A heatmap is a picture not a conclusion

The eye finds blocks of colour. A heatmap is built to offer the eye blocks. Clustering reorders rows and columns so that similar things sit together, the colour scale decides what counts as extreme, and row scaling can make a timid gene look as loud as a violent one. None of that is the result of an expression study. The result is the contrast you specified and the table of estimates that came from it. This page is for anyone who has to defend a figure that started as a matrix. The study around the figure is from cells to a gene expression result.

Using the picture without handing it the conclusion

A heatmap is a legitimate way to show many genes across many samples at once. It earns its place in a paper when a reader can tell what the colours mean, what order the rows are in, and which statistical list, if any, supplied the genes. It overreaches when the caption says "the heatmap reveals" a mechanism the model never tested. Your decision, when you draw one, is whether this view answers a question the table already supports, or whether it is an exploration that must be labelled as exploration.

The numbers underneath should be traceable. A log-transformed count, a z-score, and a fold change against a calibrator are different matrices. qPCR fold changes and RNA-seq counts should not share a colour bar unless you have said how each was scaled. Gene identity should resolve on a browser such as Ensembl or the UCSC Genome Browser, so a symbol in the row label is the gene you think it is. Large public compendia, including the ENCODE project's encyclopedia of DNA elements in Science, show how expression is displayed at scale. They do not make your matrix self-explanatory.

What the colour is encoding

Pick a scale and write it on the legend in the unit you actually plotted. A diverging palette around zero suits a log fold change, where blue and red can mean down and up relative to a baseline you name. A sequential palette suits a non-negative abundance. Clipping the scale, so that everything above a cutoff becomes the same saturated colour, makes modest and huge effects look identical. Stretching the scale to the single wildest cell makes every other cell look pale. Both choices are display decisions. They should be deliberate, and they should not be the only place the magnitude exists.

Colour-blind-safe palettes and a legend that does not depend on hue alone are part of a figure a colleague can read. So is a statement of whether the matrix was centred. A missing legend is an incomplete method.

Clustering is a rearrangement, not a discovery procedure

Hierarchical clustering needs a distance and a linkage rule. Change either one and the dendrogram changes. The algorithm will find groups in noise if you ask it to return groups. The blocks you then see are partly the order it chose. That is helpful for spotting sample swaps and unwanted structure. It is a weak way to declare cell types, disease subtypes, or pathways, unless a separate analysis with a stated criterion supports the same split.

Column clustering is where batch effects go to look biological. Samples prepared on the same day resemble each other, the dendrogram glues them together, and a stripe of genes lights up under that glue. Put the preparation date, the operator and the kit lot under the columns, as urged in batch effects in an expression study. If the stripe matches the date, you have displayed a batch. If the columns stay in design order and the treatment groups still separate inside each date, the picture is doing harder work.

Row clustering groups genes with similar profiles. Those genes may share a regulator, or they may share a length, a GC content, or a sensitivity to degradation. The cluster is a hypothesis generator. The test of a named gene set is a different analysis, described in pathway lists and the multiple-testing problem.

Row scaling and the branch when picture and table disagree

Row scaling subtracts each gene's mean and divides by its spread across the samples in the figure. After that, every gene uses the full colour range. Patterns of shape become easy to see. Effect size becomes hard to see. A transcript that moves five percent and a transcript that moves fifty-fold can occupy the same red. If your claim is about magnitude, the unscaled log fold change, with a legend in that unit, is the picture that matches the claim. If your claim is about which samples share a profile, say that you scaled rows.

When a gene is brilliant on the heatmap and missing from the differential table, believe the inclusion rule of the table first. Check for an outlier sample, a clipped colour scale, and a multiple-testing threshold. Then decide whether the gene belongs in an exploratory panel. When a gene is in the table and invisible on the heatmap, the scale or the row scaling is hiding a real estimate. Fix the figure. Do not drop the gene from the result because the picture was quiet.

A useful branch is to draw two heatmaps from the same matrix: one with samples locked to the sample-sheet order and labelled by group and batch, and one clustered. If only the clustered version tells a story, the story may be the clustering. Show the locked version in the supplement so a reader can see the design.

Patterns the display can manufacture

An outlier sample with a failed library stretches the colour scale and paints the other samples as a flat field, or it forms its own cluster that looks like a subtype. Check library metrics before you name the cluster. A filter that keeps only the most variable genes guarantees a busy heatmap. Say the filter. Imputation of missing qPCR wells, if someone filled them with a group mean, manufactures agreement. Leave the well missing or show the imputation.

Combining RT-qPCR and RNA-seq on one heat map without a shared, declared scale invites a false visual consensus. A careful confirmation is a small table: the RNA-seq contrast, the qPCR fold change, the reference genes, and whether the direction agrees. That table is allowed to be dull. Dull agreement is a better result than a saturated picture.

Research figures are not diagnostic images

A heatmap in a research paper illustrates a stated analysis. It does not diagnose a donor, and it does not replace the ethics and biosafety framework that governed the samples. If the matrix includes human data, the display should follow the consent you actually have, including what identifiers sit under the columns. This page does not grant that consent.

A shipment stripe is still a batch

When samples arrive from another building or another city, the courier leg is a processing step. A clustered heatmap that groups every tube from one warm-season shipment is showing handling. Humidity, a delayed cold box, and a weekend in a receiving freezer can write that stripe. Label the shipment next to the preparation date. The differential table, built with that factor in mind, is where you ask whether any transcript effect survives it. The picture's job is to make the stripe obvious enough that nobody mistakes it for a subtype.

What to ask for when a figure is part of the work

If you request analysis or sequencing, say which contrast must be in the table, which gene filter is allowed, and whether heatmaps should lock sample order to the design. Those sentences keep the picture in its place. Catalogue classes for the sequencing side are in the genomics and sequencing catalogue. The sample path is the nucleic acid analysis pathway.

The mRNA sequencing enquiry reference and the differential expression analysis enquiry reference are enquiry references for discussing the measurement and the model. Put the contrast, the sample labels and the figure rules on the quote request, and ask whether a quotation is possible. A display can be discussed. The conclusion stays with the table you specified.