Skip to contents

cell:cell conjugate segmentation

segment_cell.Rd

This function performs segmentation of a cell:cell conjugate graph into its respective cell types and the interface between them (optional).

Usage

segment_cell(
  cg,
  w,
  k = 2L,
  detect_interface = TRUE,
  k_interface_expansion = 0L,
  keep_largest_comp = TRUE,
  min_comp_size = 10,
  spatial_smoothing_iter = 5L,
  verbose = TRUE
)

Arguments

cg

A CellGraph object

w

A matrix of NMF weights with rows corresponding to proteins and columns corresponding to the two interacting cell types.

k

The number of hops to use when computing local node abundance profiles used to compute the NNLS projection scores.

detect_interface

Whether to identify the interface between the two cell types.

k_interface_expansion

The number of hops to expand the interface between the two cell types. The default (0) classifies nodes that are directly connected to both cell types as interface. Values greater than 0 can be set to expand the interface further by including nodes that are within k hops of the initial interface nodes.

keep_largest_comp

Whether to keep only the largest connected component in each segmented cell-type graph. If FALSE, all connected components with at least min_comp_size nodes are kept.

min_comp_size

The minimum size of connected components to keep when keep_largest_comp is FALSE.

spatial_smoothing_iter

The number of iterations to perform for spatial smoothing of node neighborhoods scores. This helps smoothen out local noise in the graph and improve classification.

verbose

Print updates during the segmentation process.

Value

A CellGraph object containing the compartment labels from the segmentation as a node attribute.

Details

Requirements:

  • A CellGraph object containing the graph representation of the cell:cell conjugate. Only two cell types should be present in the graph.

  • A matrix of NMF weights (w) for the two cell types, which can be derived using the cc_protein_weights function.

The segmentation is performed in several steps:

  1. Classify nodes into "cell1" and "cell2" using Non-Negative Least Squares (NNLS). This method is described in detail below. After classification, the Largest Connected Component (LCC) is kept for each cell type to remove small disconnected subgraphs.

  2. (Optional) Identify the interface nodes between the two cell types. These are defined as nodes that are directly connected to both cell types. Optionally, the interface can be expanded by including neighboring nodes within k_interface_expansion hops. By setting detect_interface to FALSE, the interface detection step is skipped.

Nodes that cannot be classified with confidence are labeled as "other".

NNLS method for classifying membrane nodes

To segment cell:cell conjugates, we first need to fit an NMF model to the abundance matrix (or local node neighborhood abundance profiles) of the two cell populations using the cc_protein_weights function. This gives us the protein weights for each cell type (w) that we can use for classification. As a rule of thumb, a good NMF model should have high weights for proteins that are mutually exclusive between the two cell types. For instance, if we are segmenting a B cell:T cell conjugate, we would expect the NMF model to give high weights to CD10, CD20 and CD21 for the B cell identity and CD3e, CD2 and CD5 for the T cell identity. For the membrane segmentation step, we create a neighborhood profile for each node by summing the abundance of its neighbors in the graph. We then project the NMF weights (w) onto these neighborhood profiles using NNLS to get a score for each node and cell type.

Examples

library(dplyr)
se <- ReadPNA_Seurat(minimal_pna_pxl_file()) %>%
  LoadCellGraphs(cells = colnames(.)[2], verbose = FALSE)
cg <- CellGraphs(se)[[2]]

# Get cell type weights
se$cell_type <- c("Mono", "pDC", "CD4T", "CD4T", "CD4T")
w <- cc_protein_weights(
  se,
  group_by = "cell_type",
  population_1 = "Mono", population_2 = "CD4T",
  show_plot = FALSE
)
#> ! Population "Mono" has less than 20 cells, which may lead to unstable results.
#> ! Population "CD4T" has less than 20 cells, which may lead to unstable results.

cg <- segment_cell(
  cg,
  w = w
)
#> ℹ Classifying "Mono" and "CD4T" nodes using NNLS
#> ℹ Performing spatial smoothing of projection scores with 5 iterations
#> ℹ Defining interface nodes between "Mono" and "CD4T"
#> ℹ Mapping nodes to compartments