Train a CellTypist model

Train a CellTypist model from a Seurat object, AnnData object, or h5ad file.

Usage

TrainCellTypist(
  srt = NULL,
  adata = NULL,
  h5ad = NULL,
  assay = "RNA",
  layer = "data",
  labels,
  genes = NULL,
  transpose_input = FALSE,
  with_mean = TRUE,
  check_expression = TRUE,
  C = 1,
  solver = NULL,
  max_iter = NULL,
  n_jobs = 1,
  use_SGD = FALSE,
  alpha = 1e-04,
  use_GPU = FALSE,
  mini_batch = FALSE,
  batch_number = 100,
  batch_size = 1000,
  epochs = 10,
  balance_cell_type = FALSE,
  feature_selection = FALSE,
  top_genes = 300,
  date = "",
  details = "",
  url = "",
  source = "",
  version = "",
  model_path = NULL,
  return = c("summary", "path", "model"),
  verbose = TRUE
)

Arguments

srt: A Seurat object. Default is NULL. If provided, adata will be ignored.
adata: Optional AnnData object used as input.
h5ad: Optional path to an input .h5ad file.
assay: Which assay to use. Default is "RNA".
layer: Which layer to use. Default is data.
labels: Cell labels used for training. Can be a metadata column name when srt/adata is supplied, or a vector aligned with cells.
genes: Optional gene names. Usually inferred from the input object.
transpose_input: Whether to transpose the input matrix before training.
with_mean: Whether to center features during scaling.
check_expression: Whether to validate expected CellTypist input format.
C: Inverse regularization strength for logistic regression.
solver: Optional solver passed to CellTypist.
max_iter: Optional maximum iterations.
n_jobs: Number of CPUs used by CellTypist training.
use_SGD: Whether to use SGD training.
alpha: Regularization strength for SGD training.
use_GPU: Whether to use GPU for over-clustering. Default is FALSE.
mini_batch: Whether to enable mini-batch training.
batch_number: Number of batches per epoch for mini-batch training.
batch_size: Batch size for mini-batch training.
epochs: Number of epochs for mini-batch training.
balance_cell_type: Whether to balance cell types during mini-batch training.
feature_selection: Whether to run CellTypist feature selection.
top_genes: Number of top genes used during feature selection.
date, details, url, source, version: Free-text metadata stored in the model.
model_path: Optional output path for the trained model.
return: Return mode. One of "summary", "path", or "model". Default is "summary".
verbose: Whether to print the message. Default is TRUE.

Value

Depends on return: a summary list, the model path, or a Python model object.

Examples

if (FALSE) { # \dontrun{
data(pancreas_sub)
pancreas_sub <- standard_scop(pancreas_sub)

model_info <- TrainCellTypist(
  srt = pancreas_sub,
  labels = "SubCellType",
  model_path = tempfile(fileext = ".pkl")
)

data(panc8_sub)
genenames <- make.unique(
  thisutils::capitalize(
    rownames(panc8_sub),
    force_tolower = TRUE
  )
)
names(genenames) <- rownames(panc8_sub)
panc8_sub <- RenameFeatures(
  panc8_sub,
  newnames = genenames
)
panc8_sub <- standard_scop(panc8_sub)

pancreas_sub <- RunCellTypist(
  srt = pancreas_sub,
  model = model_info
)
CellDimPlot(
  pancreas_sub,
  group.by = c("SubCellType", "celltypist_predicted_labels")
)

panc8_sub <- RunCellTypist(
  srt = panc8_sub,
  model = model_info
)
CellDimPlot(
  panc8_sub,
  group.by = c("celltype", "celltypist_predicted_labels")
)

ht <- CellCorHeatmap(
  srt_query = panc8_sub,
  srt_ref = panc8_sub,
  query_group = "celltypist_predicted_labels",
  ref_group = "celltype",
  width = 4,
  height = 3
)
ht$plot
} # }

Usage

Arguments

Value

See also

Examples