Skip to contents

Run Seurat CCA integration

seurat_cca_sc.Rd

This function implements the canonical correlation analysis (CCA) anchor integration from Stuart, et al. For each pair of batches a shared CCA embedding is computed, mutual nearest neighbours in that embedding become anchors, these are filtered in gene space, scored by shared neighbours and finally used to apply a kernel-weighted correction on the union PCA embedding. Batches are merged in the order of their pairwise anchor counts.

This port deviates from Seurat in two places. It skips the per-gene ScaleData step and works from per-cell standardised log-normalised HVG expression, and it never materialises the N1 x N2 cross-product (the canonical correlations come from a matrix-free randomised SVD). The correction runs on the embedding, not on full log-expression. In practice the anchor structure comes out close to identical at a fraction of the memory.

Usage

seurat_cca_sc(
  object,
  batch_column,
  batch_hvg_genes,
  cca_params = params_sc_seurat_cca(),
  use_precomputed_pca = FALSE,
  seed = 42L,
  .verbose = TRUE
)

Arguments

object

SingleCells class.

batch_column

String. The column with the batch information in the obs data of the class.

batch_hvg_genes

Integer vector. These are the highly variable genes, identified by a batch-aware method. Please refer to find_hvg_batch_aware_sc() for more details. These genes have to be 0-indexed!

cca_params

A list, please see params_sc_seurat_cca(). The list has the following parameters:

  • num_cc - Integer. Number of canonical correlation dimensions. The effective rank used is max(num_cc, dims).

  • dims - Integer. Dimensions used for the anchor kNN queries and size of the returned embedding.

  • k_anchor - Integer. Neighbourhood size for the anchor search.

  • k_filter - Integer. Neighbourhood size for the gene-space filter.

  • k_score - Integer. Neighbourhood size for the anchor scoring.

  • k_weight - Integer. Neighbourhood size for the kernel weights.

  • n_top_features - Integer. Top-loading genes for the gene-space filter.

  • l2_norm - Logical. L2-normalise the CCA embedding per cell.

  • sd - Numeric. Bandwidth divisor of the Gaussian kernel.

  • knn - List of kNN parameters. See params_knn_defaults() for available parameters and their defaults.

  • pca - List of PCA parameters, see params_sc_pca() for available parameters and their defaults.

use_precomputed_pca

Boolean. Should the PCA in the object be used if found. If you decide to do this, make sure that you have run the PCA on the batch-aware HVG ideally. Note that CCA still needs the PCA loadings for the gene-space filter, so this saves less work than it does for fastMNN.

seed

Integer. Random seed.

.verbose

Boolean or integer. Controls verbosity and returns run times. FALSE -> quiet, TRUE or 1L -> normal verbosity, 2L -> detailed verbosity.

Value

The object with the added "cca" embedding.

References

Stuart, et al., Cell, 2019

Examples

# CCA anchor integration over batch aware highly variable genes
sc <- demo_single_cells(
  syn_data_params = params_sc_synthetic_data(
    n_cells = 600L, n_genes = 50L, n_batches = 3L
  )
)
hvg <- find_hvg_batch_aware_sc(
  sc, hvg_no = 20L, batch_column = "batch_index", .verbose = FALSE
)
sc <- seurat_cca_sc(
  sc,
  batch_column = "batch_index",
  batch_hvg_genes = hvg$hvg_gene_idx,
  cca_params = params_sc_seurat_cca(num_cc = 10L, dims = 10L),
  .verbose = FALSE
)
dim(get_embedding(sc, "cca"))
#> [1] 600  10

unlink(sc@dir_data, recursive = TRUE, force = TRUE)