Skip to contents

Identify clusters of cells by graph-based clustering.

calculate_clusters.Rd

Builds a shared nearest neighbor (SNN) graph from reduced dimensions and identifies clusters over a range of resolutions.

Usage

calculate_clusters(x, ...)

# S3 method for class 'Seurat'
calculate_clusters(
  x,
  assay,
  reduction,
  num_dim,
  num_neighbors = 20,
  res = NULL,
  algorithm = 4,
  log_file = NULL,
  max_clusters = length(get_color_scheme("clusters")),
  ...
)

Arguments

x

A matrix of cell embeddings (cells as rows) or a Seurat object.

...

Arguments passed to the individual methods.

assay

Assay logged as the one this clustering run is over (Seurat method). Purely for the log message - clustering itself runs on Embeddings(x, reduction = reduction), not on assay data, so nothing here reads it back off the object. No default: DefaultAssay(x) is mutable state that need not match the assay reduction actually came from, so the caller states it explicitly rather than have the log silently report the wrong one.

reduction

Reduction to take the cell embeddings from (Seurat method). No default - always state it explicitly.

num_dim

Number of dimensions to use.

num_neighbors

Number of neighbors (k.param) used to build the SNN graph.

res

Clustering resolution(s). If NULL, a range of resolutions is used.

algorithm

Clustering algorithm: 1 = Louvain, 2 = Louvain with multilevel refinement, 3 = SLM, 4 = Leiden (the default). See Seurat::FindClusters().

log_file

Filename for the log file.

max_clusters

Drop resolutions yielding this many clusters or more. Defaults to the length of the cluster color scheme, since resolutions that exceed it cannot be plotted (Seurat method).

Value

Cluster assignments as a data frame (default method) or the Seurat object with the cluster assignments added to its metadata (Seurat method).

Details

For a Seurat object, the cluster assignments are added to the object metadata as res.<resolution> columns, the labels are converted to the C01, C02, ... convention (1-based, zero-padded, and C-prefixed to block downstream numeric coercion), and resolutions that do not yield any new clusters are dropped.