Score cells with the models trained by train_weights()

Scores each cell with the linear predictors of the per-cell-type models persisted by train_weights() with return_models = TRUE (methods "glmnet" and "lr" only): for a cell with expression x of the model features, the score of cell type t is eta_t = (Intercept)_t + sum_f coef_tf * (x_f - center_f) / scale_f, where center/scale are the per-gene centering/scaling recorded when the models were trained. Genes of the model features missing from exprs contribute 0; extra genes of exprs are ignored.

Usage

hitype_score_models(exprs, models, margin = 0)

Arguments

exprs

Input scRNA-seq expression matrix (genes x cells, the same convention as hitype_score()).

models

The model bundle returned by train_weights() with return_models = TRUE (the models element).

margin

Cells whose top-minus-second score (margins) is below margin are assigned "Unknown" instead of their top cell type. 0 (default) assigns every cell.

Value

A list with:

scores

A matrix (cells x cell types) of the linear predictors, with the cell types in the order of the model bundle.

assignments

A named vector with the top-scoring cell type of every cell (or "Unknown" for cells below the margin).

margins

A named numeric vector with the top-minus-second score of every cell.

Details

The centering/scaling is folded into the coefficients, so a sparse exprs matrix is never densified by the (x - center) / scale shift: eta_t = (Intercept)_t - sum_f b_tf * center_f + sum_f b_tf * x_f with b_tf = coef_tf / scale_f.