Bar plots as Venn diagram alternatives

Venn diagrams are a type of figure that I see a lot in the published literature. While Venn diagrams are a great way to conceptualize the goal of looking for shared or unique transcriptional responses to experimental manipulations, I don’t think they are the best way to vizualize empirical data.

I searched for “Venn” in Garrett Grolemund & Hadley Wickham’s R for Data Science book and in Claus Wilke’s Fundamentals of Data Visualization book, for guidance alternatives to the Venn diagram. Neither book discusses Venn diagrams, but they both provide valuable insight into creating bar plots. Interestingly, the R for Data Science book does use Venn diagrams to illustrate the differences between inner_join and full_join, so that provides some evidence that Venn diagrams are useful for conceptualizing ideas about data.)

The UpSet plot has gained some popularity since 2016. I’ve made a handful of UpSet plots with upsetR. While it is possible to modify the aesthetics of an UpSet plot, it is very challenging to manipulate the size, shape, colors, and font to the degree that is needed to combine many plots into one multi-panel figure for a manuscript.

I quite like the bar charts in the UpSet plot, so I set out to recreate them with ggplot2 and cowplot. Here, I convert 12 Venn diagrams from published research articles (from labs that I admire and respect) to 4 bar plots that I made with tidyverse and cowplot. I think these bar plots are simple, flexible, effect, and reproducibility alternatives than a Venn diagram. You can explore all the code in my vennbar repository on GitHub, and you can run the code in the cloud with Binder.

Example 1 (above) is from Cognitive specialization for learning faces is associated with shifts in the brain transcriptome of a social wasp by Berens et al. 2017. In this example, the circles do not represent a meaningful quantity. A stacked bar plot can use color, space, and text to highlight patterns in the data, which in this case appears to be a greater overlap than expected. I wanted to make the bar plot mirror the Venn diagram as closely as possible, but I changed the order of the factors so that “both” category was plotted first. This was necessary for adding text to the bar chart because the values for P. metricus and both were overlapping when I kept the “both” category in the middle. Here is the code for my bar plot alternative.

Example 2 is from Transcriptomics of an extended phenotype: parasite manipulation of wasp social behaviour shifts expression of caste-related genes by Geffre et al.. As with example 1, the proportions of the circles are all equivalent, so they don’t convey any meaning about the size of the sets they represent. The bar chart really highlights that fact that that “worker versus gyne” (or “WorkGyne” for short) comparison yields the most differentially expressed genes. However, that was not the main message of the Venn diagram. The authors used an arrow an * to highlight one particular comparison, so I used a bright red fill to highlight the parasite manipulation candidate genes. Here is the code for my bar plot alternative.

Example 3 is from Sex-biased transcriptomic response of the reproductive axis to stress by Calisi et al. 2017. This weighted Venn diagram is the highest quality Venn diagram I’ve ever seen, but it still has limitations. If you only focus on the green pies, they seem to do a good job of conveying relative size, but if you look closely, you’ll notices that the size of the grey circles is misleading. The original Venn diagram conveys information about tissue (hypothalamus, pituitary, gonad), sex (male and female), and response (up- or down-regulation), so I used facetting and a combination of stacked and side-by-side bar plot to visualize all the complexity. I think that this bar chart does a better job of conveying the magnitude of differences between the sexes and across tissues and the bias toward up-regulation of gene expression. Here is the code for my bar plot alternative, as well as a less-squished version of the bar plot.

Example 4 is from Early life stress alters transcriptomic patterning across reward circuitry in male and female mice by Peña et al. 2017. Here, the Venn diagrams are also scaled, but each of the six subplots has their own scaling, so you cannot compare all six visually. Moreover, the authors report the percent of shared gene expression in the manuscript, but their Venn diagrams show counts. So, I calculated the percent of shared differentially expressed genes (%DEGs) and colored all overlapping responses in shades of purple and the unique responses in shades of orange. Now, once can spot interesting trends in the data, like increase response to early and late life stress (ELS + STVS) in the female PFC. This is something that is present but not emphasized in the Venn. Here is the code for my bar plot alternative.

So, that’s my comparison of four Venn diagrams and four possible bar plot alternatives. What do you think? Do you see the advantages? What other alternatives do you suggest? Comments welcome! (Thanks to Titus Brown, Claus, Wilke, and Victoria Farrar for insightful comments on an earlier draft.)

References

1. Berens et al. 2017: http://jeb.biologists.org/content/220/12/2149
2. Geffre et al. 2017: https://royalsocietypublishing.org/doi/full/10.1098/rspb.2017.0029
3. Calisi et al. 2017: https://www.sciencedirect.com/science/article/pii/S0018506X17302696?via%3Dihub)
4. Peña et al. 2019: https://www.biorxiv.org/content/10.1101/624353v1