ggtext and parsing html with latest version of ggplot

[Philip-Leftwich]

I found a problem with ggplot and rendering html text with ggtext or marquee

I expected to parse character string in html format and have the style implemented.
This works for my code in ggplot 3.5.2 but not in ggplot 4.0.0

Here is the code to reproduce the bug:

library(ggplot2)
library(ggtext)

# Minimal dataframe
df <- data.frame(
  gRNA_type = rep(c("cd<sup><i>g338-384</i></sup>", 
                    "wt<sup><i>control</i></sup>"), each = 5),
  win = c(95, 88, 92, 90, 94, 45, 50, 48, 52, 47),
  loss = c(5, 12, 8, 10, 6, 55, 50, 52, 48, 53),
  id = 1:10
)

# Test plot
ggplot(df, aes(x = gRNA_type, y = win)) +
  geom_point() +
  theme_minimal() +
  theme(axis.text.x = element_markdown())

^{Created on 2025-11-19 with reprex v2.1.1}

[teunbrand]

Thanks, I can confirm this on my end. The issue is element inheritance between S3 and S7 elements. element_markdown() isn't propagated to axis.text.x.bottom because theme_minimal()$axis.text.x.bottom already exists. The issue can be fixed by using theme(axis.text.x.bottom = element_markdown()).

[teunbrand]

I'm closing this issue as I'm assuming the thumbs-up means it was solved on your end. The path forward is for extensions to implement S7-based element extensions, which isn't really something ggplot2 can do for them.

[clauswilke]

@teunbrand I'm not sure I've followed all the recent changes. Is there something ggtext should do that it hasn't done yet?

[teunbrand]

The backward compatibility mechanisms in place mostly preserve how S3 elements work. The only thing it doesn't do is cross-inheritance between S3 and S7 theme elements. In a perfect world, extensions like ggtext would update to S7 elements. However, this would come at the cost of incrementing the ggplot2 version requirement. I can prepare a PR for ggtext and see where we want to go.

[clauswilke]

Sounds good, thanks. It looks to me like a sufficiently big regression that paying the price of increasing the minimum ggplot2 version is probably warranted. People can always install an earlier version of ggtext if they need to.

[teunbrand]

I found out that updating elements to S7 will not be sufficient.
This is how it should work: a child inherits from parent correctly.

library(ggplot2)

element_foo <- S7::new_class("element_foo", parent = element_text)

# Works
t <- theme_gray() +
  theme(axis.text.x = element_foo())

e <- calc_element("axis.text.x.bottom", t)
S7::S7_inherits(e, element_foo)
#> [1] TRUE

In the example below the axis.text.x.bottom is element_text(), so it will override the element_foo(). You may argue that the 'richer' class should take over the 'slender' class, but that is not how it works right now.

# Does not  work
t <- theme_minimal() +
  theme(axis.text.x = element_foo())

e <- calc_element("axis.text.x.bottom", t)
S7::S7_inherits(e, element_foo)
#> [1] FALSE

^{Created on 2025-11-26 with reprex v2.1.1}

The real change that broke this was in #6071 that bestowed theme_minimal() with a non-NULL axis.text.x.bottom element.
In any case, I'm reopening this issue because we could do with a little bit of thinking of what should happen here.

[teunbrand]

I don't think there is a nice solution for this.
If we propagate rich classes like element_markdown() down the theme hierarchy, we'd be ignoring user input when a user states:

theme(
  axis.text = element_markdown(),
  axis.text.x = element_text()
)

This is essentially the same as in this issue, except that the 'user' is theme_minimal().

In theory you can make the case that 'we can propagate if axis.text.x has been populated by a default theme, but not when hand-rolled by a user'. However, we'd need to create a new property to keep track of this, which is a pain in itself, and the semantics are still murky at best: if a plot-creating function includes a custom theme, does that count as user or default?

The unintuive nature of some theme inheritance is something we already accepted and made peace with. This issue is an extension of that nature, so I think this can be laid to rest.

[teunbrand]

I apologise for flip-flopping on this issue, but once again I've changed my mind.
The reason I'm changing this time is because #5486 specifically had such a case in mind, and that mechanism no longer works correctly due to S7/S3 class comparison issues.