Tips for scientific figures

Sometimes I get asked for tips or guidelines on how to create figures for scientific publications.

I do have a system to do this. It is not perfect, but it fits my workflow and I do improve it over time.

The “dream” of having a full final figure generated as a whole from a piece of code is just not worth it to the type of figures my papers usually have. Therefore only in very rare cases I even attempt to create a full figure or several panels directly from code. Since I have become very familiar with Inkscape, I can very quickly assemble plots into a figure with infinite customization which would be very painful to do in code. Plus building the figures in a graphical interface really helps me ‘play’ with the pieces of data and make a visual narrative for the paper, which then I translate to text - I am a visual person and every story starts with images for me.

In concept, my approach is quite simple:

1. generate plots individually as data is analyzed using matplotlib/seaborn;
2. assemble plots into a figure using Inkscape;
3. uniformly process all figures into various formats (separate PNG, PDF, joint PDF).

1. Making plots (Python, matplotlib, seaborn)

• Try to compartmentalize data processing from visualization (not just the early ‘pipeline’-like processing, but also the analytical analysis) - but that does not mean that auxiliary/intermediate plots can’t be generated as data is analyzed.
• Each script produces a set of plots into a results directory, with subdirectories matching the script or analysis name (more hierarchical if needed, e.g. different datasets).
• Each plot file name should be self explanatory and contain enough information to track down its origin.
• Except for obvious groupings or interrelations, avoid making figures with many subplots.
• Create figures in matplotlib/seaborn always explicitly: fig, ax = plt.subplots(figsize=(3, 3)).
• For the most part aim for each subplot to have a square shape with about 3 by 3 inches. If creating a figure with multiple subplots, scale the figure size accordingly: fig, axes = plt.subplots(4, 2, figsize=(2 * 3, 4 * 3).
• Always label the axes, and make use of a single statement to set many properties at once: ax.set(xlabel="Time", ylabel="Expression (log)", yscale="log").
• For plot elements with many objects (e.g. scatter), rasterize that specific single element in order to reduce the size of the figure: g = sns.clustermap(...); g.ax_heatmap.set(rasterized=True). Do not rasterize the whole axes as that will make e.g. text element uneditable.
• Color:
  • Choose consistent colors across a project. This can be achieved for example by importing a config file with preset colors for each factor level.
  • Use categorical colormaps for factors without implicit order. In general do not use more than 8 or 10 different colors.
  • Consider thinking of all factors being highlighted in the paper before starting the project. See colormaps here: https://matplotlib.org/stable/gallery/color/colormap_reference.html.
  • Avoid situations where a color is suggestive or is associated with human values such as heteronormative, ageist or racial biases (e.g. gender: blue/pink, age: bright to dark).
  • Use continuous colormaps for continuous variables (magma, viridis), especially if they are fully non-negative. In general, use the brightest color for higher values.
  • Use divergent colormaps (coolwarm, RdBu_r, PuOr_r) only when a central value has meaning (e.g. a Z-score).
• Export figures as SVG format, but explicitly set dpi=300 to make sure any rasterized elements have good quality. Set bbox_inches='tight to make sure the whole content is visible.

2. Assemble plots manually into a figure (inkscape)

• Check the journal requirements and limitations for figures and their dimensions. Use A4 by default, not letter.
• Here is a template figure: https://gist.github.com/afrendeiro/2d9975793629774ad73ea9d18bd7abf3.
• Add a plot to the canvas, resize it to an approximate desired size, remove all groupings, remove redundant objects, possibly despine plot (i.e. remove top right, top axes). Make whole plot into a group (or layer).
• Tip: after having several groups, you can still select a single element of a group without ‘entering’ the group by using Ctrl+click.
• Align elements of plots/panels to each other with the align tool in Inkscape (the ‘Last selected’ option helps).
• Be wary of resize operations that change aspect ratio. Be sure not to create artifacts. Here’s some help: https://github.com/burghoff/Scientific-Inkscape (hasn’t worked recently for me).
• Use a consistent font family (Arial) for all text elements and only one or two font sizes (12 and 10).
• Tip: You can use the Find/Replace tool with Object types = 'text' to find all text objects, or for example find all elements of the same color or linestyle (leave out the “ff” part of a RGBA string when searching properties).
• Add a lowercase letter label to each panel (font 16).
• Add a label to the top of the Figure (i.e. Figure 1).
• Name the figure files consistently ‘Figure01.svg’, ‘Figure02.svg’, ‘SupplementaryFigure01.svg’, ‘SupplementaryFigure02.svg’, etc.

3. Automatic assembly and conversion of file types (bash, inkscape, minify, pdfunite)

• Add all figures to a svg directory (itself under figures in the project directory).
• Make sure figures are labeled consistently and with zero padded numbering (if needed).
• Use script here: https://gist.github.com/afrendeiro/a656eca139381eec55da6357e33173ce.
• Use generated individual PDFs to embed in a manuscript tex file.
• Use generated individual PNGs to embed in a manuscript file like docx.
• Use the joint PDFs to print, share or submit to journal.

Final thoughts

• Learning the basics of matplotlib and seaborn is essential to make useful plots that display the data in a clear way.
• Learning the basics of inkscape is essential to make figures without too much hassle. There is only a few key hidden tricks (e.g. Ctrl+click) to make the process super easy.
• Ultimately each person has different needs and preferences and this is just what I do. Try out some of the elements and see what works for you.