Skip to contents

Run SCENIC GRN inference

scenic_grn_sc.Rd

Runs SCENIC GRN inference on the provided genes using the specified transcription factors as predictors. Returns a ScenicGrn object containing the TF-gene importance matrix for further processing.

Usage

scenic_grn_sc(
  object,
  tf_ids,
  scenic_params = params_scenic(),
  genes_to_take = NULL,
  cells_to_take = NULL,
  streaming = NULL,
  random_seed = 42L,
  .verbose = TRUE
)

Arguments

object

SingleCells, MetaCells (or potentially other) class.

tf_ids

Character vector. Transcription factor gene identifiers to use as predictors. Must be a subset of gene identifiers present in the object.

scenic_params

List. SCENIC parameters, see params_scenic().

genes_to_take

Optional character vector. Target gene identifiers. If NULL, genes are selected automatically via scenic_gene_filter_sc() using the min_counts and min_cells thresholds in scenic_params.

cells_to_take

Optional string vector. Cell identifiers to restrict to. If NULL, defaults to all filtered cells in the class.

streaming

Optional Boolean. Shall the data be streamed in. Useful for larger data sets where you wish to avoid loading in the whole data. If NULL, will automatically detect. Ignored when applied to MetaCells.

random_seed

Integer. Used for reproducibility. Defaults to 42L.

.verbose

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

Value

A ScenicGrn object.

Details

TF identifiers that are not found in the object's gene list are silently dropped with a warning indicating how many were removed. TF indices are intersected with the target gene indices so that TFs not passing the gene filter are excluded from the predictor set but remain as potential targets if present in genes_to_take. You have the option to generate the TF-gene importance values with three distinct methods. For the random_forest and the extratrees version, a batching strategy is applied in the default settings. Correlated genes are identified and clustered together via k-means clustering on the feature loadings of the PCA. These are then divided into batches of gene_batch_size and the regression learners are leveraging multi-target regression to fit all genes in the batch in one go. This massively accelerates the algorithm and the importance values per gene-TF pair are calculated then individually. Due to the batching by similar gene, the signal dilution is limited. If you wish to run the traditional approach, you can set gene_batch_size to 1L or use the grnboost2 learner that can only fit one gene at a given time.

Examples

# TF to gene importances from a small random forest
sc <- demo_single_cells()
res <- scenic_grn_sc(
  sc,
  tf_ids = get_gene_names(sc)[1:5],
  scenic_params = params_scenic(
    min_counts = 100L,
    learner_params = list(n_trees = 20L)
  ),
  .verbose = FALSE
)
res
#> ScenicGrn (GRN results)
#>   No genes:                 50 
#>   No TFs:                   5 
#>   Applied learner:          randomforest 
#>   TF to gene generated:     FALSE 
#>   CisTarget res generated:  FALSE 

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