Train weights for the markers

Usage

train_weights(
  path_to_gs,
  exprs,
  level = 1,
  scaled = FALSE,
  clusters = NULL,
  range = c(1, 5),
  data_split = c(0.7, 0.2, 0.1),
  epochs = 20,
  batch_size = 32,
  run_weights_on_test = TRUE,
  cv_folds = 1,
  method = c("glmnet", "lr", "rf", "xgb", "lrp", "correlation", "uniform")
)

Arguments

path_to_gs

Path to the gene set file without weights

exprs

The expression matrix, or a seurat object (rows: genes, columns: samples/cells)

level

The level of the gene sets to train weights for if you have multiple levels of gene sets.

scaled

Whether the expression matrix is scaled

clusters

A named vector of cluster ids If exprs is a seurat object, this is ignored. The cluster ids are taken from the seurat object.

range

The range of the weights

data_split

A vector of fractions for training, validation and testing. If only two fractions are provided, no testing set will be used.

epochs

The number of epochs to train (lrp method only)

batch_size

The batch size (lrp method only)

run_weights_on_test

Whether to run the weights on the test set. Requires that data_split has three elements.

cv_folds

Number of cross-validation folds for weight estimation. When > 1, weights are averaged across folds for stability. Default is 1 (no cross-validation). Used by lr, glmnet, and lrp methods.

method

The weight learning method. One of:

"uniform": All markers get equal weight (= 1). Fast baseline.
"correlation": Pearson correlation between each marker and the binary cluster indicator. Simple, interpretable.
"lr": Logistic regression coefficients (one-vs-rest). Classic ML approach. Use cv_folds for stability.
"glmnet": Sparse logistic regression with elastic net penalty (alpha = 0.5). Automatically zeros out uninformative markers. Recommended for most users. Requires the glmnet package.
"rf": Random forest permutation importance. Captures non-linear marker interactions. Requires the ranger package.
"xgb": XGBoost gain-based feature importance. State-of-the-art tree method. Requires the xgboost package.
"lrp": Neural network + Layer-wise Relevance Propagation. Deep learning approach. Requires the keras and innsight packages.

Value

A data frame with the weights, that can be used directly by gs_prepare().