Wrapper function for scDblFinder doublet detection parameters

Constructor for the scDblFinder parameters. This method combines cluster-aware doublet simulation with a gradient-boosted classifier trained on engineered features.

Usage

params_scdblfinder(
  n_genes = 1352L,
  doublet_ratio = 1,
  heterotypic_bias = 1,
  cluster_resolution = 1,
  cluster_iters = 10L,
  fast_cluster = FALSE,
  n_iterations = 3L,
  gbm_n_trees = 200L,
  max_depth = 4L,
  learning_rate = 0.3,
  min_samples_leaf = 20L,
  subsample_rate = 0.75,
  cv_folds = 5L,
  cv_early_stop = 2L,
  se_fraction = 1,
  include_pcs = 19L,
  expected_doublet_rate = NULL,
  cxds_genes = NULL,
  manual_threshold = NULL,
  normalisation = list(mean_center = TRUE),
  pca = list(),
  knn = list(k = 0L),
  fast_cluster_params = list()
)

Arguments

n_genes

Integer. Number of top-expressed genes to use as features. Defaults to 1352L.

doublet_ratio

Numeric. Ratio of simulated doublets to observed cells. Defaults to 1.0.

heterotypic_bias

Numeric. Fraction of simulated pairs forced to come from different clusters (0-1). Defaults to 1.0.

cluster_resolution

Numeric. Resolution for the initial Louvain clustering. Defaults to 1.0.

cluster_iters

Integer. Number of Louvain iterations per clustering step. Defaults to 10L.

fast_cluster

Boolean. Shall fast Louvain clustering be applied, i.e., k-means clustering and use the centroids for kNN graph generation and Louvain clustering with then backpropagating the membership based on centroid proximity. Defaults to FALSE.

n_iterations

Integer. Number of refinement iterations. Typically 2-3. Defaults to 3L.

gbm_n_trees

Integer. Maximum number of boosting rounds for the GBM classifier. Defaults to 200L.

max_depth

Integer. Maximum tree depth. Shallow trees (3-5) work best. Defaults to 4L.

learning_rate

Numeric. Shrinkage applied to each tree. Defaults to 0.3.

min_samples_leaf

Integer. Minimum training samples per leaf. Defaults to 20L.

subsample_rate

Numeric. Fraction of samples used per tree. Defaults to 0.75.

cv_folds

Integer. Number of cross-validation folds for boosting round selection. Defaults to 5L.

cv_early_stop

Integer. Early stopping patience per CV fold. Defaults to 2L.

se_fraction

Numeric. Multiplier on the standard error for the SE rule used in round selection. Defaults to 1.0.

include_pcs

Any. Number of leading principal components to include as classifier features. Defaults to 19L.

expected_doublet_rate

Numeric or NULL. Expected doublet rate as a percentage. If not provided, will be calculated internally. Defaults to NULL.

cxds_genes

Integer or NULL. Number of CXDS genes to consider. If not provided, defaults to 500L. Defaults to NULL.

manual_threshold

Numeric or NULL. Manual score threshold. If NULL (default), expected-rate thresholding is used. Defaults to NULL.

normalisation

List. Optional overrides for normalisation parameters. See params_norm_doublets_defaults(). See params_norm_doublets_defaults() for the available elements. Defaults to list(mean_center = TRUE).

pca

List. Optional overrides for PCA parameters. See params_pca_defaults(). See params_pca_defaults() for the available elements. Defaults to list().

knn

List. Optional overrides for kNN parameters. See params_knn_defaults(). NNDescent works better for the larger k-values often used here. See params_knn_defaults() for the available elements. Defaults to list(k = 0L).

fast_cluster_params

List. Optional overrides for the fast clustering parameters. Only relevant if fast_cluster = TRUE. See params_fast_cluster_default() for available parameters: km_type, n_centroids, kmeans_iters and batch_size. See params_fast_cluster_default() for the available elements. Defaults to list().

Value

A named list with the following elements:

• The elements of params_norm_doublets_defaults(), overridden by normalisation, spliced in at this position.

• The elements of params_pca_defaults(), overridden by pca, spliced in at this position.

• The elements of params_knn_defaults(), overridden by knn, spliced in at this position.

• The elements of params_fast_cluster_default(), overridden by fast_cluster_params, spliced in at this position.

• n_genes - Integer. Number of top-expressed genes to use as features. Defaults to 1352L.

• doublet_ratio - Numeric. Ratio of simulated doublets to observed cells. Defaults to 1.0.

• heterotypic_bias - Numeric. Fraction of simulated pairs forced to come from different clusters (0-1). Defaults to 1.0.

• cluster_resolution - Numeric. Resolution for the initial Louvain clustering. Defaults to 1.0.

• cluster_iters - Integer. Number of Louvain iterations per clustering step. Defaults to 10L.

• fast_cluster - Boolean. Shall fast Louvain clustering be applied, i.e., k-means clustering and use the centroids for kNN graph generation and Louvain clustering with then backpropagating the membership based on centroid proximity. Defaults to FALSE.

• n_iterations - Integer. Number of refinement iterations. Typically 2-3. Defaults to 3L.

• gbm_n_trees - Integer. Maximum number of boosting rounds for the GBM classifier. Defaults to 200L.

• max_depth - Integer. Maximum tree depth. Shallow trees (3-5) work best. Defaults to 4L.

• learning_rate - Numeric. Shrinkage applied to each tree. Defaults to 0.3.

• min_samples_leaf - Integer. Minimum training samples per leaf. Defaults to 20L.

• subsample_rate - Numeric. Fraction of samples used per tree. Defaults to 0.75.

• cv_folds - Integer. Number of cross-validation folds for boosting round selection. Defaults to 5L.

• cv_early_stop - Integer. Early stopping patience per CV fold. Defaults to 2L.

• se_fraction - Numeric. Multiplier on the standard error for the SE rule used in round selection. Defaults to 1.0.

• include_pcs - Any. Number of leading principal components to include as classifier features. Defaults to 19L.

• expected_doublet_rate - Numeric or NULL. Expected doublet rate as a percentage. If not provided, will be calculated internally. Defaults to NULL.

• manual_threshold - Numeric or NULL. Manual score threshold. If NULL (default), expected-rate thresholding is used. Defaults to NULL.

• cxds_genes - Integer or NULL. Number of CXDS genes to consider. If not provided, defaults to 500L. Defaults to NULL.