Data Visualization for Biological Research: Choosing Figures That Preserve Meaning

If you have ever stared at a PCA plot that seems to hide your key biological finding or struggled to explain why a log‑transformed histogram is more informative than raw counts, this guide is for you. It is written for bench scientists, bioinformaticians, and students who need to translate complex biological data into clear, honest figures. We will cover how to select a plot type, when and why to transform data, how to show uncertainty, how to make visuals accessible, and why an exploratory figure should not be used as a final presentation. The goal is not to prescribe one “correct” plot but to give you decision criteria you can apply to your own data. For a broad, authoritative foundation on data handling in biology, see the NCBI Bookshelf which hosts free textbooks on statistics and bioinformatics. For hands‑on training in plotting and analysis, the EMBL‑EBI Training platform offers structured courses on visualization best practices.

At a Glance

Plot Selection for Biological Data

Every figure should start with a clear question: Are you comparing groups, showing a relationship, or looking at structure in high‑dimension data? The answer determines the plot type.

For comparing distributions (e.g., gene expression across conditions), boxplots, violin plots, or stripcharts are standard. Stripcharts can show every point when sample sizes are small, but they become unreadable above a few dozen points. Violin plots add density information, boxplots are compact. The Galaxy Training Network provides workflows that often produce such plots after differential expression analysis, and you can see examples of how each plot type behaves with real RNA‑seq data.

For continuous x versus continuous y (e.g., correlation between two metabolites), simple scatter plots are best, with a trend line (smoother or linear) if appropriate. Avoid using a heatmap for this relationship, heatmaps are for matrices (e.g., expression values across samples and genes).

For high‑dimensional data (e.g., RNA‑seq samples in thousands of genes), PCA plots and t‑SNE/UMAP projections are common. But be aware that t‑SNE can create false clusters when perplexity is mis‑set. A recent study in PNAS introduced provable cluster‑preserving visualizations with curvature‑based stochastic neighbor embeddings [8] that aim to reduce such artifacts. If your data has known cluster structure, consider validating with multiple methods.

For gene‑set or pathway‑level results, enrichment maps show overlapping gene sets as nodes connected by edges. The MondrianMap tool [11] offers multi‑resolution enrichment maps that let you navigate hierarchies without cluttering the figure. That kind of hierarchical visualization is especially useful when you have hundreds of significant gene sets.

Transformations and Scaling

Raw biological measurements often follow a log‑normal or negative binomial distribution. Count data from RNA‑seq or ChIP‑seq can span orders of magnitude. Plotting raw values will compress most observations into a narrow range, hiding patterns. A log transformation (log2 or log10) spreads out the low end and pulls in the high end, making differences visible.

For RNA‑seq, the standard is to use a variance‑stabilizing transformation (VST) or a regularized log transformation (rlog) before PCA or heatmaps. The Bioconductor project (especially the DESeq2 and limma packages) provides functions for these transformations and includes documentation on why they are needed. The transformed values are no longer counts, but they preserve the rank ordering and allow distance‑based methods to work properly.

When you transform, always label the axes clearly: “log2 normalized counts” not just “expression”. And note that transformation changes the visual interpretation. A fold change of 2 between groups on a log scale is a constant distance, while on a raw scale it can look different at high vs low expression. The Sequence Read Archive [5] stores raw sequencing data, and if you download example datasets you can practice transforming and plotting to see this effect yourself.

One common mistake is to transform only the axis but not the underlying calculation. For instance, bar charts of log2 values are acceptable if you explain the transformation, but the bar heights then represent geometric means, not arithmetic means. Better to show the raw data points with a transformed y‑axis tick marks if possible.

Communicating Uncertainty

A figure that only shows point estimates (e.g., fold change or mean expression) without variability can mislead. Always include some measure of uncertainty. For comparisons between groups, error bars (standard deviation or standard error) are standard, but they come with caveats. Standard error bars are narrower and can make differences look more significant than they are, standard deviation bars show the spread of the data. A better option is to show the full distribution with a violin or boxplot.

For estimates from many bootstrapped samples, consider using a “bean plot” or “half‑violin” with a dot for the median. The EMBL‑EBI Training materials on statistical graphics emphasize that confidence intervals (CIs) are more informative than p‑values alone. If you show a CI, label it clearly (e.g., “95% CI”). For large datasets, a simple scatter of all points with a trend line and shaded confidence band is often the most honest visualization.

Uncertainty also applies to sequencing data. For example, a lowly expressed gene has higher biological variability. When you plot normalized counts, you can overlay error bars that are based on the negative binomial dispersion. The Galaxy Training Network RNA‑seq tutorials show how to generate such plots from DESeq2 output.

Making Visualizations Accessible

Accessibility is not an afterthought. Many readers (including yourself when the figure is printed in black and white or projected in a dim room) need high contrast, clear labeling, and color choices that accommodate color‑vision deficiency. The most common form is red‑green color blindness, which affects about 8% of men. Using red and green together for key comparisons can hide your findings from that audience.

Instead, use color‑blind‑friendly palettes, such as the “viridis” or “cividis” scale from the ggplot2 package in R. These are perceptually uniform and work in grayscale. For categorical variables, use shapes as well as colors. Always add axis labels in a legible font size (at least 12 pt for journal figures). Titles should be descriptive but not redundant.

One aspect often missed is the readability of text on screen for people using screen readers. Avoid tiny axis tick marks and use text that can be zoomed. If you embed figures in a web report, include alt text. The interdisciplinary article “Picture This: Enhancing Biology Education through Artistic Expression” [6] discusses how visual art principles can improve the clarity of biological figures, which directly applies to accessibility.

Exploration vs Presentation Figures

In exploratory data analysis (EDA), you generate many quick plots to understand structure, spot outliers, and decide on analysis steps. These plots may be messy, have overlapping points, or use default colors. That is fine. Presentation figures for a paper, grant, or talk need to be polished, annotated, and stripped of unnecessary elements.

Do not try to combine both in one figure. For example, a t‑SNE plot that you used to check for batch effects in EDA might show 20,000 points all overlapping. For a presentation, you would subset, color by a clear condition, and add a legend and possibly density contours. The difference is intentional design.

The tools you use for each phase often differ. In R, ggplot2 is great for both, but you can export a clean theme (theme_bw()) and then edit labels with labs(). The Bioconductor ecosystem provides many packages that return ggplot objects, making it easy to refine them for presentation.

One practical workflow is to create a separate “figures” folder after analysis, and write a script that generates every figure with consistent settings (font size, color palette, resolution). Then you can re‑render them quickly if data changes. The Galaxy Training Network offers workflows that end with a visual output, and you can adapt their approach to your own pipeline.

Common Mistakes

• Using a bar chart for continuous data when sample size is small. With 3 to 5 replicates per group, a bar with error bars can mislead because the error intervals overlap. Instead, show all points with a mean line.
• Zooming into a PCA plot to show separation. If you must zoom, also show a full‑scale version or a note that the axes are truncated. Better to use a different projection.
• Ignoring overplotting. With thousands of single‑cell data points, a scatter plot can become a black blob. Use transparency, density plots, or binning.
• Using pie charts. Biologists rarely have data that fits a pie chart well. Small differences in angles are hard to perceive.
• Forgetting to test for color‑blindness. Use a simulator (e.g., Color Oracle) before finalizing.

Limits and Uncertainty

No figure can capture all the nuance. A PCA plot discards variance not captured by the first two components. A heatmap depends on the distance metric and clustering method. The [Provable cluster‑preserving visualizations [8] paper shows that even careful methods can distort distances. Always state the method used and the parameters (e.g., “PCA on 500 most variable genes, using prcomp in R with centering and scaling”). When possible, show a measure of how much variance is explained (e.g., PC1: 23% variance). If your result depends on a transformation, check the sensitivity by trying another transformation or a non‑parametric method.

For data that includes many zeros, like single‑cell RNA‑seq, standard PCA can be misleading. Alternative methods like scran’s denoisePCA or zero‑inflated factor analysis exist, but each has its own assumptions. The Bioconductor single‑cell workflow documentation discusses these trade‑offs.

Uncertainty also applies to the figure’s interpretation. A small p‑value in a volcano plot does not guarantee biological importance. Always show effect sizes and consider adding a dotted line for your significance threshold so readers can see where your claims come from.

Frequently Asked Questions

1. Should I always use log transformation for sequencing data? Not always. If you are looking at very highly expressed genes, a raw scale may suffice, but for most RNA‑seq analyses, log2 or VST transformation is standard because it stabilizes variance across the expression range. Always check a density plot of raw vs transformed values.

2. What plot is best for showing many groups at once? A heatmap with a color scale can show expression values for hundreds of genes across multiple conditions. For comparing distributions across many groups, consider a violin plot with a boxplot overlay, but limit to about 10,20 groups before it becomes cluttered.

3. How do I choose between a PCA plot and a t‑SNE plot? PCA is linear and preserves global structure well, so it is good for checking batch effects and overall variance. t‑SNE better preserves local neighborhoods but can create false clusters. Use PCA first, then t‑SNE or UMAP with multiple perplexity values to see if clusters are robust.

4. Can I use Excel to make publication‑ready biological figures? Excel is fine for simple bar charts or scatter plots of very small datasets, but it lacks many features needed for biological data such as proper log axes, color‑blind palettes, and statistical overlays. Use R, Python, or a dedicated tool for serious work.

References and Further Reading

• NCBI Bookshelf , Free textbooks covering statistics, bioinformatics, and data visualization foundations.
• EMBL‑EBI Training , Practical courses on designing clear biological figures and interpreting omics data.
• Galaxy Training Network , Workflow‑based tutorials that include plotting steps for RNA‑seq and other data.
• Bioconductor , Open‑source software and vignettes for transformation, normalization, and visualization of genomic data.
• Sequence Read Archive , Public repository for raw sequencing data, useful for practicing visualization techniques.
• Picture This: Enhancing Biology Education through Artistic Expression , Integr Org Biol article on applying art principles to biological figures [6].
• Provable cluster‑preserving visualizations with curvature‑based stochastic neighbor embeddings , PNAS study on improved dimension reduction for biological data [8].
• MondrianMap: Navigating Gene Set Hierarchies with Multi‑Resolution Enrichment Maps , bioRxiv preprint describing a tool for clear pathway visualization [11].

Related Articles

• RNA Sequencing Analysis: From FASTQ Files to Biological Questions
• RNA-seq Quality Control: What to Check Before Differential Expression
• How to Plan a Bulk RNA-seq Differential Expression Study
• Single-Cell RNA-seq Workflow: A Practical Analysis Roadmap
• Single-Cell RNA-seq Quality Control: Cells, Genes, and Mitochondrial Reads