Color and Accessibility in Scientific Figures

This skill is the audit and palette layer for everything else in the visualization category. It is the answer to: "Will every reader of my figure — including the ~8% of men and ~0.5% of women with color vision deficiency (CVD), the low-vision reader, the screen-reader user, the projector-on-yellow-tint reader — extract the same message?" Use it whenever a figure is destined for publication, a talk, a poster, or a public web page.

When to use

Trigger this skill when any of the following apply:

• You are picking a palette for a new figure.
• A figure uses color as the only encoding for groups, categories, or values.
• A reviewer or co-author has flagged a figure as hard to read.
• You are publishing a figure to a journal that requires alt text and contrast guarantees.
• You are presenting at a venue with a projector of unknown color profile.
• You are designing a figure for a public-facing web page (blog, GitHub README, lab website) that screen readers and CVD readers will encounter.

When NOT to use

• Pure data exploration with seaborn defaults — use ors-scientific-visualization-figure-design and pick a default palette.
• A schematic / workflow figure where most elements are not data — use ors-scientific-visualization-schematics-diagrams and apply accessibility at the end.
• A slide deck or poster — use ors-scientific-visualization-slides-design or ors-scientific-visualization-poster-design (which link back here for the color audit).

Prerequisites

• A draft of the figure (matplotlib, seaborn, ggplot2, plotly, or a saved PNG/PDF).
• A simulator: a CVD simulator (Coblis, Sim Daltonism, or the colorblind Python package), a greyscale conversion (gs or ImageMagick), and a contrast checker (WebAIM).
• A target contrast standard: WCAG 2.1 AA (4.5:1 for normal text, 3:1 for large) is the minimum; AAA (7:1) is the recommended target for body text in scientific figures.
• Knowledge of the data type: ordinal (low/medium/high), categorical (mutually exclusive groups), sequential (continuous, one direction), or diverging (continuous, with a meaningful center).

Core workflow

1. Identify the data type. Ordinal, categorical, sequential, or diverging. The data type picks the palette family.
2. Pick a perceptually uniform palette by default. For sequential and diverging data, start with viridis, cividis, plasma, or magma. These are CVD-friendly by design.
3. For categorical data, start with Okabe-Ito (8 CVD-safe colors) or a ColorBrewer qualitative palette. Do not exceed 6–8 groups per figure; split if you must.
4. Audit with a CVD simulator. Render the figure under deuteranopia, protanopia, and tritanopia simulations. Any pair of groups that becomes indistinguishable needs a redundant encoding (shape, line style, label).
5. Test in greyscale. Convert the figure to greyscale and confirm the data is still readable. If it is not, your encoding is too color-dependent.
6. Check contrast. Text and important graphical elements must meet WCAG 2.1 AA at minimum. Body text → 4.5:1; large text (≥ 18 pt or ≥ 14 pt bold) → 3:1; graphical elements → 3:1.
7. Add redundant encoding. Pair color with shape, line style, marker, fill pattern, or direct label. The combination is the message; the color alone is decoration.
8. Write alt text. 1–3 sentences that name the figure type, the main message, and the key elements. Alt text is for screen readers and for the journal's accessibility metadata.
9. Audit under projector conditions. A 4.5:1 contrast on a laptop may wash out to 3:1 on a projector with a yellow tint. Verify on the actual projector or a calibrated sRGB → projector profile.
10. Document the palette. Note the palette name and source in the figure caption or methods. This makes the figure reproducible.

Code patterns

A. Data type → palette family decision tree

What does the color encode?
│
├── A categorical group (e.g., cell type, treatment)
│     → Okabe-Ito or ColorBrewer qualitative
│     → Max 6–8 groups per figure
│
├── A continuous variable, low → high
│     → viridis, cividis, plasma, magma, inferno
│     → Sequential, perceptually uniform
│
├── A continuous variable with a meaningful center (e.g., log2FC, correlation)
│     → RdBu_r, BrBG, PuOr, Spectral (with caveats), coolwarm
│     → Diverging
│
├── A binary on/off
│     → Two Okabe-Ito colors with strong contrast (e.g., #0072B2 vs. #D55E00)
│
└── A heatmap with no natural center
      → viridis (default), then cividis, then magma
      → Avoid jet, rainbow, and "Spectral" for heatmaps

B. Okabe-Ito palette (the categorical default)

The Okabe-Ito 8-color palette was designed to be distinguishable under the three main forms of CVD.

okabe_ito = {
    "black":      "#000000",
    "orange":     "#E69F00",
    "sky_blue":   "#56B4E9",
    "bluish_grn": "#009E73",
    "yellow":     "#F0E442",
    "blue":       "#0072B2",
    "vermilion":  "#D55E00",
    "r_purple":   "#CC79A7",
}
# In matplotlib
import matplotlib.pyplot as plt
plt.rcParams["axes.prop_cycle"] = plt.cycler(color=list(okabe_ito.values()))
# In seaborn
import seaborn as sns
sns.set_palette(list(okabe_ito.values()))

R / ggplot2:

okabe_ito <- c("#000000","#E69F00","#56B4E9","#009E73",
               "#F0E442","#0072B2","#D55E00","#CC79A7")
scale_color_manual(values = okabe_ito)

C. viridis and friends (sequential, CVD-friendly by design)

import matplotlib.pyplot as plt
import numpy as np

fig, axes = plt.subplots(1, 4, figsize=(10, 2.2))
for ax, cmap in zip(axes, ["viridis", "cividis", "plasma", "magma"]):
    img = np.linspace(0, 1, 256).reshape(1, -1)
    ax.imshow(img, cmap=cmap, aspect="auto")
    ax.set_title(cmap, fontsize=10)
    ax.set_xticks([]); ax.set_yticks([])

These four palettes are perceptually uniform, CVD-friendly, and print legibly in greyscale.

D. CVD simulation (Python)

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.colors as mcolors

# Simulate deuteranopia / protanopia / tritanopia
# by transforming the RGB image through a CVD matrix.
def simulate_cvd(rgb, cvd_type="deuteranopia"):
    """
    Apply a color-vision-deficiency simulation matrix to an RGB image.
    cvd_type in {'deuteranopia','protanopia','tritanopia','achromatopsia'}.
    """
    # Brettel/Vienot/Mollon CVD matrices — loadable from a library
    # (e.g., `colorblind` package) or from a published lookup table.
    # For brevity, use the `colorblind` package:
    from colorblind import simulate as cb
    return cb.colorblindness(rgb, cvd_type=cvd_type)

img = plt.imread("figure.png")
plt.imsave("figure_deuteranopia.png", simulate_cvd(img, "deuteranopia"))
plt.imsave("figure_protanopia.png",   simulate_cvd(img, "protanopia"))
plt.imsave("figure_tritanopia.png",   simulate_cvd(img, "tritanopia"))
plt.imsave("figure_greyscale.png",    simulate_cvd(img, "achromatopsia"))

The colorblind package, the viscm library, and the colorspacious library all expose CVD matrices; the Brettel-Vienot-Mollon transform is the standard reference.

E. WCAG 2.1 contrast — programmatic check

def hex_to_rgb(h):
    h = h.lstrip("#")
    return tuple(int(h[i:i+2], 16) for i in (0, 2, 4))

def relative_luminance(rgb):
    """WCAG 2.1 relative luminance from sRGB 0–255."""
    def channel(c):
        c = c / 255.0
        return c / 12.92 if c <= 0.03928 else ((c + 0.055) / 1.055) ** 2.4
    r, g, b = (channel(c) for c in rgb)
    return 0.2126 * r + 0.7152 * g + 0.0722 * b

def contrast_ratio(fg_hex, bg_hex):
    L1 = relative_luminance(hex_to_rgb(fg_hex))
    L2 = relative_luminance(hex_to_rgb(bg_hex))
    lighter, darker = max(L1, L2), min(L1, L2)
    return (lighter + 0.05) / (darker + 0.05)

# Example
ratio = contrast_ratio("#000000", "#FFFFFF")
print(f"Black on white: {ratio:.2f}:1")
# Black on white: 21.00:1   (passes AAA)

WCAG 2.1 targets:

For axis labels, legends, and figure captions, target AA at minimum, AAA when feasible. For a graphical line on a white background, target 3:1 minimum (line) and 4.5:1 for the legend text.

F. Greyscale conversion (for the greyscale test)

# Ghostscript
gs -sDEVICE=pngalpha -sColorConversionStrategy=Gray \
   -o figure_greyscale.png figure.pdf

# ImageMagick
convert figure.png -colorspace Gray figure_greyscale.png

If the figure tells its story in greyscale, the encoding is robust. If not, add redundant encoding.

G. Texture and pattern as redundant encoding

When color alone is not enough, pair it with a pattern, marker shape, or line style. This is critical for figures that may be printed in greyscale.

import matplotlib.pyplot as plt
import numpy as np

fig, ax = plt.subplots(figsize=(4, 3))
x = np.arange(5)
ax.bar(x - 0.2, [1, 2, 3, 4, 5], width=0.4,
       color="#0072B2", hatch="///",  label="Control")
ax.bar(x + 0.2, [2, 3, 1, 5, 4], width=0.4,
       color="#D55E00", hatch="\\\\", label="Treatment")
ax.set_xticks(x); ax.set_xticklabels(["A", "B", "C", "D", "E"])
ax.legend(frameon=False)

Hatch patterns in matplotlib: /, \\, |, -, +, x, ., o, O, *. Use 2–3 patterns at most.

Line styles in matplotlib: '-' (solid), '--' (dashed), '-.' (dash-dot), ':' (dotted). Pair each line style with a marker for maximum redundancy.

Marker shapes: 'o', 's', '^', 'D', 'v', 'P', '*'. Match marker shape to line style or color for clarity.

H. Alt text (writing the figure description)

A figure's alt text should answer three questions: what type of figure is this?, what is the main message?, and what are the key elements? It should be 1–3 sentences for a simple figure, up to 5 for a multi-panel composite.

Examples:

• "Volcano plot. Differentially expressed genes (n = 1,243; padj < 0.05, |log2FC| > 1) are shown in red; non-significant genes in gray. The strongest induction is gene X (log2FC = 4.2, padj = 1e-30)."
• "Multi-panel figure with 4 UMAP plots. Panel A shows the full dataset (50,000 cells). Panel B highlights cluster 7 (cyan, 1.2% of cells). Panel C shows that cluster 7 is enriched in patients with outcome Y. Panel D shows the marker genes for cluster 7."
• "Schematic workflow. FASTQ files are QC-filtered with fastp, aligned with STAR, quantified with featureCounts, and tested for differential expression with DESeq2. The output is a results table feeding a volcano plot and a GO enrichment analysis."

Do not start with "Image of" or "Figure showing"; screen readers already announce that the element is a figure.

CVD simulator matrix

Run every important figure through at least one of these before submission. If two groups become indistinguishable under deuteranopia (the most common CVD, ~6% of men), the figure is not safe for the global readership of a journal.

Color choices to avoid

Some palettes are widely used in science but fail accessibility:

• Jet / rainbow: a 1970s MATLAB default that introduces false boundaries (perceptual non-uniformity) and is especially bad for greyscale and CVD. Use viridis or cividis.
• Spectral: a ColorBrewer diverging palette that is not perceptually uniform. Use RdBu_r or BrBG instead.
• Red-green only: distinguishes two groups by hue alone, and fails for ~8% of male readers. Use Okabe-Ito orange/blue or a CVD-safe pair.
• Pure red on pure green, pure red on pure blue, or pure black on pure red: low contrast for any reader. Use a CVD-safe pair with at least 3:1 contrast.
• Saturated colors on saturated backgrounds: the high-saturation + high-saturation combination vibrates and is hard to read on a projector. Use a tinted background (off-white) instead of pure white, and a desaturated foreground.
• Traffic-light colors for non-traffic-light data: the cultural association (red = bad, green = good) is not universal. For continuous data, do not encode "low = red, high = green" without explicit labels; use the full hue range so that the perceptual ordering is preserved.

Common pitfalls

• Color as the only encoding: red dots for cases, blue dots for controls. Add markers, panels, or direct labels.
• Six colors that look fine in sRGB and collapse in print: pick a palette that survives the journal's color profile (usually CMYK). The Okabe-Ito and viridis palettes are print-safe.
• Background mismatch: a figure designed on a white background viewed on a black slide-deck. Audit the figure against the actual background it will be presented on.
• A 4.5:1 ratio that drops to 3:1 under projector gamma: re-test on the actual projector.
• ColorBrewer output as raw hex without testing: ColorBrewer is a starting point, not a guarantee. Always CVD-test.
• Alt text that is just the filename: "figure1.png" tells a screen reader nothing.
• Inconsistent palette across the manuscript: figure 1 uses viridis, figure 2 uses jet, figure 3 uses a custom palette. Pick a project-wide palette and use it.
• Saturation over hierarchy: every category is at 100% saturation, and nothing is the most important. Reduce saturation for non-emphasis, keep one or two elements at full saturation as the anchor.

Validation

• CVD simulation: render the figure under deuteranopia, protanopia, tritanopia, and achromatopsia. Every group or value must remain distinguishable.
• Greyscale test: convert to greyscale (gs -sColorConversionStrategy=Gray or ImageMagick). The figure must still tell its story.
• Contrast test: every text and graphical element meets WCAG 2.1 AA at minimum. Body text → 4.5:1; graphical → 3:1.
• Projector test: view the figure on the actual projector or a similar one. The figure is not ready until it survives the venue.
• Alt-text test: paste the figure (no caption) into a document and read the alt text to a colleague. The colleague should be able to sketch the figure from the alt text alone.
• Print test: print a color figure on a black-and-white printer. Important information must not be lost.
• Color-blind colleague test: the most reliable test is to show the figure to a colleague with CVD. If they can read the figure without explanation, it passes.

Open alternatives

References

Internal cross-links to other ors-* skills:

• ors-scientific-visualization-figure-design — apply the chosen palette to a publication-ready figure.
• ors-scientific-visualization-schematics-diagrams — apply the palette to a schematic / workflow.
• ors-scientific-visualization-slides-design — apply the palette to a deck.
• ors-scientific-visualization-poster-design — apply the palette to a poster.

External resources (do not fabricate exact paths):

• ColorBrewer 2.0 (Cynthia Brewer) — colorbrewer2.org
• viridis colormap family — official documentation and motivation
• Okabe-Ito palette (Masataka Okabe and Kei Ito, 2008) — jfly.uni-koeln.de/color
• Coblis color-blindness simulator — color-blindness.com/coblis-color-blindness-simulator
• Sim Daltonism — michelf.ca/projects/sim-daltonism
• Color Oracle — colororacle.org
• WebAIM contrast checker — webaim.org/resources/contrastchecker
• WCAG 2.1 — w3.org/TR/WCAG21/
• colorblind Python package — pypi.org/project/colorblind
• colorspacious Python package — pypi.org/project/colorspacious
• Servier Medical Art (CC-BY biomedical icon library) — smart.servier.com
• National Eye Institute, color vision deficiency statistics (prevalence of red-green CVD in the population)

Changelog

• 1.0.0 (2026-06-10): Initial adaptation by Pradyumna Jayaram from scientific-visualization and bio-data-visualization-color-palettes (K-Dense Inc.). Merged the two source skills into a single color-and-accessibility skill, added a CVD simulator matrix, redundant-encoding patterns, WCAG 2.1 contrast code, an alt-text template, and a "colors to avoid" section with rationale.