Train Cell Type-Specific NMF Models for Recovering Spatial EcoTypes

This function trains cell type-specific NMF (Non-Negative Matrix Factorization) models to recover SE-specific cell states from single-cell data, as part of the Spatial EcoTyper analysis workflow. It downsamples cells for training when the dataset size is large, and selects a subset of features with the highest specificity.

Usage

NMFGenerateWList(
  scdata,
  scmeta,
  CellType = "CellType",
  SE = "SE",
  scale = TRUE,
  Sample = NULL,
  balance.sample = TRUE,
  nfeature = 2000,
  nfeature.per.se = 50,
  min.cells = 20,
  downsample = 2500,
  ncores = 1,
  seed = 2024
)

Arguments

scdata: Numeric matrix containing single-cell expression data.
scmeta: Data frame containing metadata information associated with single-cell data, including cell types and spatial clusters.
CellType: Character string specifying the column name in the metadata data frame containing cell type annotations.
SE: Character string specifying the column name in the metadata data frame containing spatial ecotype annotations.
scale: Boolean specifying whether to perform univariance normalization before training the models (default: TRUE).
Sample: Character string specifying the column name in the metadata data frame containing sample annotations. If specified, the univariance normalization will be performed within each sample.
balance.sample: Boolean specifying whether to perform balance the cells from all samples before training the models (default: TRUE).
nfeature: Integer specifying the top variable features for training the models (default: 2000).
nfeature.per.se: Integer specifying the maximal number of features to select for each SE (default: 50).
min.cells: Integer specifying the minimal number of cells required for each SE cell state.
downsample: Integer specifying the number of cells per cell type (downsampling) for training the NMF models (default: 2500).
ncores: Integer specifying the number of CPU cores to use for parallel processing.
seed: Integer specifying the seed for random sampling during downsampling.

Value

A list of cell type-specific NMF models, each represented by its corresponding factorization matrix W.