Skip to contents

build tree with AGNES clustering

Source: R/aggo_tree.R
build_aggo_tree.Rd

Agglomerative hierarchical clustering is a method of clustering that can produce a tree-like structure. Here, we construct a distance matrix based, typically, on copy number states of bins, then feed that distance matrix to AGNES clustering.

Usage

build_aggo_tree(
  reads_df,
  cut_k = 8,
  state_col = "state",
  sample_col = "cell_id",
  chrom_col = "chr",
  by_ploidy_change = FALSE,
  by_cn_change = FALSE
)

Arguments

reads_df

data.frame. bin based reads data

cut_k

int. level to cut the tree at.

state_col

str. Column to use for the clustering. Default: "state"

sample_col

str. Column of the sample labels. Default: "cell_id"

by_ploidy_change

bool. Use CN change from cell mode ploidy as the feature to cluster sample by. This can help group clones that are WGD of lower ploidy clones.

by_cn_change

bool. cluster based on changes in CN state along chromosomes.

Value

list. Two elements. $phylo: the tree; $clones: tibble of clone ids of tip labels based on tree cutting.

Details

This function will also do a preliminary tree cutting to provide groups within the tree, which could be interpreted as copy-number clones of cells. You can always re-cut the returned tree with stats::cutree(), which is all that is used internally here.

There are 3 options of values to use for clustering:

  1. states

  2. by_cn_change: changes from one state to another state are marked as 1, and all other bins as 0. In essence, it's marking the breakpoints and clustering based on shared breakpoints.

  3. by_ploidy_change: finds mode ploidy of each cell, and then bin states are converted into their difference from the mode ploidy.

Options 2 and 3 seem to work well to group WGD cells of lower ploidy clones together into clades. The also circumvent the problem that a straight distance of states doesn't always correspond to biological reality. For example, from a state of 2, a state of 3 and and state of 4 can both be 1 mutational step away, the former being an amplification event and the latter being a WGD event. Thus, bins of 3 and 4 are potentially equally distant from a bin of 2, making distance based metrics using CN states dubious (see example).

Examples

mm <- matrix(c(2, 3, 4), byrow = TRUE)
rownames(mm) <- c(paste0("cell-", LETTERS[1:3]))
dist(mm)
#>        cell-A cell-B
#> cell-B      1       
#> cell-C      2      1