cellassign: Annotate cells to cell types using cellassign

In Irrationone/cellassign: Automated, probabilistic assignment of scRNA-seq to cell types

Description

Automatically annotate cells to known types based on the expression patterns of a priori known marker genes.

Usage

cellassign(
  exprs_obj,
  marker_gene_info,
  s = NULL,
  min_delta = 2,
  X = NULL,
  B = 10,
  shrinkage = TRUE,
  n_batches = 1,
  dirichlet_concentration = 0.01,
  rel_tol_adam = 1e-04,
  rel_tol_em = 1e-04,
  max_iter_adam = 1e+05,
  max_iter_em = 20,
  learning_rate = 0.1,
  verbose = TRUE,
  sce_assay = "counts",
  return_SCE = FALSE,
  num_runs = 1,
  threads = 0
)

Arguments

exprs_obj	Either a matrix representing gene expression counts or a SummarizedExperiment. See details.
marker_gene_info	Information relating marker genes to cell types. See details.
s	Numeric vector of cell size factors
min_delta	The minimum log fold change a marker gene must be over-expressed by in its cell type
X	Numeric matrix of external covariates. See details.
B	Number of bases to use for RBF dispersion function
shrinkage	Logical - should the delta parameters have hierarchical shrinkage?
n_batches	Number of data subsample batches to use in inference
dirichlet_concentration	Dirichlet concentration parameter for cell type abundances
rel_tol_adam	The change in Q function value (in pct) below which each optimization round is considered converged
rel_tol_em	The change in log marginal likelihood value (in pct) below which the EM algorithm is considered converged
max_iter_adam	Maximum number of ADAM iterations to perform in each M-step
max_iter_em	Maximum number of EM iterations to perform
learning_rate	Learning rate of ADAM optimization
verbose	Logical - should running info be printed?
sce_assay	The assay from the input#' SingleCellExperiment to use: this assay should always represent raw counts.
return_SCE	Logical - should a SingleCellExperiment be returned with the cell type annotations added? See details.
num_runs	Number of EM optimizations to perform (the one with the maximum log-marginal likelihood value will be used as the final).
threads	Maximum number of threads used by the algorithm (defaults to the number of cores available on the machine)

Details

Input format exprs_obj should be either a SummarizedExperiment (we recommend the SingleCellExperiment package) or a cell (row) by gene (column) matrix of raw RNA-seq counts (do not log-transform or otherwise normalize).

marker_gene_info should either be

• A gene by cell type binary matrix, where a 1 indicates that a gene is a marker for a cell type, and 0 otherwise

• A list with names corresponding to cell types, where each entry is a vector of marker gene names. These are converted to the above matrix using the marker_list_to_mat function.

Cell size factors If the cell size factors s are not provided they are computed using the computeSumFactors function from the scran package.

Covariates If X is not NULL then it should be an N by P matrix of covariates for N cells and P covariates. Such a matrix would typically be returned by a call to model.matrix with no intercept. It is also highly recommended that any numerical (ie non-factor or one-hot-encoded) covariates be standardized to have mean 0 and standard deviation 1.

cellassign A call to cellassign returns an object of class cellassign. To access the MLE estimates of cell types, call fit$cell_type. To access all MLE parameter estimates, call fit$mle_params.

Returning a SingleCellExperiment

If return_SCE is true, a call to cellassign will return the input SingleCellExperiment, with the following added:

• A column cellassign_celltype to colData(sce) with the MAP estimate of the cell type

• A slot sce@metadata$cellassign containing the cellassign fit. Note that a SingleCellExperiment must be provided as exprs_obj for this option to be valid.

Value

An object of class cellassign. See details

Examples

data(example_sce)
data(example_marker_mat)

fit <- em_result <- cellassign(example_sce[rownames(example_marker_mat),],
marker_gene_info = example_marker_mat,
s = colSums(SummarizedExperiment::assay(example_sce, "counts")),
learning_rate = 1e-2,
shrinkage = TRUE,
verbose = FALSE)

Irrationone/cellassign documentation built on April 23, 2020, 3:10 p.m.