Skip to contents

Partial Least Squares Regression for background factor regression

isotype_pls.Rd

This function performs Partial Least Squares (PLS) regression to model and extract a background score using isotype controls from single-cell data. It can optionally residualize the data against covariates before fitting the PLS model.

Usage

isotype_pls(
  object,
  isotype_markers,
  model_mat = NULL,
  remove_covariates = FALSE,
  layer = "scale.data"
)

Arguments

object

Seurat object containing single-cell data.

isotype_markers

A character vector of isotype control marker names.

model_mat

An optional matrix of covariates for residualization (cells × covariates).

remove_covariates

Logical; if TRUE, residualizes data against model_mat before PLS.

layer

String; the data layer in object to use (default is "scale.data").

Value

A list containing the PLS model, scores, and loadings.

Details

The PLS model is fit using the isotype control markers as the response variable and the non-isotype markers as predictors. A single factor is extracted from the model, which represents the background signal of all markers in each cell, captured by the markers' covariance with isotype controls. The resulting background score describes the amount of background in each cell, and can be used for downstream analyses or filtering. The model loadings can also be examined to identify which markers are most associated with the background signal. The scores and loadings are oriented such that higher scores correspond to higher background signal, and positive loadings indicate markers that covary positively with the isotype controls.

Note that this method assumes that the isotype control markers capture the background signal of interest. If there are other sources of variation that are not captured by the isotype controls, or if the isotype controls do not adequately represent the background signal, then the resulting scores may not fully capture the background variation. If remove_covariates is TRUE, the method residualizes the abundance data against the specified covariates before fitting the PLS model. This can help to isolate the background signal by removing variation associated with known covariates (e.g., different samples, cell types, or conditions).

Examples


library(pixelatorR)

seur <-
  ReadPNA_Seurat(
    minimal_pna_pxl_file(),
    overwrite = TRUE,
    load_proximity_scores = FALSE,
    verbose = FALSE
  )

pls_results <-
  isotype_pls(
    object = seur,
    isotype_markers = c("mIgG1", "mIgG2a", "mIgG2b"),
    model_mat = NULL,
    remove_covariates = FALSE,
    layer = "counts"
  )

seur$sample <- c("S1", "S1", "S2", "S2", "S2")

# Residualizing
model_mat <- model.matrix(~ 0 + seur$sample)

pls_results_resid <-
  isotype_pls(
    object = seur,
    isotype_markers = c("mIgG1", "mIgG2a", "mIgG2b"),
    model_mat = model_mat,
    remove_covariates = TRUE,
    layer = "counts"
  )