Figure Legend: What It Must Contain and How to Write One for a Journal

The components above hold everywhere. Where they go, and how much detail belongs in the legend rather than the Methods, changes by journal. Three current author guides show how far apart the rules sit:

Source: Nature formatting guide; eLife author guide; PLOS ONE figure guidelines (all accessed September 21, 2026)

Two more rules affect how you draft. Nature asks that text, including keys to symbols, go in the legend rather than on the figure itself where possible. The ICMJE Recommendations say the same about titles and detailed explanations, which belong in legends and not on the illustrations. Many medical journals follow ICMJE, so a plot crowded with explanatory text boxes is working against the house style there.

The practical consequence: pick the journal before you write the legends. Converting a Nature-style legend into an eLife-style one isn't a trim. It means pulling replicate structure, exclusion criteria and exact P values out of the Methods and back into every legend, which touches every figure in the paper and is an easy place to introduce a mismatch.

Here is a legend of the kind that turns up in first drafts. It's an illustration built for this page, not a real paper.

Figure 3. Drug X suppresses inflammation. (A) Representative western blot. (B) Quantification. n = 3. *P < 0.05.

Read it as a reviewer would:

• The title outruns the data. "Suppresses inflammation" is a claim about a biological process. The panels measure one protein in one cell type.
• Panel A names nothing. Which proteins, which cells, which treatment, which loading control? "Representative" of how many repeats?
• Panel B has no denominator. Quantification of which band, normalized to what, relative to which group?
• n = 3 has no unit. Three lanes on one blot and three independent mice are very different evidence.
• The statistics are half there. There are no error bars defined, no test is named, and the asterisk has a threshold but no stated comparison.
• Nothing says where the lanes came from. If the control lane was cut in from another membrane, Nature's policy expects the legend to say so.

The rewrite:

Figure 3. Drug X lowers IL-6 protein in LPS-stimulated mouse macrophages. (a) Immunoblot of IL-6 and β-actin (loading control) in bone-marrow-derived macrophages pretreated with vehicle (0.1% DMSO) or drug X (10 µM) for 6 h, then stimulated with LPS (100 ng/ml) for 4 h. The blot shown is one of the four biological replicates quantified in b; all lanes come from the same membrane, and uncropped images are in Supplementary Fig. 7. (b) IL-6 band intensity normalized to β-actin and expressed relative to vehicle. Bars show mean ± s.d.; each dot is cells from one mouse (n = 4 mice per group). Two-tailed unpaired t-test; *P < 0.05 versus vehicle.

That's about 110 words, well inside Nature's cap. It is also longer than it needs to be for Nature. Because Nature keeps methods detail out of legends when a Methods section exists, the doses and timings would move to Methods there, while the cell type, the comparison, the loading control, n, the error-bar definition and the test stay put. For eLife, you'd keep everything and add the exact P value.

Notice what changed in the title. It went from a claim about inflammation to a claim about one measured protein. That edit alone answers the first objection this legend would have drawn, before anyone reads the panels.

When you can't decide whether a detail belongs in the legend, sort it by who needs it to read the plot:

A useful test for the middle category: if a reader could misread the plot without the detail, it belongs in the legend at any journal that allows it.

Some figures get longer legends every draft because the legend is patching a problem in the figure. Think twice before adding another sentence if any of these is true:

• It needs more than 300 words to make the figure readable. At Nature that's over the cap. Anywhere else it usually means the figure carries too many panels; split it, or move the supporting panels to the supplement.
• It explains what the axis should say. "The y axis shows IL-6 relative to vehicle" is a relabeling job, not a legend sentence. Put the variable and its units on the axis.
• It tells readers which bars to compare. If the comparison only becomes visible after reading prose, redraw the plot so the two groups sit side by side.
• The title only stays true with a hedge. "Drug X may partly reduce inflammation" is a figure being asked to carry a claim it doesn't support. Fix the claim in the Results and abstract first, then retitle.

A longer legend can't repair these failures. It adds words and leaves the risk where it was.

In our pre-submission review work, legends get read against two things: the figure they describe and the Results text that cites them. The $39 Full Review runs a figure pass over the figure pages it detects in your manuscript, and the checks behind it map onto four named patterns:

The undefined error bar. Error bars are drawn, but the legend never says whether they show s.d., s.e.m. or a 95% confidence interval. The figure pass flags panels where that definition is missing, along with panels that state no n per group or report a P value without naming the test. For small groups it also asks for the individual data points to be shown, not just a bar.

The unexplained asterisk. Stars appear above bars with no threshold given, or with thresholds but no stated comparison. The review treats significance markers as undefined until the legend says what each one means.

The title that outruns the panels. This is the Figure 3 problem above. The figure pass reads each legend title against the evidence in the panels and flags titles that claim causation where the data show correlation, claim a mechanism the figure doesn't validate, or use "demonstrate" and "reveal" for indirect evidence. It then proposes a narrower title.

The legend that disagrees with the Results. The legend says n = 6 and the Results say five animals, or the P value in the text isn't the one in the figure. The review cross-checks quantitative values that appear in both places, including n, P values, fold changes and correlation coefficients, and flags the mismatches it finds. For western blots it also looks for a loading control, band quantification and uncropped blots in the supplement.

One guard runs in the other direction. Before the Full Review is delivered, an evidence audit checks every statement that a panel is unquantified or lacks statistics against that panel's legend. It exists so that a legend which does report the n and the test isn't wrongly called out for missing them.

The $99 Submission-Ready Dossier adds a separate vision critique of the rendered figure pages, checking axis labels and units, scale bars on microscopy, whether significance markers are defined, whether the color palette holds up for colorblind readers, resolution and cropping, and whether what the figure visibly shows matches what its caption says. Neither product is image forensics. For duplicated or spliced panels, use the steps in our guide to checking figures for image integrity.