Gains:
- Ability to choose basic graphic types of biology such as volcano plot, heat map, box/violin plot and PCA for the purpose.
- Ability to apply honesty principles such as axis starting from zero, definition of error bar, n information, and accessible palette
- Ability to avoid misleading graphs that hide the distribution, cut the axis, and beautify the data
A biological finding, no matter how important, cannot be understood if it is not visualized well. At the same time, a bad or misleading graph may misrepresent the data; This is both an ethical problem and a scientific mistake. In this unit, we will discuss the types of graphs most used in biology, how to produce them quickly with artificial intelligence, and what to pay attention to to ensure that the graph is honest.
AI produces broadcast quality plots in seconds with matplotlib/seaborn code; but it is your job to decide whether the graph accurately represents the data.
Basic types of graphs in biology
- Volcano plot: Comparison of effect size (log2FC) and significance (-log10 p) in differential expression. Shows changed genes at a glance.
- Heatmap: Expression pattern of multiple genes/samples in colors. It reveals patterns through clustering.
- Box/violin plot: Comparing the distribution of groups. It is more honest than the average bar because it shows the spread.
- PCA chart: Reducing high-dimensional data to 2 dimensions and showing clustering of samples (principal component analysis).
- Phylogenetic tree: Showing the evolutionary relationship of species/sequences through branching.
- Survival curve (Kaplan-Meier): Showing the probability of an event (e.g., death) over time.
Tip: Simple bar charts plotted on the mean are often misleading in biology because they obscure individual data points and distribution. If possible, opt for a box plot or “beeswarm” that also shows dots. If there are few data points, show them all.
Honest graphics principles
- Let the axis start from zero (especially in bar charts); the truncated axis exaggerates the difference.
- Specify what the error bar is: standard deviation, standard error, or confidence interval? These are very different.
- Show n: How many samples were obtained should be in the graph or in the title.
- Use a colorblind friendly palette (e.g. viridis); Do not rely on the red-green distinction.
- Do not beautify the data: Deleting the outlier, moving the axis, or only showing the beautiful repetition is scientific misconduct.
Step by step: producing a volcano chart
- Prepare the data: table with gen, log2FC, padj columns.
- Transformation: Calculate -log10(padj) for the y-axis.
- Set thresholds: |log2FC|>1 and padj<0.05.
- Color it: Up, down, meaningless three categories.
- Label: Name a few (not all) of the strongest genes.
- Title and axis: Clear label, n information, threshold description.
Copiable prompt templates
Role: You are a scientific visualization assistant. Task: Draw a volcano plot from my differential expression table (gen, log2FC, padj). Threshold: padj<0.05 and |log2FC|>1. Color the three categories and label the 10 most significant genes. Colorblind friendly palette, clear axis label, add n information to header. matplotlib, broadcast quality. Commented code.
Draw a heat map: rows are the 50 most variable genes, columns are samples. Perform row normalization with Z-score, add hierarchical clustering, use viridis palette. Show sample groups with a color stripe.
Explain whether the error bar in my graph should be standard deviation or standard error, depending on the purpose of the experiment. Explain the difference between the two and the consequences of choosing the wrong one.
Make this bar chart more honest: add individual data points on top, start axis at zero, show n. Available code: [code]
Weak prompt / Strong prompt
Weak: "Plot the yield."
Strong: "Have enzyme activity data for 4 groups (n=8 each). Draw a violin chart comparing groups; overlay individual data points for each group; show median line; start y-axis at zero; add unit to axis labels (nmol/min); use colorblind-friendly palette. Make sure the chart represents the distribution fairly."
Difference: The powerful prompt has the chart type, n, honesty policies and unit. The model produces a graphic that is informative, not misleading.
three mini cases
Case 1 — Truncated axis: In one presentation, a 3% difference between two groups was made to appear huge by squeezing the axis between 95-100. When the AI started the axis from scratch, it turned out that the difference was actually small. Lesson: axis manipulation misleads the viewer.
Case 2 — Hidden distribution: A bar chart showed two groups “significantly different”; but when the points were added, it was seen that the groups overlapped to a large extent, and the difference came from a few outlier points. When the model suggested adding points, the real story came out. Lesson: a chart that hides distribution lies.
Case 3 — Colorblind problem: A heat map was drawn with a red-green palette; Approximately 8% of the audience (colorblind males) could not see the distinction. When I switched to the Viridis palette, the chart became readable for everyone. Lesson: accessible pallet should be standard.
comparison chart
Purpose
suitable graphic
To be avoided
differential expression
volcano chart
raw p list
Group distribution
box/violin+dots
plain stick
multi gene pattern
Heat map (clustered)
long table
Sample clustering
PCA
—
time-event
Kaplan-Meier
average bar
Common mistakes
- Truncated axis: Exaggerating the difference.
- Hide the distribution: Show only the average bar.
- Not defining the error bar: SD/SE/GA confusion.
- Not specifying n: The reader cannot evaluate reliability.
- Inaccessible color: Excluding color blind people with a red-green palette.
- Data beautification: Selective representation is scientific misconduct.
Attention: Any arrangement of the graph that makes the data look different than it is (cutting the axis, hiding the outlier, selective repetition) is a violation of scientific integrity. AI can technically produce a “nice” chart; but it is your responsibility to make sure that chart represents the data fairly.
Phylogenetic tree and relationship graphs
A visualization specific to biology is the phylogenetic tree, which shows the evolutionary relationship of species or sequences. Branching pattern indicates relatedness, and branch lengths indicate the amount of change. The AI writes code that builds a tree from aligned sequences (e.g. with Biopython Phylo or ete3) and draws the tree in a readable format. However, there are two pitfalls in tree interpretation: ignoring branch length and looking only at branching, and not specifying the bootstrap (the value that indicates how reliable the branch is). A branch with low support is not sufficient to claim definitive kinship.
Draw a phylogenetic tree from aligned sequences. Show branch lengths to scale, add bootstrap support values to nodes, mark branches with low support below 70%. Approach low support branches with caution when interpreting the tree.
Table with graph: which when
Not every data requires graphics. Where exact numbers are important (e.g. exact values of several groups, statistical results) a well-organized table is superior to a graph. The chart shows the pattern and comparison; The table gives the exact value. When asked to artificial intelligence, "should this data be displayed as a graph or a table?" and asking for justification strengthens your presentation.
Finally, the chart needs to be self-explanatory. A reader should be able to understand what is shown just by looking at the graph and its title without reading the text: axes should be labeled and with units, groups should be defined, n should be specified, statistical significance should be marked. Ask the AI your chart “is it self-contained with its title and tags?” Having it inspected is a good final check.
Ready-to-publish chart: technical details
There are some technical details that take a graphic from looking good on screen to usable in broadcast, and AI can add these to the code. First, resolution: journals often require high resolution (300 DPI and above) or vector format (PDF, SVG); This ensures that the graphic does not blur in print. Second is the font size: a tag that can be read on the screen may not be read when it is reduced on the article page; axis and label points should be adjusted accordingly. Third, consistency: using the same color coding (e.g., control always gray, treatment always blue) across all graphs in the same study prevents the reader from redecoding each graph. You can add these details in one step by telling the artificial intelligence "make this graphic ready for publication: 300 DPI, vector output, large font size, consistent color palette".
Get my chart ready for publication: 300 DPI and also save as vector (PDF), increase axis and label sizes to print-readable size, apply consistent color palette across the entire run (control=gray, treatment=blue). Give commented code with matplotlib.
In summary
A good scientific graph displays data both clearly and honestly. Volcano plot, heat map, box/violin plot and PCA are basic tools of biology. The AI quickly writes the code for these graphs, but it's your job to follow principles of honesty like starting the axis at zero, showing the distribution, defining the error bar, specifying n, and the accessible palette. Beautiful graphics are not honest graphics.
Application task
With a data set you have (or a sample), have the AI produce two charts: a violin/box plot with data points showing the group distribution, and a volcano plot from a differential expression table. In both, start axis from zero (if appropriate), add n to title, use colorblind friendly palette. Then ask the model to produce a “misleading” and “honest” version of the same bar chart and explain the difference.
checklist
- [ ] I chose the chart type according to the data and purpose.
- [ ] I have properly initialized the axis from scratch.
- [ ] I showed the distribution/data points, not just the average.
- [ ] I specified what the error bar is (SD/SE/GA).
- I added [ ] n information to the chart.
- [ ] I used a colorblind friendly palette and did not beautify the data.