Graph Construction

Purpose

Graph construction converts low-dimensional cell embeddings into a cell-cell graph.

In Seurat, this is usually done with FindNeighbors().

The graph is then used by:

• FindClusters()
• graph-based community detection
• downstream interpretation of cell neighborhoods

Where It Fits

Typical position:

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

After Harmony integration:

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

FindNeighbors() should use the same reduction and dimensions chosen for downstream analysis.

What FindNeighbors Does

FindNeighbors() uses a low-dimensional reduction to find nearby cells.

It commonly builds:

• KNN graph: k-nearest neighbor graph
• SNN graph: shared nearest neighbor graph

The SNN graph is usually the graph used by FindClusters().

Choose Reduction And Dims

The two most important inputs are:

• reduction: which embedding to use
• dims: which dimensions to include

Common choices:

Workflow	reduction	dims
Classic PCA	"pca"	selected PCs, often 1:30
SCT PCA	"pca"	selected PCs, often 1:40
CCA/SCT integration	"pca"	PCs computed from integrated assay
Harmony integration	"harmony"	selected Harmony dimensions

Choose dims from PCA diagnostics, elbow plot, variance explained, and biological interpretation.

Run FindNeighbors

For classic, SCT, or integrated assay PCA:

dims <- 1:30

seu <- Seurat::FindNeighbors(
  object = seu,
  reduction = "pca",
  dims = dims,
  k.param = 20,
  verbose = TRUE
)

For Harmony:

dims <- 1:30

seu <- Seurat::FindNeighbors(
  object = seu,
  reduction = "harmony",
  dims = dims,
  k.param = 20,
  verbose = TRUE
)

k.param = 20 is the default number of nearest neighbors used to build the KNN graph.

Check Graphs

Check graph names stored in the Seurat object:

names(seu@graphs)

Common graph names include:

RNA_nn
RNA_snn
SCT_nn
SCT_snn
integrated_nn
integrated_snn

The exact names depend on the active assay and graph construction settings.

Graph Names

If multiple graph construction steps are run in the same object, set graph names explicitly:

seu <- Seurat::FindNeighbors(
  object = seu,
  reduction = "harmony",
  dims = dims,
  k.param = 20,
  graph.name = c("harmony_nn", "harmony_snn"),
  verbose = TRUE
)

Then use the SNN graph for clustering:

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

This avoids accidentally clustering on an older graph.

Note

Graph construction is not clustering itself. It prepares the neighborhood graph that clustering uses.

Changing reduction, dims, or k.param can change the graph and therefore the final clusters.