UMAP on 30k cells with different values for min_dist and spread

plot-umap-experiment.R

library(uwot)
library(scattermore)
library(foreach)
library(doParallel)
library(data.table)
library(ggplot2)
library(scales)
library(glue)

# This code snippet assumes we have these objects:
#
# pca
# A matrix with PCA scores, one column for each PC.
#
# nn_method
# A list with two matrices "idx" and "dist"
# Each matrix has nrow = cells, ncol = num neighbors
# idx has the ids of neighbor cells
# dist has euclidean distances to each neighbor cell
#
# obs
# A table with one row per cell, and a column called "cluster"

umap_params <- expand.grid(
  spread   = c(1, 2, 4, 8, 16, 24),
  min_dist = c(0.01)
)

# Run 8 jobs in parallel
my_cluster <- parallel::makeCluster(8, type = "FORK")

doParallel::registerDoParallel(cl = my_cluster)
umaps <- foreach(
  spread = umap_params$spread,
  min_dist = umap_params$min_dist,
  .packages = c("uwot", "Matrix")
) %dopar% {
  uwot::umap(
    X         = NULL,
    init      = pca[,1:2],
    nn_method = nn_method,
    spread    = spread,
    min_dist  = min_dist,
    n_threads = 2
  )
}
parallel::stopCluster(cl = my_cluster)

d <- rbindlist(lapply(seq(nrow(umap_params)), function(i) {
  data.table(
    x = umaps[[i]][,1],
    y = umaps[[i]][,2],
    spread = umap_params$spread[i],
    min_dist = umap_params$min_dist[i],
    group = as.character(obs[["cluster"]])
  )
}))

p <- ggplot(d) +
  geom_scattermore(
    mapping = aes(x = x, y = y, color = group),
    pointsize = 1,
    pixels    = c(1000, 1000)
  ) +
  theme(
    axis.text = element_blank(),
    axis.ticks = element_blank(),
    axis.title = element_blank(),
    legend.position = "none"
  ) +
  facet_wrap(~ spread, scales = "free") +
  labs(
    title = glue("{comma(nrow(umaps[[1]]))} cells in UMAP, adjusting spread")
  )

umap-min_dist-montage.pdf

umap-spread-montage.pdf