Alternative backend for scatter plot rasterization: {ggrastr} instead of {scattermore}

[pormr]

Seurat now use geom_scattermore() by default to render scatter plots with more than 100,000 cells under the hood (this can be overriden using raster = FALSE). However, the results can be suboptimal when prioritizing subpopulations with high gene expression on UMAP/t-SNE plots using order = TRUE. In these cases, positive cells blend with negative ones, resulting in a grayish and blurry appearance. This is partially due to {scattermore}'s specialized blending algorithm (this has been discussed in exaexa/scattermore#25), which treat all overlapping points the same for blending colors. This behavior is currently hardcoded, and for some reasons the package author has not implemented alternative blending options in geom_scattermore().

To address this, extra efforts are warranted to bend its low-level API, and compatibility with faceting functions would be a great problem. As an alternative, {ggrastr} can also rasterize specific layers while preserving others as editable vectors. Meanwhile, it can produces clearer, publication-ready results (at the cost of increased time and memory usage). Notably, ggrastr::geom_point_rast() serves as a drop-in replacement for geom_point() without hassle, when compared with geom_scattermore()/geom_scattermost().

However, {ggrastr} is much slower than {scattermore} and introduces dependencies like Cairo and ragg. A practical solution would be to offer a backend option: ggrastr for quality and scattermore for performance. While specialized packages like {scpubr} may be a better place to address this issue, implementing this flexibility could be a valuable addition to this package.

-- Update --
Seems that Seurat has already using {ggrastr} for rasterizing jitter points in VlnPlot() (as implemented in SingleExiPlot()). It would be nice we could do the same for scatter plots.

Codes for benchmarking:

Details

library(ggplot2)
library(ggrastr)
library(scattermore)

set.seed(42)
n <- 1e6  # 1 million points
df <- data.frame(x = rnorm(n), y = rnorm(n))

plot_ggrastr <- ggplot(df, aes(x, y)) +
  geom_point_rast(size = 0.5, alpha = 0.5)

plot_scattermore <- ggplot(df, aes(x, y)) +
  geom_scattermore()

• ggrastr::geom_point_rast() (about 22s):

  benchplot(plot_ggrastr)

  step user.self sys.self elapsed
  1 construct 0.000 0.000 0.000
  2 build 0.406 0.172 0.580
  3 render 0.222 0.012 0.237
  4 draw 21.410 0.015 21.458
  5 TOTAL 22.038 0.199 22.275

• scattermore::geom_scattermore() (less than 1s):

  benchplot(plot_scattermore):

  step user.self sys.self elapsed
  1 construct 0.000 0.000 0.000
  2 build 0.405 0.062 0.467
  3 render 0.311 0.017 0.328
  4 draw 0.028 0.001 0.029
  5 TOTAL 0.744 0.080 0.824

  Setting pointsize to values other than 0 would slightly increase total time.

[pormr]

Possibly related: #8225