
Quick answer: A volcano plot puts fold-change on the x-axis and statistical significance on the y-axis so that the genes worth caring about — big change, small p-value — rise to the top corners of the picture. A heatmap shows expression across samples for a chosen gene set, usually after variance-stabilizing transformation and row scaling. Both are two-line jobs in R or Python once your table is in the right shape; the craft is in what you scale, what you threshold, and which genes you choose to display — because those three choices decide whether the figure is evidence or decoration.
Companion to the bulk RNA-seq analysis guide on this blog — this is the visualization step that sits at the end of that pipeline.
Table of Contents
What a volcano plot is actually encoding
Take your differential-expression table with three columns per gene: log2 fold-change, p-value (or adjusted p), and gene ID. The plot maps log2FC to x and −log10(p) to y. The shape explains the name: statistically strong but biologically trivial genes hug the top centre, large-but-noisy changes drift along the bottom corners, and the genes you want sit high and far out — the eruption.
Design decisions, in order of impact:
- Use adjusted p-values, not raw ones. With 20,000 tests, raw p below 0.05 yields ~1,000 false positives by chance. Rank by FDR-adjusted p (Benjamini–Hochberg, which DESeq2/edgeR/limma all report) and say so in the caption.
- Thresholds are conventions, not laws. The common default is |log2FC| > 1 (a twofold change) and adjusted p < 0.05. Draw them as reference lines and label them — a volcano without visible thresholds hides the analyst’s choices.
- Avalanche, plateau, or no signal? The classic three clusters (up, down, non-significant) only appear with a real, reasonably powered comparison. A flat cloud usually means low depth, high dispersion, or a weak contrast — a diagnostic, not a failed figure.
The ready-made route in R is EnhancedVolcano from Bioconductor:
library(EnhancedVolcano)
EnhancedVolcano(res, # DESeq2/edgeR result table
lab = res$gene,
x = 'log2FoldChange', y = 'pvalue',
pCutoff = 0.05, FCcutoff = 1,
title = 'Treatment vs control')Or draw it yourself with ggplot2 when you want full control — ggplot(res, aes(log2FoldChange, -log10(padj))) + geom_point(aes(colour = sig), alpha = 0.4) + geom_vline(xintercept = c(-1, 1), linetype = 2) + geom_hline(yintercept = -log10(0.05), linetype = 2). The same construction in matplotlib works for Python users: ax.scatter(lfc, -np.log10(padj)) with threshold lines added as ax.axvline.
Labeling without turning it into spaghetti
Only annotate the genes a reader would ask about: the top 5–10 by adjusted p, a short named list (marker genes for your condition), or manual geom_text_repel labels for a handful. Labeling every point below 0.05 makes the figure unreadable and argues nothing.
The heatmap: transformation first, aesthetics second
Raw counts are a trap for heatmaps. Their variance grows with their mean, so a clustering built on raw counts groups genes by abundance rather than by behaviour, and the highest-expression genes dominate the colour scale. The standard order of operations:
- Filter to expressed, interesting genes — typically the 500–2,000 most variable rows, or genes significant in the DE test. A heatmap of all 20,000 genes is noise texture.
- Transform for distance computation: variance-stabilizing transform or rlog (DESeq2), log2(CPM + 1) (edgeR/limma), or TPM for descriptive plots. Never cluster raw counts.
- Scale by row so each gene’s colour encodes its deviation from its own mean across samples. Without scaling, you are colouring by absolute abundance — which the volcano already told you.
- Order and annotate columns by sample group; add annotation bars for batch, condition, or patient. Grouped columns are what makes the pattern legible at a glance.
In R:
library(pheatmap)
top <- head(rownames(res[order(res$padj), ]), 500)
mat <- vst(dds, blind = FALSE)[top, ] # variance-stabilized counts
pheatmap(assay(mat), scale = 'row', # one row = one gene, z-scored
annotation_col = coldata, # condition/batch side bars
show_rownames = FALSE, clustering_method = 'ward.D2')Python equivalents exist (seaborn.clustermap with z_score=1 for row scaling) and behave the same way.
The mistakes that make reviewers twitch
- Significance by raw p-value. The single most common error in submitted figures.
- Clustering scaled-by-sample instead of by gene (or not scaling at all), producing an abundance map presented as a pattern map.
- Circular gene selection. Picking the top genes from the same test you are visualizing is fine for exploration — but if the heatmap is then sold as independent confirmation, that is double-dipping. Say it is descriptive.
- Colour scales that imply precision the data lacks. Diverging palettes (blue–white–red) fit z-scored data; sequential greys fit abundance. Rainbow palettes invent boundaries the numbers do not have.
- Samples that cluster by batch, not condition. If your columns order themselves by lane or collection date, that is the finding — fix the model (batch covariate in the DE formula) before drawing figures.
- No reproducible figure code. Save the script and the exact gene list next to the file; a figure you cannot regenerate is a figure you cannot defend.
From figure to report
Export at vector resolution (PDF or SVG for journals, PNG ≥ 300 dpi for slides) with the thresholds, transform, and gene-count visible in the caption. A caption that reads “Genes with |log2FC| > 1, FDR < 0.05, n = 1,204; heatmap of the 500 most variable genes after VST, row-scaled” answers every question a reviewer will ask in eighteen words.
Key takeaways
- Volcano: x is effect size, y is significance — always the adjusted p-value, with thresholds drawn and stated.
- A volcano’s silhouette is a QC readout: flat or sparse clouds point at power, depth, or contrast problems.
- Heatmaps need filtering, a variance-stabilizing transform, and row scaling — raw counts cluster by abundance, not biology.
- Show 500–2,000 genes, annotate sample groups, and label only the genes a reader would name anyway.
- If samples cluster by batch rather than condition, fix the model before making figures.
- Keep the figure script and gene list with the results so every panel can be regenerated on demand.
Challenge
Take a public differential-expression table (GEO or the supplement of any RNA-seq paper) and produce both figures in one script: a volcano with FDR-based thresholds and five labeled genes, plus a row-scaled heatmap of the top 500 variable genes with sample-group annotations. Then repeat the heatmap on raw counts and compare — the difference between “clusters by condition” and “clusters by abundance” is the lesson you will remember.
Want guided practice turning expression tables into defensible figures? Ampersand Academy offers one-to-one mentoring in bioinformatics — part of Mahadhi, a development and digital marketing company.
FAQ
Should the volcano plot use p-value or adjusted p-value?
Adjusted (FDR) p-value. With tens of thousands of tests, raw p-values deliver large false-positive sets at any conventional cutoff, and most journals now expect the adjusted value stated in the caption.
What are the standard fold-change and significance cutoffs?
The common default is absolute log2 fold-change above 1 (twofold) with adjusted p below 0.05. They are conventions: tighten them for small, well-powered studies and treat them as exploratory flags when power is low.
Why does my volcano plot have no genes in the top corners?
Usually weak statistical power rather than a plotting bug — check depth, dispersion, batch effects, and whether the contrast is biologically real. A cloud hugging the axis can also come from using raw p-values on a very noisy table.
Should I plot raw counts or normalized counts in a heatmap?
Neither directly: use a variance-stabilizing transform (VST, rlog) or log2 CPM. Raw-count heatmaps cluster genes by abundance because variance rises with the mean, which drowns the condition pattern.
Do I need to scale rows in the heatmap?
Row scaling converts each gene’s values to z-scores relative to its own mean, so colour shows per-gene change instead of absolute expression. For comparing patterns across genes it is essentially required; for showing absolute abundance it is wrong.
How many genes should the heatmap include?
Typically 500 to 2,000 — the most variable rows or the significant set from your test. Beyond a few thousand rows the plot becomes a texture, and clustering stability on noisy low-expression genes adds nothing.
Is it circular to select genes by significance and then show them in a heatmap?
Not for exploration or illustration, as long as you describe it as such. It becomes a problem only if the same selection is presented as independent evidence for the result that produced it.
Related tutorials
- Batch Effect Correction in Omics Data: ComBat, ComBat-Seq and When Not to Correct — why the same genes keep dominating your plots across batches
Last updated on · Written by Dinesh Kumar R

