metric of pairwise differences between two cells.

Inspired by MSKCC SPECTRUM paper. Bins are aligned between two cells and marked for if they have the same state Then segments of matching and non-matching runs of bins are found (filtering those smaller than a specified minimum). These segments are then re-split into 500kb bins and the difference becomes the number of matching bins divided by the number of considered bins.

Usage

pairwise_bin_difference(
  bin_df,
  cells = c(),
  min_seg_length = 2500000,
  return_pairs_matrix = FALSE
)

Arguments

bin_df

a dataframe of read bins with states. Expected columns of: cell_id, chr, start, end, state

cells

optional vector specifying cells to compare. If it's blank, all cells are compared. If it's 1 cell, then that one cell is compared to all others. If it's 2 or more, then just the specified cells are compared to each other.

min_seg_length

double. This is the minium length of matching segment bins to use when measuring similarity.

return_pairs_matrix

boolean. If TRUE, returns a pairwise matrix object of distances. This is useful to then pass to functions like hclust() and so forth. Can also do afterwards with dlptools::convert_dists_to_pairwise()

Value

tibble of cell pairs and metrics about their differences.

Details

This function is slow, and the number of pairwise comparisons grows quickly. Dramatic speed improvements can be had by setting up a parallel plan for furrr like so:

future::plan(future::multicore, workers=N_CORES_YOU_WANT)

Example for 100 cells, which is 4950 pairs, this function will take 4 minutes with 4 cores.

The returned DF is organized by each cell and the distances to each other cell (so there are some redundant comparisons, like cell 1 vs cell 2 and cell 2 vs cell 1). There is also a column "nearest_neighbour" which is a boolean identifying which comparison is the minimum distance for each cell.