivors are k-means clustered into k groups whose median becomes the consensus factor.

Usage

params_nmf_consensus(
  consensus_target = c("h", "w"),
  n_neighbours = 0L,
  density_threshold = 0.5,
  kmeans_iters = 100L,
  kmeans_n_init = 3L
)

Arguments

consensus_target

String. "h" clusters the gene programmes (the spectra), which is what cNMF does and what you almost always want. "w" clusters in sample space instead, which on single cell data means cell space: it pools a dense (k * n_runs) x n_cells matrix and runs an exhaustive cosine search over it, so it gets expensive fast. One of c("h", "w"). Defaults to "h".

n_neighbours

Integer. Neighbours used for the local density estimate. 0L picks ceiling(0.3 * n_runs) for you. Defaults to 0L.

density_threshold

Numeric. Components whose mean cosine distance to their neighbours exceeds this are dropped as unstable. Cosine distance cannot exceed 2, so any value >= 2 disables the filter entirely. Defaults to 0.5.

kmeans_iters

Integer. Maximum k-means iterations. Defaults to 100L.

kmeans_n_init

Integer. Number of k-means restarts. Defaults to 3L.

Value

A named list with the following elements:

  • consensus_target - String. "h" clusters the gene programmes (the spectra), which is what cNMF does and what you almost always want. "w" clusters in sample space instead, which on single cell data means cell space: it pools a dense (k * n_runs) x n_cells matrix and runs an exhaustive cosine search over it, so it gets expensive fast. One of c("h", "w"). Defaults to "h".

  • n_neighbours - Integer. Neighbours used for the local density estimate. 0L picks ceiling(0.3 * n_runs) for you. Defaults to 0L.

  • density_threshold - Numeric. Components whose mean cosine distance to their neighbours exceeds this are dropped as unstable. Cosine distance cannot exceed 2, so any value >= 2 disables the filter entirely. Defaults to 0.5.

  • kmeans_iters - Integer. Maximum k-means iterations. Defaults to 100L.

  • kmeans_n_init - Integer. Number of k-means restarts. Defaults to 3L.

References

Kotliar et al., eLife, 2019