ggplot2: Labeling noteworthy points in scatterplots (geom_noteworthy)

[friendly]

An important feature of graphical methods for datavis is automatic labeling of unusual or "noteworthy" points
in scatterplots. This is handled automatically in base R for plot(model), and much more generally in the
car package, where an id = list(n, method, ...) argument to many functions provides for most use cases
via a function car::showLabels().

I use this extensively with base R graphs in my packages, e.g, heplots::cqplot(), which produces chisquare QQ plots for a multivariate
linear model, labeling the 3 points with largest Mahalanobis $D^2$:

library(heplots)
iris.mod <- lm(as.matrix(iris[,1:4]) ~ Species, data=iris)
cqplot(iris.mod, id.n=3)

A nice feature of car::showLabels() is that it has an argument location for positioning the text label relative to the point to avoid
overplotting these.

Correct me if I'm wrong, but I do not find any similar facility for easily labeling points selectively in statistical
applications of ggplot2. It can be done, but is currently awkward to implement in ggplot2.

For a simple example, I want to plot Crime_pers against Literacy in the Guerry data, and label those
departments which are extreme in terms of Mahalanobis distance, to get this:

library(ggplot2)
data(Guerry, package="Guerry")

ggplot(aes(x=Literacy, y=Crime_pers/1000), data=Guerry) +
  geom_point(size=2) +
  stat_ellipse(level=0.68, color="blue", linewidth=1.2) +  
  stat_ellipse(level=0.95, color="gray", linewidth=1, linetype=2) + 
  geom_smooth(method="lm", formula=y~x, fill="lightblue") +
  geom_smooth(method="loess", formula=y~x, color="red", se=FALSE) +
  theme_bw() + theme(text = element_text(size=18))

gp <- last_plot()

Then, this is what I had to do to add the point labels:
From qchisq(0.9, df=2) = 4.6 I had to create a separate dataset gdf,
and then subset that using gdf[gdf$dsq > 4.6,]

#' ## Use Mahalanobis Dsq to label unusual points
gdf <- Guerry[, c("Literacy", "Crime_pers", "Department")]
gdf$dsq <- heplots::Mahalanobis(gdf[, 1:2])

library(ggrepel)
gp + 
	   geom_label_repel(data = gdf[gdf$dsq > 4.6,], aes(label=Department))

What I'm looking for is something like a geom_noteworthy() function. It would provide similar methods to those in
car::showLabels() to select points to be
identified in a plot, and a geom argument that controls how the identified points are labeled in a plot.

geom_noteworthy(method = c("dsq", "mahal", "x", "y", "r", "ry"),
                n =            # number of points to label,
                geom =         # e.g.,   "text", "label", "repel",
                ...)           # other args passed to the geom

To this end, in the latest release (v 1.7.4) of heplots
I've written a compute function, heplots::noteworthy(x, y, n, method, ...) that tries to be very general,
https://github.com/friendly/heplots/blob/master/R/noteworthy.R.

The method argument can be any of the following:

• an integer vector of row numbers: all are chosen
• a vector of length(x) numbers: the largest n are chosen, e.g., cooks.distance(model) to label influential
• a text string selecting a built-in method: 'x', 'y', 'mahal', 'dsq', 'r', 'ry'

It returns the row IDs of the selected points

ids <- noteworthy(Guerry$Literacy, Guerry$Crime_pers, method = "mahal", n = 7) |> print()
# [1]  7 21 66 86  8 23 37
Guerry[ids, c("Literacy", "Crime_pers", "Department")]
#    Literacy Crime_pers Department
# 7        67      35203   Ardennes
# 21       23      37014     Creuse
# 66       71       7343  Haut-Rhin
# 86       49       2199      Corse
# 8        18       6173     Ariege
# 23       73      11560      Doubs
# 37       73      26221       Jura

If I had a geom_noteworthy(), I could simply do something like:

gp + geom_noteworthy(aes(label=Department), method="mahal", geom="repel", n=7)

Is there anything like this in the ggplot2 world? I'm not sure where I would begin to try to implement this.

[teunbrand]

I suppose it depends on what 'easily labeling points selectively' means. To my knowledge there are few tools out there:

• ggforce::geom_mark_*() functions, which is more for identifying groups than individual points but has a selection aesthetic.
• {ggrepel} does the 'easily labeling points' bit well but 'selectively' is a slight inconvenience.
• {directlabels} is also 'easily places labels' but is also more group oriented.

With regards to a ggplot2 friendly interface to heplots::noteworthy(), I think having made geom_noteworthy(geom) an argument to flexibly swap out label geoms is indicative that you might want to implement a stat instead. Do you imagine selecting the points work for the entire plot, per panel or per group?

A few loose remarks:

• A metric I entertained a bit is the inverse of 2D KDE or the inverse of the number of nearest neighbours (within some threshold distance), so that the points that are in the least dense parts of the plot or have the fewest neighbours would be selected.
• I always detested having to pick points to label. I used to do plenty of volcano plots, such as the one below, where you want to usually pick the top n points (either in effect size or significance) to label at each side.

[friendly]

Do you imagine selecting the points work for the entire plot, per panel or per group?

That's a good point. I hadn't thought about faceting or groups, but I can see that this would be useful in a general approach. What would have to change in your sketch below for this?

[teunbrand]

You'd swap out the compute_group() method for compute_panel() or compute_layer()

[teunbrand]

I'm not sure where I would begin to try to implement this.

Here is a quick sketch:

library(ggplot2)
library(ggrepel)

# Class
StatNoteworthy <- ggproto(
  "StatNoteworthy", Stat,
  
  required_aes = c("x", "y"),
  
  compute_group = function(data, scales, method = "dsq", n = 10) {
    idx <- heplots::noteworthy(x = data$x, y = data$y, n = n, method = method)
    data[idx, , drop = FALSE]
  }
)

# Constructor: standard stuff + method + n
stat_noteworthy <- function(
    mapping = NULL, data = NULL, geom = "text", position = "identity",
    ..., method = "dsq", n = 10,
    na.rm = FALSE, show.legend = NA, inherit.aes = TRUE
) {
  layer(
    data = data, mapping = mapping, stat = StatNoteworthy, geom = geom,
    position = position, show.legend = show.legend, inherit.aes = inherit.aes,
    params = rlang::list2(
      method = method,
      n = n,
      ...
    )
  )
}

# Example
ggplot(mpg, aes(displ, hwy)) +
  geom_point() +
  stat_noteworthy(aes(label = model), geom = "text_repel")

^{Created on 2025-04-16 with reprex v2.1.1}

[EvaMaeRey]

if 'text_repel' is the anticipated use and you willing to take that dependency, maybe just make that 'text_repel' as the default too.

# Constructor: standard stuff + method + n
stat_noteworthy <- function(
    mapping = NULL, data = NULL, geom = "text_repel", position = "identity",
    ..., method = "dsq", n = 10,
    na.rm = FALSE, show.legend = NA, inherit.aes = TRUE
) {
  layer(
    data = data, mapping = mapping, stat = StatNoteworthy, geom = geom,
    position = position, show.legend = show.legend, inherit.aes = inherit.aes,
    params = rlang::list2(
      method = method,
      n = n,
      ...
    )
  )
}

ggplot(mpg) +
  aes(displ, hwy, label = model) +
  geom_point() +
  stat_noteworthy()

Iif a geom_ prefix feels friendlier, (I think it does for many of people), you could follow the springs case study 'constructors' https://ggplot2-book.org/ext-springs.html#constructors and just switch what's fixed (Geom is fixed):

# Constructor: standard stuff + method + n
geom_noteworthy <- function(
    mapping = NULL, data = NULL, stat = "noteworthy", position = "identity",
    ..., method = "dsq", n = 10,
    na.rm = FALSE, show.legend = NA, inherit.aes = TRUE
) {
  layer(
    data = data, mapping = mapping, stat = stat, geom = GeomTextRepel,
    position = position, show.legend = show.legend, inherit.aes = inherit.aes,
    params = rlang::list2(
      method = method,
      n = n,
      ...
    )
  )
}

This feels a little weird, because you can remove the noteworthy-ness of the layer (the stat).

So another approach is to do verbatim alias - essentially the approach of ggbump::geom_bump(); and what I recommend in easy-geom-recipes.. Using stat_ or geom_ if you're not going to change any defaults should feel identitcal, so name doesn't really matter.

geom_notable <- stat_notable

This has it's own weirdness to it, but it's a concise weirdness - you can go back and do verbose construction later?

[EvaMaeRey]

I like the alternative universe in which we just expose all the main characters (stat, geom, and position) in the constructor function...

[friendly]

Thanks so much @teunbrand @EvaMaeRey !!

• I've implemented this as a stat, stat_noteworthy(), with geom = "text" and n=5 as the defaults for now.
• Positioning the text labels wrt the points is important in this application, and geom_text(), ..., geom_label_repel() provide many optional arguments for this purpose. Do I have to allow explicitly for these in stat_noteworthy() or can I assume they would somehow be inherited via ...?
• Also added the level argument, which, in the case of Mahalanobis $D^2$ allows to filter selected cases according to a p-value.
• Tested this for nearly all use-cases.
• Still no idea how to incorporate groups or facets, or situations where there is a color|fill|shape aesthetic defining groups. The usual ggplot-thinking approach would be to identify n unusual points within each group, but a stats/datavis approach would want to see just those points identified in a global plot.

My current code is at https://github.com/friendly/heplots/blob/master/dev/stat_noteworthy.R
(Should probably be on a dev branch, not a folder.)

Here are the cases I've tested, all with expected results:

# Examples
gp <- ggplot(mpg, aes(x = displ, y = hwy, label = model)) +
  geom_point() 

# default: method = "mahal"
gp + stat_noteworthy()

# specify level to filter by qchisq() > level 
gp + stat_noteworthy(n = 10, level = 0.99)

# use ggrepel
gp + stat_noteworthy(geom = "text_repel")
gp + stat_noteworthy(geom = "label_repel")

gp + stat_noteworthy(method = "x")
gp + stat_noteworthy(method = "y")
gp + stat_noteworthy(method = "ry")

# select cases using row IDs
gp + stat_noteworthy(method = seq(10, 100, 20))

# select cases using a computed vector
mod <- lm(hwy ~ displ, data=mpg)
gp + stat_noteworthy(method = cooks.distance(mod))
gp + stat_noteworthy(method = hatvalues(mod))

[teunbrand]

Positioning the text labels wrt the points is important in this application, and geom_text(), ..., geom_label_repel() provide many optional arguments for this purpose. Do I have to allow explicitly for these in stat_noteworthy() or can I assume they would somehow be inherited via ...?

You don't need to do any extra infrastructure. Everything that is passed via params = list2(...) is automatically distributed to the stat and geom of a layer.

situations where there is a color|fill|shape aesthetic defining groups

This happens automatically when other aesthetics are discrete. This is neither under control of the Stat or Geom but under Layer.

[friendly]

I may be asking for way too much, but here is the example that prompted me to raise this idea.

For a plot of the Penguins data, plotting bill_length against bill_depth, I wanted to label the points with largest $D^2$ in the full dataset, to get this plot (to illustrate that multivariate outliers may not appear unusual in 2D views)

First, calculate the $D^2$ values and augment the data set

data(peng, package = "heplots")
# use all numeric variables
DSQ <- heplots::Mahalanobis(peng[, 3:6])
ids <- order(DSQ, decreasing = TRUE)[1:3] |> print()
#> [1] 283  10  35

# augment the dataset to add `note` variable for selecting labels
peng_plot <- peng |>
  tibble::rownames_to_column(var = "id") |> 
  mutate(note = id %in% ids)

Then, my ggplot2 step used geom_text() on the subset:

ggplot(peng_plot, 
       aes(x = bill_length, y = bill_depth,
           color = species, shape = species, fill=species)) +
  geom_point(aes(size=note), show.legend = FALSE) +
  scale_size_manual(values = c(1.5, 4)) +
  geom_text(data = subset(peng_plot, note==TRUE),
            aes(label = id),
            nudge_y = .4, color = "black", size = 5) +
  geom_smooth(method = "lm", formula = y ~ x,
              se=FALSE, linewidth=2) +
  stat_ellipse(geom = "polygon", level = 0.95, alpha = 0.1) +
  # theme_penguins() +
  theme_bw(base_size = 14) +
  theme(legend.position = "inside",
        legend.position.inside=c(0.85, 0.15))

Here is my attempt at a simplified version using stat_noteworthy(), leaving out changing the point size for the labeled points,
(and also testing what happens with color|fill|shape aesthetics)

ggplot(peng, 
       aes(x = bill_length, y = bill_depth,
           color = species, shape = species, fill=species)) +
  geom_point() +
  geom_smooth(method = "lm", formula = y ~ x,
              se=FALSE, linewidth=2) +
  stat_ellipse(geom = "polygon", level = 0.95, alpha = 0.1) +
  stat_noteworthy(method = ids, label = ids)

But this generates an error:

Error in `stat_noteworthy()`:
! Problem while setting up geom aesthetics.
ℹ Error occurred in the 4th layer.
Caused by error in `check_aesthetics()`:
! Aesthetics must be either length 1 or the same as the data (9).
✖ Fix the following mappings: `label`.
Run `rlang::last_trace()` to see where the error occurred.

[EvaMaeRey]

you could define StatNoteworthy panel-wise and compute a row number along the way (id), and have the default_aes for label be the row number.

StatNoteworthy <- ggproto(
  "StatNoteworthy", Stat,
  
  required_aes = c("x", "y"),
  
  compute_panel = function(data, scales, method = "dsq", n = 10) {
    
    data <- data |>
      mutate(id = row_number())
    
    id_rows <- heplots::noteworthy(x = data$x, y = data$y, n = n, method = method)
    
    data[id_rows, , drop = FALSE]
    
  },
  default_aes = aes(label = after_stat(id))   # << 
)

peng |> 
  select(x = bill_length, y = bill_depth) |>
  StatNoteworthy$compute_panel()


ggplot(peng, 
       aes(x = bill_length, 
           y = bill_depth,
           color = species, 
           shape = species, 
           fill = species)) +
  geom_point() +
  geom_smooth(method = "lm", 
              formula = y ~ x,
              se = FALSE, 
              linewidth = 2) +
  stat_ellipse(geom = "polygon", level = 0.95, alpha = 0.1) +
  geom_text(stat = StatNoteworthy, method = ids, n = 3, color = "black")

But I'm not quite sure I understand the intent. It seems like you want Mahalanobis computed in multiple dimensions. But does the noteworthy function work in 2D? Is the Stat doing any work related to noteworthiness in this example? I think it is subsetting the data based on IDs pulled out prior to plotting. Is that right? Sorry for my confusion!

[teunbrand]

Does using aes(label = ids) solve the issue? Assuming that ids is a vector parallel to the rows of your data, if you pass this outside aes(), it gets added after the stat is computed and you have less rows than in the full data.

[friendly]

@EvaMaeRey Thanks for this suggestion. In nearly all use cases, noteworthy points are determined by (x,y) within the plot, This example is unusual (and possibly confusing) in that I want to identify noteworthy points in the full space of the numeric penguin variables.

[EvaMaeRey]

@friendly, thanks for the explanation! I don't think it would be that strange for others to want to do the same (look at noteworthy points in a 2-D projection). I think there are some interesting plotting/ggplot2 extension problems if you think about a Stat to do the compute internally -- something for when some free time happens to find us. 🙃😭 I think the main extension issue might be bringing a bunch of vars from the data verbatim to the plotting space, like a aes_as_is(). @yjunechoe showed me another approach #18 (reply in thread), (vars_pack) but I've wondered about something that mirrors aes(sex = sex, species = species) where the syntax would be aes_as_is(sex, species) to scoot the variables over to the plotting environment without having to repeat yourself. For now, I think @teunbrand is right that if you used peng_plot for example (which has the 'id' var), you could do with the Stat as you'd specified it earlier:

ggplot(peng_plot, 
       aes(x = bill_length, 
           y = bill_depth,
           color = species, 
           shape = species, 
           fill = species,
           label = id)) +                        #<<
  geom_point() +
  geom_smooth(method = "lm", 
              formula = y ~ x,
              se = FALSE, 
              linewidth = 2) +
  stat_ellipse(geom = "polygon", level = 0.95, alpha = 0.1) +
  geom_text(stat = StatNoteworthy, method = ids, color = "black")

[aphalo]

@friendly Something to keep in mind when using geom_text_repel() and geom_label_repel(): the data they receive must include all observations, with those not to be labeled having "" as the string mapped to label (NA does not work as ggplot removes these rows from the data seen by geoms). This is necessary to ensure that text does not overlap unlabelled observations. For the same reason all text or labels to be repulsed must be added to the plot in the same layer.

Package 'ggpp' has some tools for selective labelling, but not specific ones for statistical plots as you envision for geom_noteworthy().

The equivalent of location would be in 'ggplot2' a position function. If a segment or arrow is to be drawn connecting the observations and labels then positions and geoms from packages 'ggrepel' or 'ggpp' have to be used. Package 'ggpp' defines several position functions aiming at automating the positioning of labels, some of them computing the direction of the displacement based on the position of observations/shape of the cloud of observations. Also arbitrary focal points or lines to nudge away/towards are supported. Positions and stats from 'ggpp' are designed to work well with 'ggrepel'. Manual nudging of individual labels is possible with positions from 'ggpp' as they accept vectors as arguments for x and y parameters. (I do not know/remenber if position functions from 'ggplot2' and other extensions accept vectors.)

In scatter plots, volcano plots and quadrant plots stat_dens2d_labels() and stat_dens2d_filter() from 'ggpp' can be used to label only observations in regions with low local 2D density of observations. Matching functions for 1D density are also provided. Constraints can be passed as arguments and also manual (vector- and function-based) selection of the observations to be labeled is possible, and can be combined with the density approach in a single plot layer.

The documentation of 'ggpp' includes several examples in the help pages and in the User Guide vignette. Sections on 2D density filtering of labels, and Simple position functions are the most relevant. It also includes an article on Combining repulsion and nudging, i.e., using 'ggpp' and 'ggrepel' together. The page Volcano and quadrant plots with ‘ggpmisc’ at my website has several examples also relevant in other contexts.

Although 'ggpp' does not directly implement what geom_noteworthy() aims at, there maybe be some useful ideas in 'ggpp' as the features implemented in it have been evolving over the course of several years of use and development, including collaboration with Kamil Slowikoski, the developer of 'ggrepel' to get 'ggpp' and 'ggrepel' working together smoothly.

[aphalo]

Here I may be repeating in part what has been discussed earlier in this thread. Hopefully, not too much.

gp + geom_noteworthy(aes(label=Department), method="mahal", geom="repel", n=7)

geoms do not have a geom parameter, so this would be stat_noteworthy(). Anyway my own view is that it is preferable to define geoms with stat_identity() as default and define statistics for anything that modifies data with a suitable geom as default argument.

If used with repulsive geoms, as I mentioned above, selecting the rows in data would not work always as desired. Instead of

data = gdf[gdf$dsq > 4.6,]

something like data$label <- ifelse(gdf$dsq > 4.6, data$label, "") would be the approach to use with repulsive geoms. However, this approach would be wasteful with other geoms. Selecting rows should work with non-repulsive geoms. In 'ggpp' I implemented these two approaches, each in a separate stat, but they could be combined. I think that even switching the default approach based on the name of the geom passed as argument to the stat should be doable.

In 'ggpp' stat_dens2d_labels() does replacement of labels, and stat_dens2d_filter() selects rows. The sources are at GitHub: stat-dens2d-labels.r and stat-dens2d-filter.r. The criterion used to filter/edit labels is different to what it will be in stat_noteworthy() and some of the additional tricks are most likely not needed. @friendly @teunbrand @EvaMaeRey and others, comments/criticism on my approach is warmly welcome.

Repulsive geoms need to have access to the location of observations to be able to avoid overlaps. This means that only position functions that preserve the original coordinates when displacing the initial position of labels work well with repulsion, and obviously also allow drawing connecting segments. As far as I know, only position functions defined in packages 'ggrepel' and 'ggpp' work well with repulsion. Package 'ggpp' defines in addition to many new types of positions, keep versions of most (all?) positions defined in 'ggplot2'.

(I will be travelling during the next few days, and I may be even slower than usual with replying...)

[friendly]

This topic has been dormant for quite a while, but I'd like to pick it up again and achieve closure-- hey, ship a feature!

I'm currently going over my previous attempts with a view to specifying the behavior I want for a general, easily used method for identifying "unusual" points in ggplot displays. I'll post a follow-up when this is ready.

[EvaMaeRey]

I think you are after something else, but might be worth checking out this little experiment, that uses highest cooks as qualification for notewortiness. https://evamaerey.github.io/mytidytuesday/2026-09-23-noteable-gg-friendly/noteable-gg-friendly.html

I think even if you are using another notable-ness measure, you might think about including a drop argument and an add argument for forcing a label to be dropped or added. For example we might want to drop Zimbabwe even though it has a high cooks distance (since it's in a bit of a crowd with other high cooks observations), and we might just want Argentina because that is a country of interest to us.

Compute borrows from ggxmean::geom_point_high_cooks, which was written by my student in independent study in Fall 2021 🙃. Lessons learned teaching that turned into 'easy geom recipes' V1, also drafted together with an independent study student in Spring 2022. Translated to plotnine in 2026! https://github.com/has2k1/easy-geom-recipes 🚀

[teunbrand]

I've also always liked the idea to compute a 2D density of some sort and labelling the least dense points.
Here with delayed aesthetics:

library(ggplot2)
library(ggpointdensity)
library(ggrepel)

gapminder::gapminder |>
  dplyr::filter(year == 2002) |>
  ggplot(aes(gdpPercap, lifeExp)) +
  geom_point() +
  stat_pointdensity(
    geom = "text_repel",
    method = "kde2d",
    aes(label = stage(
      as.character(country), 
      after_stat = ifelse(scales::rescale(density) < 0.1, label, "")
    )),
    colour = 'black'
  ) +
  scale_x_log10()

^{Created on 2026-09-24 with reprex v2.1.1}

You can also move the 'hide the high-density labels' mechanism to the scales if you like that better.

gapminder::gapminder |>
  dplyr::filter(year == 2002) |>
  ggplot(aes(gdpPercap, lifeExp, label = country)) +
  geom_point() +
  stat_pointdensity(
    geom = "text_repel",
    method = "kde2d",
    aes(colour = after_stat(density))
  ) +
  continuous_scale(
    "colour",
    palette = function(x) ifelse(x > 0.1, NA, "black"),
    guide = "none"
  ) +
  scale_x_log10()

[EvaMaeRey]

hello stat_pointdensity() 👋🙌 I didn't know it - about to check it out. I'd thought a little about stat ellipse for this, but didn't get far at all - since it doesn't return point-wise values (I don't believe). stat_pointdensity seems like a great direction ... more general than cooks.

I was aware of ggpointdensity::geom_pointdensity() - really love the visuals it helps produce!

[EvaMaeRey]

lol... I missed the ggpointdensity import. And geom_pointdensity and stat_pointdensity are basically alliases.

[teunbrand]

Yes, no exotic voodoo witchcraft here I'm afraid. Just garden variety voodoo witchcraft :D

[aphalo]

@EvaMaeRey In 'ggpp' there is an implementation based based on local density. As Teun's example shows using ifelse(), in many cases it is easy enough to achieve the selection based on density without defining new stats. While reading this thread, I realised that the stats from 'ggpp' could be modified to take the function used to do the selection as an argument instead of having the computation of density hard-wired. A couple of things the 'ggpp' stats do is provide different ways of specifying which labels to keep, including your wish of adding some manually. Anyway, these stats are mostly a convenience. In contrast, the positions that I wrote to go with them, turned out to be more useful than what I had expected because of the synergy with 'ggrepel'. I mentioned them here in April last year, when the thread seems to have been dormant. Here are some additional details.

• stat_dens1d_filter() and stat_dens2d_filter() subset data and work well with geom_text() and geom_label(). These two stats do the computations by panel. stat_dens1d_filter_g() and stat_dens2d_filter_g() do the computations by group.
• stat_dens1d_labels() and stat_dens2d_labels() instead of subsetting data replace the labels to be "hidden" by "". They are meant to be used with geom_text_repel() and geom_label_repel(), because 'ggrepel' recognizes labels set to "" as points to repulse away from, thus ensuring that labels do not overlap unlabelled observations.
• 'ggpp' also defines special nudge position functions, which compute the nudge, for example, radially away from the centroid or away from a fitted line. A small nudge usually makes repulsion start in a favourable direction.

Yes, I have also found that sometimes manually setting which observations are be label is useful, so this is also supported.

I have quite a few examples of combining nudging and repulsion in in an article in the on-line documentation. In the vignette there are some examples not using 'ggrepel' to avoid the dependency. In my web site I have additional examples, from which I copy here one plot.

This example uses 'ggplot2', 'ggrepel', 'ggpp' and 'ggpmisc'.

ggplot(volcano_example.df, 
       aes(logFC, 
           PValue, colour = outcome2factor(outcome),
           label = gene)) +
  geom_point() +
  stat_dens2d_labels(geom = "text_repel", 
                     size = 2,
                     colour = "black",
                     position = position_nudge_centre(x = 0.1, 
                                                      y = 0.1, 
                                                      direction = "radial"),
                     keep.fraction = 0.03, # 3% of data points
                     min.segment.length = 0) +
  scale_x_logFC(name = "Transcript abundance%unit") +
  scale_y_Pvalue() +
  scale_colour_outcome() +
  stat_quadrant_counts(data = function(x) {subset(x, outcome != 0)},
                       aes(label = after_stat(count.label))) +
  theme_bw()