A figure can pass every journal’s technical checklist — correct DPI, an accepted file format, a properly labeled panel scheme — and still fail at its actual job, which is letting a reader understand a result in a few seconds without misreading it. That is a design problem, not a compliance problem, and it is solved earlier in the process, at the point where a researcher decides what kind of chart to draw, what to emphasize, and what to leave out. This guide covers that design and creation stage: choosing the right chart type for the data, building visual hierarchy so the reader’s eye goes to what matters, choosing a colorblind-accessible palette as a design decision (not just a submission requirement), and avoiding the handful of design pitfalls — chartjunk, truncated axes, 3D effects — that distort or obscure data even in a technically compliant file. For the separate question of resolution, file format, panel labeling, numbering conventions, and image-integrity rules journals check before acceptance, see CASRAI’s Scientific Figure Rules for Journal Submission guide. Read the two together: this one covers making a figure good; that one covers making it compliant.
Choosing the Right Chart Type for Your Data
The most common figure-design failure is not a bad execution of the right chart — it is choosing the wrong chart type for what the data actually is. Each of the standard chart types encodes a specific kind of comparison, and picking one that doesn’t match the underlying question forces the reader to do work the figure should be doing.
Bar charts: comparing discrete categories
Bar charts are for comparing a measured quantity across a small number of discrete, unordered or nominally ordered categories — treatment groups, species, conditions, time points treated as categories rather than a continuum. The length of the bar is the only thing being judged, which is why bar charts should almost always start at zero: truncating the baseline (covered below) turns a length comparison into a misleading one. Grouped or stacked bars extend this to a second categorical variable, but stacking should only be used when the total (the sum of the stack) is itself a meaningful quantity — if it isn’t, grouped bars side by side are usually clearer than a stack.
Line charts: showing a trend over a continuous variable
Line charts are for showing how a measured quantity changes across a continuous or ordered variable, most often time. The connecting line itself carries meaning — it implies the values in between the plotted points are part of the same continuous process — which is why a line chart is the wrong choice for genuinely discrete, unordered categories (that’s what a bar chart is for). Multiple lines on one axis work well for comparing a small number of trends (roughly five or fewer before color and line-style discrimination becomes difficult); beyond that, small multiples (a grid of separate panels, one trend per panel, sharing the same axes) usually communicate more clearly than crowding every line onto one plot.
Scatter plots: showing the relationship between two variables
Scatter plots are for showing the relationship between two continuous variables measured on the same set of observations — whether they’re correlated, and roughly how strongly. Each point is one observation, which is the scatter plot’s core strength (it doesn’t summarize away individual data points the way a bar chart of means does) and its core weakness at scale: with a large number of observations, points overlap and overplotting hides the actual density of the data. Partial transparency (alpha blending), point-size reduction, hexagonal binning, or a 2D density/contour overlay are standard fixes for overplotting; a scatter plot of thousands of fully opaque points is a common design failure that makes dense regions look like a single solid blob.
Heatmaps: encoding a third variable in color across a matrix
Heatmaps are for data that naturally forms a matrix — two categorical or ordered axes (genes and samples, conditions and time points, a correlation matrix) with a continuous value encoded as color at each cell. Because a heatmap asks the reader to judge magnitude by color rather than by position or length — a comparison people are measurably worse at making precisely — the colormap choice matters more here than in almost any other chart type. A perceptually uniform, continuous colormap (see the palette section below) is close to a requirement for a heatmap to be read accurately, not just an accessibility nicety.
A simple decision framework
Before opening plotting software, it helps to answer three questions: What kind of variable is on each axis (categorical, continuous, ordered, time)? What comparison does the reader actually need to make (a magnitude between categories, a trend, a relationship between two variables, a pattern across a matrix)? And how many series or categories are being shown at once? Bar charts answer category-to-category magnitude questions; line charts answer trend questions; scatter plots answer relationship questions; heatmaps answer matrix/pattern questions. A figure that tries to answer more than one of these at once is usually a sign it should be split into two panels rather than forced into one chart type.
Visual Hierarchy and Clarity
Statistician and information-design writer Edward Tufte’s The Visual Display of Quantitative Information (1983) introduced two ideas that remain the working vocabulary for figure design decades later: the data-ink ratio — the proportion of a figure’s ink (or, on screen, pixels) that represents actual data rather than decoration, borders, or redundant gridlines — and chartjunk, Tufte’s term for decorative elements that add visual noise without adding information. The practical version of this principle for a research figure is straightforward: every gridline, border, background shade, drop shadow, gradient fill, or decorative icon should earn its place by helping the reader interpret the data; anything that doesn’t should be removed.
Visual hierarchy is the related, complementary idea: a well-designed figure guides the reader’s eye to what matters first, through position, size, contrast, and color, and lets secondary information recede. Concretely, this means: the data itself (bars, lines, points) should have the highest visual contrast in the figure; axis lines, tick marks, and gridlines should be thin, light, and unobtrusive relative to the data; legends and labels should be legible at the figure’s final printed size, not just on a laptop screen at full zoom; and a figure should generally make one primary point, with secondary comparisons handled by panels or a separate figure rather than crowding everything into a single dense plot. Journals’ own author guidance frequently states a minimum legible font size for figure text at final print size (commonly around 5–8 pt at the figure’s final column width, though the exact number varies by publisher) precisely because labels that are legible during figure preparation at screen size routinely become illegible once resized to a print column — always check the target journal’s specific figure guidelines for its stated minimum, and test a figure at its actual final printed dimensions before submission, not just on screen.
Choosing a Colorblind-Accessible Palette
Color choice is a design decision as much as it is a compliance one: roughly 1 in 12 men and a much smaller share of women (commonly cited around 1 in 200) have some form of red-green color vision deficiency, which means a meaningful fraction of any journal’s readership cannot reliably distinguish red from green in a figure. The design question is which kind of palette a given figure needs, since the three palette types encode different kinds of data and choosing the wrong type is itself a design error independent of accessibility:
- Qualitative (categorical) palettes — distinct hues with no implied order, for distinguishing unordered categories (treatment groups, species, conditions). The Okabe-Ito eight-color palette is the widely adopted colorblind-safe standard here, popularized across scientific publishing following Bang Wong’s 2011 Nature Methods “Points of view” column on color blindness.
- Sequential palettes — a single hue (or a smooth gradient) that varies in lightness or saturation, for a continuous variable that runs low to high (concentration, expression level, density). Viridis and similar perceptually uniform colormaps are the standard choice: equal steps in the underlying data correspond to equal perceived steps in color, which a traditional “rainbow” or “jet” colormap does not achieve, distorting how differences in the data are perceived even before accessibility is considered.
- Diverging palettes — two contrasting hues meeting at a neutral midpoint, for data with a meaningful zero or reference value in the middle (a correlation matrix running from -1 to +1, a fold-change value, a difference from baseline). A diverging palette should be symmetric around its true midpoint, and should avoid red-green as the two end colors for the same accessibility reason as categorical palettes.
Two practical design habits catch most remaining accessibility problems: run the finished figure through a colorblindness simulator before finalizing it (several free browser-based and desktop tools exist for this), and never rely on color as the only channel encoding a distinction — pairing color with a second redundant cue (different point shapes, line styles, or direct labels) means the figure still communicates correctly even if a specific color pair happens to be hard to distinguish for a given reader. For the specific hex values, journal-by-journal accessibility policies, and how this fits into a technical pre-submission check, see the color-accessibility section of the Scientific Figure Rules for Journal Submission guide.
Common Design Pitfalls That Undermine a Figure
Chartjunk and unnecessary decoration
Gradient fills, drop shadows, decorative background images, unnecessary borders around every panel, and ornamental icons are all chartjunk in Tufte’s sense: they consume the reader’s attention without conveying information, and at worst they actively compete with the data for visual priority. The test is simple — if removing an element doesn’t remove any information, remove it.
3D effects that distort data
3D bar charts and 3D pie charts are a design pitfall specifically because the third dimension distorts the very comparison the chart exists to convey: perspective makes bars or wedges nearer the viewer appear larger than equal-value bars or wedges farther away, and the added depth cue makes precise magnitude judgment measurably harder than the flat 2D equivalent. Unless a figure is genuinely representing three continuous spatial dimensions (a 3D structural rendering, a surface plot of an actual three-variable relationship), a 3D chart is very rarely conveying real three-dimensional information — it is a 2D comparison rendered with a distortion that makes it harder, not easier, to read accurately. The flat 2D version of the same chart is almost always the correct choice.
Truncated or misleading axes
A bar chart’s central visual claim is that length is proportional to value, which only holds if the axis starts at zero. Truncating the y-axis — starting it above zero to exaggerate small differences between bars — is one of the most common ways a technically accurate figure creates a misleading visual impression: a genuinely small difference in the underlying numbers can be made to look dramatic simply by cropping the baseline. Line charts are more tolerant of a non-zero baseline since the reader is judging the trend’s shape and slope rather than each point’s absolute length from a shared origin, but even there, an axis break, an inconsistent scale, or a log scale used without clearly labeling it as such can mislead a reader who is (reasonably) assuming a linear scale by default. Whenever an axis doesn’t start at zero or isn’t linear, that has to be visually obvious and explicitly labeled, not something the reader has to notice by checking the tick values.
Dual y-axes and other correlation-implying tricks
A figure with two different y-axes (one on the left, one on the right, each with its own scale) plotted against a shared x-axis lets an author choose the scaling of each axis independently — and a poor or manipulative choice of relative scaling can make two genuinely unrelated trends appear to move together, or two related trends appear to diverge, purely as an artifact of where each axis happens to start and how it’s stretched. Where a dual-axis figure is genuinely necessary (two variables measured in different, non-comparable units, both relevant to the same x-axis), the safest practice is to anchor both axes at a meaningful, consistent reference point (such as both starting at zero, or both showing percent change from baseline) rather than independently rescaling each one to maximize apparent visual overlap.
Overloading a single figure with too much information
A figure that tries to show every result from an experiment in one panel — many colors, many chart types combined, dense annotation — usually communicates less than a figure that shows one clear result and moves everything else to a second panel, a supplementary figure, or a table. If a figure needs an extended paragraph in the caption just to explain what’s being shown, that is often a sign the figure itself should be redesigned or split rather than more thoroughly captioned.
A Practical Figure-Design Checklist
- Match the chart type to the actual comparison being made: bar charts for categorical magnitude, line charts for trends, scatter plots for relationships, heatmaps for matrix patterns.
- Start bar-chart axes at zero; make any non-zero or non-linear axis explicit and clearly labeled, never something the reader has to infer from tick values.
- Remove any element — gridline, border, gradient, shadow, decorative icon — that doesn’t convey information.
- Give the data itself the highest visual contrast in the figure; keep axis lines, tick marks, and gridlines light and unobtrusive.
- Choose the right palette type for the data: qualitative for unordered categories, sequential for a continuous low-to-high variable, diverging for data with a meaningful midpoint — and make each one colorblind-safe (e.g., Okabe-Ito for categorical, viridis for continuous).
- Pair color with a second redundant visual cue (shape, line style, direct labeling) rather than relying on color alone to carry a distinction.
- Avoid 3D bar and pie charts unless the figure is genuinely representing three spatial dimensions.
- Test the figure’s text and line weights at its actual final printed size, not just on a full-screen monitor during preparation.
- Aim for one clear primary message per figure; split a crowded figure into panels or a supplementary figure rather than layering in more information.
- Once the design is right, run it against the target journal’s technical requirements — resolution, file format, panel labeling — covered in the Scientific Figure Rules for Journal Submission guide.
Frequently Asked Questions
How do I choose the right chart type for my data?
Start from the comparison the reader actually needs to make, not the chart type you’re used to using. Bar charts compare a quantity across discrete categories; line charts show a trend across a continuous or ordered variable (usually time); scatter plots show the relationship between two continuous variables measured on the same observations; heatmaps show a continuous value across a matrix of two categorical or ordered axes. If a figure seems to need more than one of these at once, that’s usually a sign to split it into separate panels.
What is chartjunk?
Chartjunk is Edward Tufte’s term, from The Visual Display of Quantitative Information (1983), for decorative elements in a chart — gradient fills, drop shadows, unnecessary borders, ornamental icons — that add visual noise without conveying any additional information. The related concept, data-ink ratio, is the proportion of a figure’s visual content that actually represents data; the practical design rule is to remove any element that doesn’t carry information.
Why are truncated y-axes considered misleading?
A bar chart’s core visual claim is that bar length is proportional to the underlying value, which only holds if the axis starts at zero. Starting the axis above zero exaggerates the visual difference between bars relative to the actual difference in the numbers, which is why truncated bar-chart axes are treated as a design (and sometimes integrity) problem rather than a neutral stylistic choice. Any non-zero or non-linear axis should be made visually obvious and explicitly labeled.
Should I use 3D bar charts or 3D pie charts in a scientific figure?
Generally no. Perspective in a 3D chart distorts the apparent size of bars or pie wedges based on their position relative to the viewer, making magnitude comparisons measurably harder and less accurate than the equivalent flat 2D chart — without adding real information, since most 3D bar and pie charts are not actually representing three spatial dimensions of data. Reserve 3D rendering for figures genuinely showing a three-dimensional spatial relationship, such as a molecular structure or a true three-variable surface plot.
How is this guide different from CASRAI’s Scientific Figure Rules guide?
They cover two different stages of figure preparation. This guide covers design and creation: choosing the right chart type, building visual hierarchy, picking an appropriate and colorblind-accessible palette, and avoiding pitfalls like chartjunk, truncated axes, and distorting 3D effects — the decisions that make a figure clear and honest in the first place. The Scientific Figure Rules for Journal Submission guide covers what happens once a figure is designed: the resolution, file format, panel-labeling, numbering, and image-integrity requirements a specific journal checks before accepting it. A well-designed figure still needs to pass the technical and integrity checks; a technically compliant figure still needs good design to communicate clearly.
