Skip to contents

Generate a weighted nearest neighbour (WNN) graph

generate_wnn_graph_sc.Rd

This function implements the approach from Hao et al., to generate a weighted nearest neighbour graph given two 'omics modalities (for example RNA and ADT). Per-cell modality weights are computed by comparing within- and cross-modality neighbourhood distances, and these weights are used to fuse the two modalities into a single multimodal kNN graph.

The per-modality kNN graphs are always recomputed internally from the chosen embeddings at knn_range neighbours (see params_sc_wnn()); the kNN graphs previously stored on the object via find_neighbours_sc() are not reused, as WNN requires a larger candidate pool than the default neighbour search.

The result is stored in the object's other_data under "wnn", containing the fused kNN graph as a SingleCellNearestNeighbour (with a kernelised pseudo-distance metric), the sNN graph as an igraph and a table of per-cell modality weights. This graph can subsequently be used for clustering and 2D embedding via the "wnn" modality.

Usage

generate_wnn_graph_sc(
  object,
  modality_1 = "rna",
  modality_2 = "adt",
  embd_to_use_1 = "pca",
  embd_to_use_2 = "pca",
  no_embd_to_use_1 = NULL,
  no_embd_to_use_2 = NULL,
  wnn_params = params_sc_wnn(),
  full_snn = TRUE,
  pruning = 1/15,
  snn_similarity = "jaccard",
  seed = 42L,
  .verbose = TRUE
)

Arguments

object

SingleCellsMultiModal class to which to add the WNN.

modality_1

String. First modality. Defaults to "rna".

modality_2

String. Second modality. Defaults to "adt".

embd_to_use_1

String. The embedding to use for the first modality. Must be available in the object for modality_1. Defaults to "pca".

embd_to_use_2

String. The embedding to use for the second modality. Must be available in the object for modality_2. Defaults to "pca".

no_embd_to_use_1

Optional integer. Number of embedding dimensions to use for embd_to_use_1. If NULL, all will be used.

no_embd_to_use_2

Optional integer. Number of embedding dimensions to use for embd_to_use_2. If NULL, all will be used.

wnn_params

Named list. Controls the parameters for the WNN generation, see params_sc_wnn().

full_snn

Boolean. Shall the full shared nearest neighbour graph be generated that generates edges between all cells instead of between only neighbours.

pruning

Numeric. Weights below this threshold will be set to 0 in the generation of the sNN graph. Seurat uses for example 1/15 with k = 20.

snn_similarity

String. One of c("rank", "jaccard"). The Jaccard similarity calculates the Jaccard between the neighbours, whereas the rank method calculates edge weights based on the ranking of shared neighbours. For the rank method, the weight is determined by finding the shared neighbour with the lowest combined rank across both cells, where lower-ranked (closer) shared neighbours result in higher edge weights Both methods produce weights normalised to the range ⁠[0, 1]⁠.

seed

Integer. For reproducibility.

.verbose

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

Value

The SingleCellsMultiModal object with the WNN graph and per-cell modality weights added to other_data[["wnn"]].

References

Hao et al., Cell, 2021

Examples

# a fused RNA + ADT graph, both modalities reduced with PCA first
rna <- generate_single_cell_test_data()
adt <- generate_single_cell_test_data_adt()
dir <- tempfile("bixverse_mm")
dir.create(dir)
object <- load_r_data(
  SingleCellsMultiModal(dir_data = dir),
  counts = rna$counts,
  obs = rna$obs,
  var = rna$var,
  sc_qc_param = params_sc_min_quality(min_unique_genes = 5L),
  .verbose = FALSE
)
object <- add_adt_counts_sc(object, adt_counts = adt$counts, method = "clr")
object <- find_hvg_sc(object, hvg_no = 30L, .verbose = FALSE)
object <- calculate_pca_sc(object, no_pcs = 15L, .verbose = FALSE)
object <- calculate_pca_adt_sc(object, no_pcs = 10L)
object <- generate_wnn_graph_sc(object, .verbose = FALSE)
get_snn_graph(object, modality = "wnn")
#> IGRAPH eebffd3 U-W- 1000 35379 -- 
#> + attr: weight (e/n)
#> + edges from eebffd3:
#>  [1]  1-- 4  5-- 8  3-- 9  1--13  4--13 13--16  1--16 14--17  3--18 10--19
#> [11] 14--20 17--20  5--20 12--21  1--22 16--22  5--23 20--23  8--23 15--24
#> [21] 12--24  9--24 21--24  4--25  5--26 20--26 23--26 17--26 12--27 21--27
#> [31] 24--27  6--27 25--28  2--29 12--30 15--30 24--30 28--31 17--32  9--33
#> [41] 15--33 31--34  4--34 32--35 11--35 15--36 27--36 30--36 24--36  1--37
#> [51] 13--37 16--37 22--37 28--37 14--38 17--38 20--38 27--39 33--39  9--39
#> [61] 24--39 12--39 36--39 15--39 28--40 31--40  2--41 29--41 11--41 35--41
#> [71] 18--42  3--42  9--42 22--43 10--43  1--43 19--43 32--44 35--44 17--44
#> + ... omitted several edges

unlink(dir, recursive = TRUE, force = TRUE)