Blog

Multi-Panel Figures in R

Background

This post will detail how to make some nice, (nearly) publication-ready, multi-panel figures in R. I want to stress that this isn’t a post intended for R beginners. I will be mixing some standard R code with tidyverse syntax, using the latter’s almighty pipe operator (%>%) and some left-to-right assignment (->).

As the title suggests, this post will be about that most dreaded part of preparing a manuscript for publication – making multi-panel figures. These are often an intimidating part of writing any article; if you’re too much of a perfectionist like me, it can be hard enough to make a single graph that looks just right, let alone making multiple graphs and fitting them all together.

When I first started using R, I was at a loss when it came time to put multiple graphs together. At first I would just create them individually, pasting them together afterwards in image editing software. Unfortunately this technique often leads to things that don’t quite look right – panels that don’t exactly line up, lines/points/labels that are different sizes, etc.

Eventually I learned about packages such as lattice and ggplot, which allow you to take the data stored within a dataframe, split it up into pieces based upon a factor variable, and then plot each piece. So, here’s an example using lattice and some car engine data from the famous Motor Trend dataset:

library(lattice)
data(mtcars)
xyplot(hp ~ disp | factor(cyl), pch = 16, data = mtcars)

The xyplot() function is plotting engine displacement on the x-axis against horsepower on the y-axis, and splitting the data up by number of cylinders. The result:

Here’s how to make approximately the same thing using ggplot, and the output:

library(ggplot2)
data(mtcars)
ggplot(mtcars, aes(x = disp, y = hp)) +
  geom_point(color = "blue", fill = "blue") +
  facet_grid(~ cyl) +
  theme_bw()

Not too bad – the ggplot code is a little more involved, but offers nearly endless options for customization. I want to emphasize nearly, because there are certain things that ggplot and lattice can do with multiple panels, and certain things they can’t. If you’re ever using ggplot and you think to yourself “this graph would be great if I could just change the scale on only the left pane” you’ll quickly learn that it’s relatively difficult to do. This is by design – “customized” multi-pane figures are one way in which people commonly try to cook up some significant looking findings.

In the example below, I’ll show you how to mix and match different plots together. I want to emphasize that this is often a bad idea, but just like everything in life, there are exceptions. Ultimately it is up to you to decide when it is appropriate to use this technique. However, even if you are working with graphs that aren’t drastically different, this can be a good method to let you format and customize each one individually to get everything looking good. I’ll first go through the example, and at the end give my justification for why I think the practice is okay in this situation.

For the example, I will create a dataset consisting of 5 sunflower genotypes tested across 4 environments (props if you can figure out which state I pulled these town names from). I will simulate two traits for all genotype/environment combinations – plant height, and seed oil content. Then I’ll make genotype-by-environment interaction plots for both traits. The code will use ggplot in conjunction with the cowplots package to arrange plots into a nice, professional-looking grid.

Data Simulation

The first step is to create our simulated data. I’ll define a function, dat_sim(), and then use it to create a dataset “gxe”, consisting of the environment name, genotype name, and then the oil and height data. The dat_sim() function will create random data defined by both the standard deviation between environments, and the standard deviation between genotypes within environments. So, I’ll give the oil content trait some messy genotype-by-environment interaction – the trait variation will be much greater across environments than across genotypes. Conversely, I’ll give the height trait highly stable interaction, with trait variation across environments much less than the variation across genotypes:

library(tidyverse)
library(cowplot)
set.seed(99)

dat_sim <- function(n_genos, n_envs, grand_mean, env_sd, geno_sd) {
  vec <- abs(rnorm(n_envs, mean = grand_mean, sd = env_sd))
  out_list <- list()
  for (i in 1:length(vec)) {
    out_list[[i]] <- rnorm(n_genos, vec[i], sd = geno_sd)
  }
  return(do.call(c, out_list))
}

gxe <- data.frame("env" = sort(rep(c("Alpena", "Mellette", "Kranzburg", "Centerville"), 5)),
                  "geno" = rep(c("Zebulon", "Chocolate", "Sunbright", "Firecracker", "Greenburst"), 4),
                  "oil" = dat_sim(5, 4, 0.25, 0.08, 0.03),
                  "height" = dat_sim(5, 4, 180, 30, 10))

One thing you’ll notice is that the trait units here are completely different. The oil content is expressed as a fraction, and the height data is expressed in centimeters.

The Plotting Function

Now I’m going to create a quick-and-dirty plotting function that uses ggplot() internally. This function, plot_gen(), will have input arguments consisting of the gxe dataframe (df), the trait name (trait), and a label for the trait (ylabel). It will then perform a few steps:

1. Find the mean trait value of all genotypes within each environment
2. Sort the environments from the lowest mean to the highest mean
3. Sort the dataframe by environment to match this ordering
4. Plot a regression for each genotype across the environments

Keep in mind that we will be running this function separately for each trait.

plot_gen <- function(df, trait, ylabel) {
  
  ## Find the order of environment means
  group_by(df, env) %>%
    summarise_at(.vars = trait, .funs = mean) %>%
    arrange_(trait) %>%
    pull(env) %>%
    as.character() ->
    env_order

  ## This assigns the env variable levels in the order of
  ## their mean values before sorting
  gxe$env <- factor(gxe$env, levels = env_order)
  gxe <- arrange(gxe, desc(env), geno)

  ## Plot regressions
  ggplot(gxe, aes_string(x = "env", y = trait)) +
    geom_smooth(aes(group = geno, color = geno), size = 1, method = "lm", se = FALSE) +
    xlab("Environment") +
    ylab(ylabel) +
    theme_bw() ->
    plot_out
  
  return(plot_out)
}

IMPORTANT: Notice that when I sort the dataframe in line 14 above, I am first sorting by environment, and then by genotype. In this case, this is a critical step. The genotype column in our dataframe must be sorted the same way for each plot that is created, because ggplot will read the genotype names in the order that it encounters them. If the ordering of this column doesn’t stay consistent, the coloring of the genotypes will change between plots. It doesn’t matter what order the environments are in, as long as the genotypes are always listed in the same order within each environment.

The method that I’m using – sorting environments by their mean values, and then performing a regression for each genotype – is similar to a method for more formally assessing genotype-by-environment interaction, first described by Finlay and Wilkinson (1963).

The advantage to doing our plotting inside a function is that it makes it easy to reproduce the same plot for different traits. So, to make the plot for oil content (our low-heritability trait), we can just write:

plot_gen(gxe, "oil", "Oil Content")

Which gives us:

Nice. Most of the genotypes are behaving alike across environments. There’s just that one troublemaker -the genotype ‘Greenburst’ (which conveniently got colored green), which is exhibiting strong crossover interaction with all the other genotypes.

Assembling the Plot

Making our oil content plot is all well and good, but now we can get fancier by plotting the two traits together. This is where the cowplots package comes into play. It contains a function, plot_grid(), which allows us to easily smoosh together plots by supplying them as arguments. You can also specify the layout, and give each plot labels. So, to plot our two traits together, it’s as simple as:

plot_grid(plot_gen(gxe, "oil", "Oil Content"),
          plot_gen(gxe, "height", "Plant Height (cm)"),
          ncol = 2,
          labels = c("A", "B"))

The output:

So… it’s not great. The big problem is that the legend appears twice. This is where it comes in handy to have all our plotting wrapped up in our plot_gen() function. First, to show an example of how we could tweak the plots themselves, I’ll rotate the x-axis labels 45 degrees, just for the heck of it. This just requires adding an additional line to our call to ggplot() – line 22 in the box below, so that plot_gen() becomes:

plot_gen <- function(df, trait, ylabel) {
  
  ## Find the order of environment means
  group_by(df, env) %>%
    summarise_at(.vars = trait, .funs = mean) %>%
    arrange_(trait) %>%
    pull(env) %>%
    as.character() ->
    env_order

  ## This assigns the env variables levels in the order of
  ## their mean values before sorting
  gxe$env <- factor(gxe$env, levels = env_order)
  gxe <- arrange(gxe, desc(env), geno)

  ## Plot regressions
  ggplot(gxe, aes_string(x = "env", y = trait)) +
    geom_smooth(aes(group = geno, color = geno), size = 1, method = "lm", se = FALSE) +
    xlab("Environment") +
    ylab(ylabel) +
    theme_bw() +
    theme(axis.text.x = element_text(angle = 45, hjust = 1)) ->
    plot_out
  
  return(plot_out)
}

NOTE that our use of theme() in line 22 above comes after our use of theme_bw() on the line above it. If theme_bw() comes after, it will override any custom theme options that we have set – you can tell I know this by hard-won experience.

Now to take care of the real problem – the multiple legends. The cowplot package has another really handy function – get_legend(). As the name implies, this function pulls the legend out of a figure generated by ggplot. So, in the code block below, we will:

1. Generate a dummy figure (just one of the plots that will be in the final figure)
2. Pull the legend out of it using get_legend()
3. Change the title of the legend from “geno” to “Genotype”
4. Assemble the final multi-panel figure using our two traits and our extracted legend
5. Label the two panels “A” and “B”

Note that we can also assign the portion of the overall plotting space that is taken up by each panel. So below, I am dividing the space up into fifths, and assigning each plot to 2/5 of the overall space, and the legend to 1/5.

dummy_plot <- plot_gen(gxe, "oil", "Oil Content") +
  scale_color_discrete(name = "Genotype")
legend <- get_legend(dummy_plot)

plot_grid(plot_gen(gxe, "oil", "Oil Content") + theme(legend.position = "none"),
          plot_gen(gxe, "height", "Plant Height (cm)") + theme(legend.position = "none"),
          legend,
          ncol = 3,
          rel_widths = c(2, 2, 1),
          labels = c("A", "B", ""))

Our final multi-panel figure:

Discussion

That gets us pretty close to what we want. You’ll notice that not only are the units different for each plot, but the ordering of environments on the x axes are different as well. If there are additional tweaks that we want, we can always easily modify the plot_gen() function.

I won’t go into the details here, but as a final thought, I always suggest saving images in svg format. This allows us to both scale the image to any size without loss in fidelity, and to further tweak things by hand in a program like Adobe Illustrator or Inkscape. A great example of an easy svg tweak appears in our final image above – you’ll notice that the x-axis is labeled with “Environment” on both plots. It would be nice to have that appear just once, centered below both plots. This is an easy fix with an svg image – we can just manually remove one of the “Environment” labels, and drag the other copy to where we want it.

The graphs look pretty close to what I imagined – our messy trait (oil content) has some crossover genotype-by-environment interaction, while our more stable trait (plant height) exhibits nice, parallel lines in its interaction plot. I’m sure any plant breeder would prefer to work with the latter trait.

Now, as promised, my explanation as to why we did this in the first place. My justification is that in this case, we are more interested in the patterns of interaction for each of these two traits, rather than specific data points. To be more specific, it’s a matter of the scope of our inference. I could imagine generating this sort of plot if we were using a model that treated both genotype and environment as random effects. In that case, we would be more interested in characterizing some pool of possible genotypes and environments that we sampled in our experiment, rather than characterizing the specific genotypes or environments that we happened to use. (I use the term pool instead of population here, because I find the latter gets confusing when we start talking about populations in the genetic sense vs. the statistical sense).

With that, you now you have everything you need to go forth and compile all manner of multi-panel figures in R – please just commit to using this power wisely.

References