How to plot a colorbar legend with `geom_scattermost`?

[davemcg]

I'm trying to implement scattermost (seems to improved plotting time from ~2 seconds to ~1 second!), but I can't figure out how to add the scale color bar. Any tips?

[exaexa]

Hello! I assume the picture you shared is geom_scattermost + the scale edited in by hand?

Anyway, geom_scattermost unfortunately cannot influence the guides/legends directly because its main purpose is to avoid ggplot touching the data. You can either use geom_scattermore if that's good enough (that will be slow though), or fake the data column as such:

xy <- cbind(rnorm(1000), rnorm(1000))  # data
val <- asinh(xy[,1]*xy[,2])   # color value (derived from data to look nicer)

ggsave("test.png", width=2, height=1.5, scale=2,   #saves the plot
  ggplot() +  #no real data to plot
  geom_scattermost(xy,    #manually plot and color the points
    col=viridisLite::inferno(100)[1+99*(val-min(val))/(max(val)-min(val))],
    pointsize=4) +
  geom_point(data=data.frame(x=double(0)), aes(x,x,color=x)) +  # add a fake empty point-style aes
  scale_color_gradientn(  #add the manual guide for the empty aes
    limits=c(min(val),max(val)),
    colors=viridisLite::inferno(100),
    name="Some value") +
  ggtitle("Manual guide with geom_scattermost")
)

Results in this:

Would that work?

[davemcg]

Oh that is perfect. I forgot you had to fake the data aes for ggplot to add a legend. I was trying to manually add the legend and was getting miffed.

The example I gave is actually the scattermore implementation, but as this is for a Shiny app speed is VERY important as the plot will get remade every time someone changes the gene.

Code for scattermost (with your improvement):

gene <- 'CRX'
# p is a data frame with UMAP_1 (x coord), UMAP_2 (y coord), and cpm (intensity value for color) that only contains values for points/cells with a cpm > 1
# meta_filter is just UMAP_1, UMAP_2 for ALL data points (plotted in gray)
color_range <- range(p$cpm)
plot <- ggplot() +
  geom_scattermost(cbind(meta_filter$UMAP_1, meta_filter$UMAP_2), color = '#D3D3D333', 
                   pointsize = 2) +
  geom_scattermost(cbind(p$UMAP_1, p$UMAP_2),
                   color = viridis::magma(100, alpha=0.3)
                   [1+99*(p$cpm-color_range[1])/diff(color_range)],
                   pointsize=2.5,
                   pixels=c(1000,1000),
                   interpolate=FALSE) + 
  geom_point(data=data.frame(x=double(0)), aes(x,x,color=x)) +
  scale_color_gradientn(  #add the manual guide for the empty aes
    limits=c(min(p$cpm),max(p$cpm)),
    colors=viridis::magma(100),
    name="log2(cpm+1)") +
  cowplot::theme_cowplot() + 
  theme(axis.line = element_blank(),
        axis.title = element_blank(),
        axis.ticks = element_blank(),
        axis.text = element_blank()) +
  annotate("text", -Inf, Inf, label = paste0(gene, '\nexpression'), hjust = 0, vjust = 1, size = 6)

Scattermore version

p %>% ggplot() + 
  geom_scattermore(data = meta_filter,
                   aes(x = UMAP_1, y = UMAP_2), 
                   pointsize = 2, color = 'gray', alpha = 0.1) +
  geom_scattermore(aes(x = UMAP_1, y = UMAP_2, colour = cpm), 
                   pointsize = 2.5, 
                   alpha = 0.3) +
  scale_color_viridis_c(option = 'magma') +
  cowplot::theme_cowplot() + 
  theme(axis.line = element_blank(),
        axis.title = element_blank(),
        axis.ticks = element_blank(),
        axis.text = element_blank()) +
  annotate("text", -Inf, Inf, label = paste0(gene, '\nexpression'), hjust = 0, vjust = 1, size = 6)

This is plotting ~550,000 points and on a 12'' MacBook also running Zoom scattermore is ~6 seconds and scattermost is ~3 seconds.

[exaexa]

Oh that is perfect.

OK, great. I'll pin the issue b/c there may be more people with the same problem.

but as this is for a Shiny app speed is VERY important

Well, Shiny is the original reason for scattermore... See https://gitlab.com/exaexa/ShinySOM :]

[exaexa]

PS. This https://github.com/exaexa/shinyAccPoints was the first attempt for plotting the data really quick with Shiny, using WebGL. It seemed to be a pretty nice starting point for various animated web output, but the loading performance is actually killed by the insane (non-)speed of R JSONifier implementation.... :]

[jhagberg]

Is there any easy way to add a legend of the colors using scattermost? Not a gradient just like a regular legend like with geom_point

[exaexa]

@jhagberg you mean for a discrete scale? yeah, basically the same way as at https://github.com/exaexa/scattermore/issues/2#issuecomment-616656555 except you likely need to supply an empty data preferably as integer(0) and your precise description of the colors. I can work out an example when I get a bit of time.

[jhagberg]

Yes. We have a Very slow scatter plot using ggplot. with X,y and grouping by "method" name . So every point from the same method should have the same color

[exaexa]

Anyway @jhagberg it was a bit harder to push ggplot into thinking that it really needs to print the guide but it can be glued together roughly like this:

xy <- cbind(rnorm(1000), rnorm(1000))
levs <- 1:4
labs <- c("a","B","c","violet")
clu <- sample(levs, 1000, replace=T)
pal <- function(n){print(n); RColorBrewer::brewer.pal(n, 'Set1');}

ggsave("test2.png", width=2, height=1.5, scale=2,
  ggplot() +
  geom_scattermost(xy,
    col=pal(length(levs))[clu],
    pointsize=4) +
  geom_point(
    data=data.frame(x=as.double(NA), y=factor(as.integer(NA), levels=levs)),
    aes(x,x,color=y)) +
  scale_color_manual(
    breaks=levs,
    values=pal(length(levs)),
    labels=labs,
    drop=F,
    name="Some category") +
  ggtitle("Manual guide with geom_scattermost")
)

I hope some ggplot specialists show up and fix the code a bit less hacky :D

[exaexa]

(side note: interesting that ggplot is okay making the colorbar for empty continuous data, but completely removes the legend for empty discrete data...)