Visual encoding, colour, and figure integrity

Summary

Colour and other visual encodings carry information. They must be chosen to reveal true signal, not distract. Always ensure sufficient contrast: colorblind-safe palettes (e.g. blue–orange). Never rely on colour alone to distinguish categories; add redundant markers (shape, line type). Keep fonts legible. Avoid chartjunk: gridlines and 3D effects that don’t encode data distort perception. Preserve data-ink ratio by minimizing unnecessary background. Each plot must stand alone: include annotations (e.g. “*” for significance only if explained; no decorative doodles).

Core rules

• Colour choice: Use palettes where every category can be differentiated in grayscale. For sequential data, use a single-hue gradient; for diverging data, use two contrasting hues with a white midpoint.
• Contrast: Ensure text and lines meet WCAG contrast ratios (≥4.5:1).
• Redundant encoding: Encode categorical differences with shape or dashing in addition to colour.
• Minimal ornamentation: Do not add 3D bars, “exploding” pie slices, or heavy shadows. Use simple lines, bars, etc.
• Direct labeling: Label lines or bars directly when possible instead of a separate legend.

Required context

• Audience’s needs (e.g. color-deficiency prevalence).
• Whether figures will be printed (grayscale test) or displayed digitally.

AI behaviour

• Palette check: If too many categories for distinct colors, raise a warning (or group categories).
• Convert to grayscale: Validate that the plot conveys information when desaturated.
• Clutter removal: Remove non-informative elements (excess ticks, backgrounds) to highlight data.
• Legible text: Check point sizes and font sizes are readable at final figure resolution.

Common failure modes

• Low contrast: Light grey text on white background or red/green on white without sufficient contrast.
• Overreliance on colour: Color differences without shapes (red vs green lines) excludes ~10% viewers.
• Chartjunk: 3D bar charts or unnecessary pictograms that do not map to data quantities.

Authoritative standards

• Accessible design: Follow WCAG guidelines for charts (contrast, redundancy).
• Perception studies: Cleveland & McGill’s hierarchy (position > length > angle > area > colour) suggests using position/length encodings over area or color for quantitative comparisons.

Examples

Example: Multi-category scatter

• Data: Expression by cell type (6 categories).
• Check: Use a colorblind-friendly palette (e.g. scale_color_viridis_d() or blue/orange pair). Use different shapes for lineages as well.
• Good outcome: Legend shows color and shape mapping, and the chart remains interpretable in grayscale.

Example: Time-series lines

• Data: Trends for 3 conditions.
• Check: Two conditions with similar color must differ in line type. Add data labels for key points if close.
• Good outcome: One line solid blue, one dashed green, one dotted red; each labelled in-plot or via clear legend.