doubletdetection.BoostClassifier#

class doubletdetection.BoostClassifier(boost_rate=0.25, n_components=30, n_top_var_genes=10000, replace=False, clustering_algorithm='phenograph', clustering_kwargs=None, n_iters=10, normalizer=None, pseudocount=0.1, random_state=0, verbose=False, standard_scaling=False, n_jobs=1)[source]#

Classifier for doublets in single-cell RNA-seq data.

Parameters

  • boost_rate (float, optional) – Proportion of cell population size to produce as synthetic doublets.

  • n_components (int, optional) – Number of principal components used for clustering.

  • n_top_var_genes (int, optional) – Number of highest variance genes to use; other genes discarded. Will use all genes when zero.

  • replace (bool, optional) – If False, a cell will be selected as a synthetic doublet’s parent no more than once.

  • self.clustering_algorithm (str, optional) – One of `[“louvain”, “leiden”,

  • implementations. ("phenograph"]`. "louvain" and leiden refer to the scanpy) –

  • clustering_kwargs (dict, optional) – Keyword args to pass directly to clusering algorithm. Note that we change the PhenoGraph ‘prune’ default to True. We also set directed=False and resolution=4 for Louvain and Leiden clustering. You must specifically include these params here to change them. random_state and key_added should not be overriden when clustering algorithm is Louvain or Leiden.

  • n_iters (int, optional) – Number of fit operations from which to collect p-values. Defualt value is 25.

  • normalizer ((sp_sparse) -> ndarray) – Method to normalize raw_counts. Defaults to normalize_counts, included in this package. Note: To use normalize_counts with its pseudocount parameter changed from the default pseudocount value to some positive float new_var, use: normalizer=lambda counts: doubletdetection.normalize_counts(counts, pseudocount=new_var)

  • pseudocount (int, optional) – Pseudocount used in normalize_counts. If 1 is used, and standard_scaling=False, the classifier is much more memory efficient; however, this may result in fewer doublets detected.

  • random_state (int, optional) – If provided, passed to PCA and used to seedrandom seed numpy’s RNG. NOTE: PhenoGraph does not currently admit a random seed, and so this will not guarantee identical results across runs.

  • verbose (bool, optional) – Set to False to silence all normal operation informational messages. Defaults to True.

  • standard_scaling (bool, optional) – Set to True to enable standard scaling of normalized count matrix prior to PCA. Recommended when not using Phenograph. Defaults to False.

  • n_jobs (int, optional) – Number of jobs to use. Speeds up neighbor computation.

all_log_p_values_#

Hypergeometric test natural log p-value per cell for cluster enrichment of synthetic doublets. Use for tresholding. Shape (n_iters, num_cells).

Type

ndarray

all_scores_#

The fraction of a cell’s cluster that is synthetic doublets. Shape (n_iters, num_cells).

Type

ndarray

communities_#

Cluster ID for corresponding cell. Shape (n_iters, num_cells).

Type

ndarray

labels_#

0 for singlet, 1 for detected doublet.

Type

ndarray, ndims=1

parents_#

Parent cells’ indexes for each synthetic doublet. A list wrapping the results from each run.

Type

list of sequences of int

suggested_score_cutoff_#

Cutoff used to classify cells when n_iters == 1 (scores >= cutoff). Not produced when n_iters > 1.

Type

float

synth_communities_#

Cluster ID for corresponding synthetic doublet. Shape (n_iters, num_cells * boost_rate).

Type

sequence of ints

top_var_genes_#

Indices of the n_top_var_genes used. Not generated if n_top_var_genes <= 0.

Type

ndarray

voting_average_#

Fraction of iterations each cell is called a doublet.

Type

ndarray

Methods table#

doublet_score()

Produce doublet scores

fit(raw_counts)

Fits the classifier on raw_counts.

predict([p_thresh, voter_thresh])

Produce doublet calls from fitted classifier

Methods#

doublet_score#

BoostClassifier.doublet_score()[source]#

Produce doublet scores

The doublet score is the average negative log p-value of doublet enrichment averaged over the iterations. Higher means more likely to be doublet.

Returns

Average negative log p-value over iterations

Return type

scores (ndarray, ndims=1)

fit#

BoostClassifier.fit(raw_counts)[source]#

Fits the classifier on raw_counts.

Parameters

raw_counts (array-like) – Count matrix, oriented cells by genes.

Sets:

all_scores_, all_log_p_values_, communities_, top_var_genes, parents, synth_communities

Returns

The fitted classifier.

predict#

BoostClassifier.predict(p_thresh=1e-07, voter_thresh=0.9)[source]#

Produce doublet calls from fitted classifier

Parameters

  • p_thresh (float, optional) – hypergeometric test p-value threshold that determines per iteration doublet calls

  • voter_thresh (float, optional) – fraction of iterations a cell must be called a doublet

Sets:

labels_ and voting_average_ if n_iters > 1. labels_ and suggested_score_cutoff_ if n_iters == 1.

Returns

0 for singlet, 1 for detected doublet

Return type

labels_ (ndarray, ndims=1)