Clustering

Purpose

Clustering groups cells based on the neighborhood graph built by FindNeighbors().

In Seurat, clustering is usually done with FindClusters().

Clustering is not cell type annotation. It creates data-driven groups that need to be interpreted later with marker genes, metadata, and biological context.

Where It Fits

Typical position:

RunPCA() -> FindNeighbors() -> FindClusters() -> RunUMAP()

After Harmony integration:

RunPCA() -> RunHarmony() -> FindNeighbors() -> FindClusters() -> RunUMAP()

FindClusters() expects a graph that has already been created by FindNeighbors().

Run Clustering

Basic clustering:

seu <- Seurat::FindClusters(
  object = seu,
  resolution = 0.5,
  algorithm = 1,
  verbose = TRUE
)

resolution controls clustering granularity:

• lower resolution gives fewer, broader clusters
• higher resolution gives more, finer clusters

There is no universally correct resolution. It should be evaluated from the dataset and analysis goal.

algorithm = 1 uses the original Louvain algorithm.

Use A Specific Graph

If the object contains multiple graphs, specify which graph to use.

Check available graphs:

names(seu@graphs)

Then cluster on the desired SNN graph:

seu <- Seurat::FindClusters(
  object = seu,
  graph.name = "SCT_snn",
  resolution = 0.5,
  algorithm = 1,
  verbose = TRUE
)

For Harmony workflows, use the graph created from the Harmony reduction:

seu <- Seurat::FindClusters(
  object = seu,
  graph.name = "harmony_snn",
  resolution = 0.5,
  algorithm = 1,
  verbose = TRUE
)

Using graph.name avoids accidentally clustering on an older graph.

Try Multiple Resolutions

It is common to test several resolutions:

resolutions <- c(0.4, 0.6, 0.8, 1.0, 1.4)

seu <- Seurat::FindClusters(
  object = seu,
  resolution = resolutions,
  algorithm = 1,
  verbose = TRUE
)

Seurat stores each result as a metadata column.

Check clustering columns:

cluster_cols <- grep(
  pattern = "snn_res",
  x = colnames(seu@meta.data),
  value = TRUE
)

cluster_cols

Check the number and size of clusters for each resolution:

graph_prefix <- "SCT_snn"

for (res in resolutions) {
  col_name <- paste0(graph_prefix, "_res.", res)

  message("Resolution = ", res)
  message("Number of clusters: ", length(unique(seu@meta.data[[col_name]])))
  print(table(seu@meta.data[[col_name]]))
}

Choose one resolution for the working cluster identity:

selected_resolution <- 0.8
selected_col <- paste0(graph_prefix, "_res.", selected_resolution)

Seurat::Idents(seu) <- selected_col
seu$seurat_clusters <- seu@meta.data[[selected_col]]

The exact metadata column name depends on the graph used and the resolution value. For CCA/SCT integration this may be integrated_snn; for Harmony this may be harmony_snn if that graph name was set explicitly.

Check Clusters

Check cluster sizes:

table(Seurat::Idents(seu))

Check the default clustering column:

head(seu$seurat_clusters)

Store a selected clustering result with a clearer name:

seu$cluster_res_0.8 <- Seurat::Idents(seu)

This makes later marker detection and plotting less ambiguous.

Algorithm

FindClusters() supports different community detection algorithms.

The default is usually sufficient for routine analysis:

seu <- Seurat::FindClusters(
  object = seu,
  resolution = 0.5,
  algorithm = 1,
  verbose = TRUE
)

Common values include:

algorithm	Meaning
1	Louvain
2	Louvain with multilevel refinement
3	SLM
4	Leiden

Leiden clustering requires the leidenbase package.

Note

Changing dims, reduction, k.param, graph.name, resolution, or algorithm can change clusters.

Choose clustering parameters by checking marker genes, biological interpretability, cluster stability, and whether clusters are over-split or under-split.