Data Visualization Best Practices for Computational Biologists
By Dr. Zubair Khalid, DVM, MS, PhD ·
Quick Answer
- Choose a visualization type based on your data structure and biological question, not on default software settings or personal habit.
- Match the figure format to the intended journal or presentation medium before you begin analysis, because resizing later distorts labels and point density.
- A clear figure requires explicit axis labels, a readable legend, and a caption that states the sample size and statistical test used.
Why Visualization Quality Matters in Computational Biology
Computational biology produces large datasets from genomics, proteomics, and other high-throughput technologies. The figures you create are the primary way reviewers, collaborators, and hiring committees evaluate your work. A poorly designed plot can obscure a real biological signal, while a well-designed one can make a complex result immediately interpretable.
The U.S. Bureau of Labor Statistics describes the broader life-science workforce as one that increasingly relies on data analysis and quantitative reasoning across research settings. The Occupational Outlook Handbook notes that many biological science occupations require a master's degree or Ph.D. for research positions, and that these roles involve designing studies, analyzing data, and communicating findings. Your ability to produce publication-quality figures is a core professional skill, not an afterthought.
The U.S. Department of Labor's O*NET OnLine lists tasks for biological technicians and related occupations that include recording data, analyzing results, and preparing reports. These tasks depend on your capacity to translate raw output into a visual form that others can verify and reuse. A figure that cannot be read accurately is a failed experiment in communication.
Core Principles of Effective Biological Data Visualization
Know Your Data Structure Before You Plot
The first decision is not which chart to make. The first decision is what kind of data you hold. Biological data falls into a few broad categories, and each category has a natural visual form.
Continuous data, such as gene expression levels, protein abundance, or time-series measurements, are best shown with scatter plots, line plots, or histograms. Categorical data, such as treatment groups, genotypes, or tissue types, are best shown with bar plots, box plots, or strip plots. Count data, such as read counts or variant calls, often require careful normalization before any visualization is meaningful.
A common error is to treat all data as if it were continuous and normally distributed. Gene expression counts, for example, are often skewed and require a log transformation before a heatmap or scatter plot becomes interpretable. If you skip this step, your figure will show a few extreme values and compress the rest of the data into an unreadable mass.
Match the Plot Type to the Question
Different biological questions require different visual logic. A scatter plot is the right choice when you want to show the relationship between two continuous variables, such as gene expression in two conditions. A volcano plot is the standard for showing both the magnitude of change and the statistical significance of many genes at once. A heatmap is the right choice when you want to show a matrix of values, such as expression levels across many genes and many samples.
The decision framework below is a practical starting point. It is not a rigid rule, but it will help you avoid the most common mismatches between data and visualization.
| Data Question | Recommended Plot Type | What It Shows | Common Use Case |
|---|---|---|---|
| How do two continuous variables relate? | Scatter plot | Direction and strength of association | Gene expression in two conditions |
| Which genes change most between conditions? | Volcano plot | Fold change and statistical significance | Differential expression analysis |
| How does a value vary across many samples? | Heatmap | Patterns and clusters in a matrix | Sample-to-sample comparison |
| How does a value change over time? | Line plot | Trend and trajectory | Time-course experiments |
| How is a single variable distributed? | Histogram or density plot | Shape of the distribution | Quality control of read counts |
Use Color With Restraint and Purpose
Color is a powerful tool, but it is also the most common source of confusion. A heatmap that uses a rainbow palette can obscure differences between adjacent values because the human eye does not perceive color transitions as evenly spaced. A better choice is a sequential palette that goes from light to dark, or a diverging palette that goes from one color through a neutral midpoint to another.
Color is also a problem for readers with color vision deficiency. A figure that relies on red and green to distinguish groups is unreadable for a substantial portion of the population. Use a palette that is colorblind-safe, or add a second encoding such as shape or line type.
The National Institutes of Health Office of Intramural Training and Education provides guidance on scientific communication and career development for researchers. The Office of Intramural Training and Education notes that clear communication is a professional skill that researchers are expected to develop. A figure that cannot be read by all members of your audience is a communication failure.
Label Everything Explicitly
Every axis must have a label that states the variable and the unit. A plot that shows "Expression" on the y-axis is incomplete. The label should read "Expression (log2 counts per million)" or "Protein abundance (relative units)." The same rule applies to legends, which must state what each color, shape, or line represents.
The caption is part of the figure. It must state the biological sample, the number of replicates, the statistical test used, and the meaning of any error bars. A caption that says "Gene expression in treated and untreated cells" is not sufficient. A complete caption says "Gene expression in three biological replicates of treated and untreated cells, measured by RNA-seq. Error bars show the standard error of the mean. Significance was assessed by a two-sided t-test."
A Practical Workflow for Creating Publication-Quality Figures
Step 1: Define the Biological Question
Write down the exact question you want the figure to answer. If you cannot state the question in one sentence, you are not ready to plot. The question determines the plot type, the axes, and the statistical summary.
Step 2: Inspect the Raw Data
Before any transformation or normalization, look at the raw data. Check the range, the distribution, and the number of missing values. A quick histogram of each variable will reveal outliers and skew. This step is the quality control that prevents a misleading figure later.
Step 3: Choose the Plot Type
Use the decision table above to select the plot type. If you are comparing many genes across many samples, a heatmap is the natural choice. If you are showing a single relationship, a scatter plot is simpler and more honest.
Step 4: Apply the Correct Transformation
Log transformation is common for expression data. Normalization is often required before comparing samples. The transformation must be stated in the caption. A figure that shows transformed data without stating the transformation is misleading.
Step 5: Build the Figure With a Reproducible Script
Write the plotting code in a script, not in a point-and-click interface. A script allows you to regenerate the figure when the data changes and to share the exact method with others. The script is part of the scientific record.
Step 6: Review the Figure for Common Errors
Check the axis labels, the legend, the color palette, and the point density. If the figure has too many points to read, consider a density plot or a sample of the data. If the figure has too few points to support a conclusion, the figure is not the problem.
Common Failure Patterns and How to Avoid Them
The Overplotted Scatter Plot
A scatter plot with tens of thousands of points is a black blob. The relationship between the variables is invisible because the points overlap completely. The solution is to use a density plot, a hexbin plot, or to sample the data. A density plot shows the distribution of points in a region, which is often more informative than the individual points.
The Heatmap With the Wrong Color Scale
A heatmap with a color scale that does not match the data distribution will hide the structure. If most values are near zero and a few are large, a linear color scale will make the heatmap look mostly one color. A log scale or a quantile-based scale will reveal the structure.
The Bar Plot With Error Bars That Overlap
A bar plot with error bars that overlap between groups is a common source of misinterpretation. The reader may conclude that the groups are different when the overlap means the difference is not statistically significant. A better choice is a box plot or a violin plot, which shows the full distribution of the data.
The Line Plot With Too Many Lines
A line plot with dozens of lines is unreadable. The reader cannot distinguish the lines or follow a single trajectory. The solution is to show a subset of lines, to show the mean and a confidence band, or to use a heatmap instead.
Observations and Measurements That Support Good Visualization
Record the Data Provenance
Every figure should be traceable to the exact data file and the exact script that produced it. This record is essential for reproducibility and for answering reviewer questions. The record should include the date, the software version, and the parameters used.
Record the Transformation and Normalization
The transformation and normalization steps are part of the data record. If you log-transform the data, record the base of the logarithm. If you normalize to a reference sample, record the reference. This record allows another researcher to reproduce your figure exactly.
Record the Statistical Method
The statistical method used to compute significance must be recorded. The method determines the interpretation of the figure. A figure that shows a significance threshold without stating the method is incomplete.
Common Failure Patterns in Computational Biology Figures
Failure to State the Sample Size
A figure that does not state the number of samples is a common failure. The reader cannot judge the reliability of the result. The sample size must be in the caption or in the figure itself.
Failure to State the Error Bar Definition
Error bars can represent the standard deviation, the standard error, or a confidence interval. These are different quantities. The figure must state which one is shown. A figure that shows error bars without a definition is ambiguous.
Failure to Use a Consistent Color Scheme
A figure that uses different colors for the same group across different panels is confusing. The reader must re-learn the color scheme for each panel. Use a consistent color scheme across all panels in a figure and across all figures in a paper.
Failure to Check the Figure at the Final Size
A figure that looks clear on a large screen may be unreadable when printed at a small size. The font size, the point size, and the line width must be legible at the final size. Check the figure at the size it will appear in the journal or presentation.
Professional Development and Career Context
Visualization Skills in the Job Market
The ability to create clear figures is a skill that employers look for in computational biology candidates. The U.S. Bureau of Labor Statistics notes that biological and medical scientists often work in research and development settings, and that their work involves analyzing data and presenting findings. The Occupational Outlook Handbook describes these roles as requiring strong analytical and communication skills.
The U.S. Department of Labor's O*NET OnLine lists the tasks and skills for biological and medical scientists. These tasks include analyzing data, preparing reports, and presenting findings. A candidate who can show a portfolio of clear, reproducible figures has a distinct advantage in the job market.
Training and Professional Development
The National Institutes of Health Office of Intramural Training and Education provides training for researchers at all career stages. The Office of Intramural Training and Education offers workshops and resources on communication, data analysis, and career development. These resources are a practical way to build visualization skills.
The American Society for Microbiology provides career resources for professionals in the life sciences. These resources include professional development, networking, and training opportunities. A computational biologist who is a member of a professional society can access workshops and webinars on data visualization.
Federal Research Careers
The National Science Foundation describes careers at NSF that involve scientific, technical, and administrative work. These roles often require the ability to analyze data and communicate findings to a broad audience. A computational biologist who can create clear figures is well suited to these roles.
Limitations and Professional Escalation Criteria
When to Seek Help
If you cannot make a figure that clearly shows the pattern in your data, the problem may not be the visualization. The problem may be the data or the analysis. A figure that shows no pattern may be an honest result, not a failure of visualization.
If you are unsure about the correct statistical method, seek help from a statistician or a colleague with statistical expertise. A figure that uses the wrong statistical method is misleading, even if the figure is visually clear.
If you are preparing a figure for a regulatory submission or a clinical trial, the requirements are different from a research paper. The requirements are jurisdiction-specific and must be confirmed with the relevant authority. The sources in this article do not provide regulatory guidance for clinical or regulatory submissions.
Professional Escalation Criteria
If you are a student or a junior researcher, and your supervisor asks you to create a figure that you believe is misleading, you should discuss the issue with your supervisor. If the issue is not resolved, you should escalate to a more senior member of the laboratory or to the institutional research integrity office.
If you are a laboratory staff member and you are asked to create a figure that does not match the underlying data, you should refuse to do so and document your refusal. This is a matter of research integrity.
A Decision Framework for Multidimensional Biological Data
When Simple Plot Types Fail
The plot selection table in the previous section works well for data with one or two dimensions. Many computational biology experiments, however, produce data that is inherently multidimensional. A single gene expression measurement may have a value, a time point, a treatment condition, and a batch identifier. A proteomics experiment may measure thousands of proteins across dozens of samples with multiple technical replicates. When your data has more than two dimensions, the simple scatter plot or bar plot becomes insufficient.
The U.S. Department of Labor's O*NET OnLine lists complex data analysis and interpretation as core tasks for biological scientists. These tasks require a framework for deciding how to reduce dimensionality without losing the biological signal. The framework below is a practical method for making that decision.
Step 1: Count the Dimensions You Must Show
Write down every variable that must be visible in the figure for the reader to understand the result. A typical list includes the primary measurement, the grouping variable, the time or condition variable, and the replicate structure. If the list has more than three items, you cannot show all of them in a single standard plot. You must either reduce the dimensionality or split the figure into panels.
A common error is to try to encode too many variables into a single plot. A scatter plot that uses color for one variable, shape for another, and size for a third is difficult to read. The human visual system can track color and position well, but shape and size are weaker encodings. If you need more than three variables, a faceted plot or a small multiple is a better choice than a single crowded plot.
Step 2: Decide Whether to Reduce or Preserve Dimensionality
If the data has many variables but the biological question is about a single relationship, reduce the data. Principal component analysis or t-distributed stochastic neighbor embedding can project high-dimensional data into two dimensions for visualization. These methods are appropriate when the goal is to show overall structure or clustering.
If the biological question is about the relationship between specific variables, preserve the dimensionality and use a faceted plot. A faceted plot shows the same plot type for each level of a grouping variable. This approach is more honest than a single plot that tries to encode everything at once.
The National Institutes of Health Office of Intramural Training and Education provides training resources that emphasize the importance of clear data presentation in research communication. The choice between dimensionality reduction and faceting is a communication decision that affects how reviewers interpret your results.
Step 3: Match the Reduction Method to the Question
A principal component analysis plot is the standard choice for showing overall sample similarity. The axes are the principal components, and the percentage of variance explained by each component must be stated in the axis label. A t-distributed stochastic neighbor embedding plot is a better choice for showing local structure, but it does not preserve distances or variance. The choice between these methods depends on whether the reader needs to see global or local structure.
A UMAP plot is another option for showing sample clusters. The parameters used to generate the plot must be recorded and stated in the caption. These methods are not interchangeable, and the choice must be justified by the biological question.
Step 4: Use a Faceted Plot for Multidimensional Comparisons
A faceted plot is a grid of smaller plots, each showing the same data structure for a different level of a grouping variable. This approach is effective for comparing many conditions without overloading a single plot. The axis limits must be the same across all facets, or the reader will compare the wrong things.
A faceted plot is also the right choice when the data has a time variable and a treatment variable. Each facet shows the time course for one treatment, and the reader can compare the trajectories across facets. The color scheme must be consistent across all facets.
Step 5: Record the Reduction Parameters
Every dimensionality reduction method has parameters that affect the result. The number of components, the perplexity, the number of neighbors, and the random seed all change the output. These parameters must be recorded in the script and stated in the caption. A figure that shows a principal component plot without stating the number of components is incomplete.
The U.S. Bureau of Labor Statistics describes the work of biological scientists as involving the analysis of complex data and the communication of findings. The Occupational Outlook Handbook notes that these roles require attention to detail and the ability to document methods. The same standard applies to the parameters of a dimensionality reduction.
A Comparison of Dimensionality Reduction Methods
| Method | What It Shows | When to Use It | What to Record |
|---|---|---|---|
| Principal component analysis | Global variance structure | When the goal is to show overall sample similarity | Variance explained by each component |
| t-distributed stochastic neighbor embedding | Local structure and clusters | When the goal is to show clusters in high-dimensional data | Perplexity, number of iterations, random seed |
| UMAP | Local and global structure | When the goal is to show clusters with a faster algorithm | Number of neighbors, minimum distance, random seed |
Common Failure Patterns in Multidimensional Visualization
The Principal Component Plot With No Variance Label
A principal component plot that does not state the percentage of variance explained by each component is a common failure. The reader cannot judge how much of the data structure is visible in the plot. The axis label must state the variance, for example "PC1 (42 percent of variance)."
The t-SNE Plot With No Parameter Record
A t-distributed stochastic neighbor embedding plot is not reproducible without the parameters. The perplexity and the random seed change the output. A figure that does not state these parameters cannot be verified by a reviewer.
The Faceted Plot With Different Axis Limits
A faceted plot with different axis limits for each facet is a common source of misinterpretation. The reader compares the shapes of the plots, not the actual values. The axis limits must be the same across all facets.
The Overloaded Single Plot
A single plot that tries to encode too many variables is a common failure. The reader cannot distinguish the groups or follow the pattern. The solution is to use a faceted plot or to reduce the dimensionality.
Records and Measurements for Multidimensional Figures
Record the Data File and Script
The script that produces the figure must be saved with the data file. The script must state the software version and the parameters used. This record allows another researcher to reproduce the figure exactly.
Record the Reduction Parameters
The parameters of the dimensionality reduction must be recorded. The number of components, the perplexity, the number of neighbors, and the random seed are all part of the record. A figure that does not have this record is not reproducible.
Record the Axis Limits and Color Scale
The axis limits and the color scale must be recorded. The axis limits determine the interpretation of the plot. The color scale must be stated in the caption.
Professional Escalation Criteria
If you are unsure whether a dimensionality reduction is appropriate for your data, seek help from a statistician or a colleague with expertise in multivariate analysis. A figure that uses the wrong reduction method is misleading, even if the figure is visually clear.
If you are preparing a figure for a regulatory submission or a clinical trial, the requirements for dimensionality reduction are different from a research paper. The requirements are jurisdiction-specific and must be confirmed with the relevant authority. The sources in this article do not provide regulatory guidance for clinical or regulatory submissions.
Frequently Asked Questions
What is the most common visualization mistake in computational biology?
The most common mistake is choosing a plot type that does not match the data structure. A bar plot for continuous data, a scatter plot for categorical data, or a heatmap for data that is not a matrix are all common errors. The decision table in this article is a practical starting point.
How do I choose between a scatter plot and a volcano plot?
A scatter plot shows the relationship between two continuous variables. A volcano plot shows the relationship between the magnitude of change and the statistical significance of that change. Use a volcano plot for differential expression analysis, and use a scatter plot for a direct comparison of two variables.
What is the best color palette for a heatmap?
A sequential palette that goes from light to dark is a good choice for a heatmap. A diverging palette is a good choice when the data has a meaningful midpoint. Avoid rainbow palettes because they are not perceptually uniform and are not accessible to readers with color vision deficiency.
How do I make a figure reproducible?
Write the plotting code in a script, and save the script with the data file. The script should be version-controlled and should state the software version and the parameters used. A reproducible figure is one that can be regenerated from the raw data with the script.
What should I do if my figure is too crowded to read?
If a scatter plot has too many points, use a density plot or a hexbin plot. If a line plot has too many lines, show a subset or show the mean and a confidence band. If a heatmap has too many rows, cluster the rows and show the clusters.
What is the difference between a box plot and a violin plot?
A box plot shows the median, the quartiles, and the outliers. A violin plot shows the full distribution of the data. A violin plot is more informative when the distribution is not symmetric, but it is also more complex to read.
How do I state the error bars in a figure caption?
The caption must state what the error bars represent. The options are the standard deviation, the standard error of the mean, or a confidence interval. The caption must also state the number of replicates.
What is the role of a figure in a computational biology paper?
A figure is the primary way to communicate a result. A figure must be clear, reproducible, and honest. A figure that is misleading is a scientific error, regardless of the quality of the underlying data.
Related Articles
- Data Visualization for Biological Research: Choosing Figures That Preserve Meaning
- Bioinformatics vs Computational Biology: Choosing Your Career Path
- Environmental Biology vs Environmental Science: Choosing the Right Career Path
- Bioinformatics vs Computational Biology: Which Degree Path Is Right for You?
- Computational Biology Jobs
References and Further Reading
- Occupational Outlook Handbook. U.S. Bureau of Labor Statistics.
- O*NET OnLine. U.S. Department of Labor.
- Office of Intramural Training and Education. National Institutes of Health.
- Careers at NSF. U.S. National Science Foundation.
- Career Services. American Society for Microbiology.
- AAVMC Career Resources. Association of American Veterinary Medical Colleges.
- NCBI Literature Resources. National Center for Biotechnology Information.
- PubMed. National Library of Medicine.
This article is educational and does not replace institutional policy, professional advice, or applicable safety and regulatory requirements.