In keyes-timothy/tidytof: Analyze High-dimensional Cytometry Data Using Tidy Data Principles

knitr::opts_chunk$set(
    collapse = TRUE,
    comment = "#>",
    fig.height = 4,
    fig.width = 4
)

library(tidytof)
library(dplyr)
library(ggplot2)

count <- dplyr::count

Often, high-dimensional cytometry experiments collect tens or hundreds or millions of cells in total, and it can be useful to downsample to a smaller, more computationally tractable number of cells - either for a final analysis or while developing code.

To do this, {tidytof} implements the tof_downsample() verb, which allows downsampling using 3 methods: downsampling to an integer number of cells, downsampling to a fixed proportion of the total number of input cells, or downsampling to a fixed cellular density in phenotypic space.

Downsampling with tof_downsample()

Using {tidytof}'s built-in dataset phenograph_data, we can see that the original size of the dataset is 1000 cells per cluster, or 3000 cells in total:

data(phenograph_data)

phenograph_data |>
    dplyr::count(phenograph_cluster)

To randomly sample 200 cells per cluster, we can use tof_downsample() using the "constant" method:

phenograph_data |>
    # downsample
    tof_downsample(
        group_cols = phenograph_cluster,
        method = "constant",
        num_cells = 200
    ) |>
    # count the number of downsampled cells in each cluster
    count(phenograph_cluster)

Alternatively, if we wanted to sample 50% of the cells in each cluster, we could use the "prop" method:

phenograph_data |>
    # downsample
    tof_downsample(
        group_cols = phenograph_cluster,
        method = "prop",
        prop_cells = 0.5
    ) |>
    # count the number of downsampled cells in each cluster
    count(phenograph_cluster)

And finally, we might also be interested in taking a slightly different approach to downsampling that reduces the number of cells not to a fixed constant or proportion, but to a fixed density in phenotypic space. For example, the following scatterplot demonstrates that there are certain areas of phenotypic density in phenograph_data that contain more cells than others along the cd34/cd38 axes:

rescale_max <-
    function(x, to = c(0, 1), from = range(x, na.rm = TRUE)) {
        x / from[2] * to[2]
    }

phenograph_data |>
    # preprocess all numeric columns in the dataset
    tof_preprocess(undo_noise = FALSE) |>
    # plot
    ggplot(aes(x = cd34, y = cd38)) +
    geom_hex() +
    coord_fixed(ratio = 0.4) +
    scale_x_continuous(limits = c(NA, 1.5)) +
    scale_y_continuous(limits = c(NA, 4)) +
    scale_fill_viridis_c(
        labels = function(x) round(rescale_max(x), 2)
    ) +
    labs(
        fill = "relative density"
    )

To reduce the number of cells in our dataset until the local density around each cell in our dataset is relatively constant, we can use the "density" method of tof_downsample:

phenograph_data |>
    tof_preprocess(undo_noise = FALSE) |>
    tof_downsample(method = "density", density_cols = c(cd34, cd38)) |>
    # plot
    ggplot(aes(x = cd34, y = cd38)) +
    geom_hex() +
    coord_fixed(ratio = 0.4) +
    scale_x_continuous(limits = c(NA, 1.5)) +
    scale_y_continuous(limits = c(NA, 4)) +
    scale_fill_viridis_c(
        labels = function(x) round(rescale_max(x), 2)
    ) +
    labs(
        fill = "relative density"
    )

Thus, we can see that the density after downsampling is more uniform (though not exactly uniform) across the range of cd34/cd38 values in phenograph_data.

Additional documentation

For more details, check out the documentation for the 3 underlying members of the tof_downsample_* function family (which are wrapped by tof_downsample):

• tof_downsample_constant
• tof_downsample_prop
• tof_downsample_density

Session info

sessionInfo()

keyes-timothy/tidytof documentation built on May 7, 2024, 12:33 p.m.