Skip to content

MA Plot vs Volcano Plot: What Each One Shows and When to Use It

Oak
Two glowing tidal pools on a reef at dusk, one a flat blue band, the other an orange cone of rising spores.

An MA plot and a volcano plot are two views of the same differential expression table, differing only in what goes on the y-axis and the x-axis. The MA plot shows log2 fold change against mean expression. The M axis stands for “minus” and the A axis for “average”. It answers whether your normalization is sound and where in the dynamic range the changes sit. The volcano plot shows log2 fold change against statistical significance, usually −log10 of the p-value or adjusted p-value, so it answers which genes survive multiple-testing correction and how large their effects are. We look at the MA plot first as a diagnostic. Only then do we look at the volcano plot as a result. Reversing that order is how people end up interpreting a normalization artifact as biology.

What MA stands for and where the axes come from

The M and A come from the original two-color microarray literature. For a gene measured in two channels with intensities R and G, M was defined as log2(R) − log2(G) and A as the average, (log2(R) + log2(G)) / 2. M is a difference of logs, which is the log ratio. A is the mean log intensity, a proxy for how abundantly the gene is expressed. The plot was originally drawn to expose intensity-dependent dye bias, which appeared as a curved trend in M across A rather than a flat band at zero.

In modern RNA-seq the same geometry survives with different inputs. In DESeq2 the x-axis is baseMean, the mean of normalized counts across all samples, plotted on a log scale. The y-axis is log2FoldChange for your chosen contrast. In limma-voom the equivalent is AveExpr against logFC. Nothing about the plot is specific to transcriptomics: the same axes work for proteomics intensities and for any abundance table where you compare two groups, which is why general omics platforms ship both plot types side by side.12

Reading an MA plot

Start with the shape of the cloud. A well-normalized two-group comparison gives a roughly symmetric band centered on M = 0, wide on the left where counts are low and tightening as you move right into well-expressed genes. That funnel is expected. Low-count genes have high relative Poisson noise, so their fold change estimates are unstable even when nothing is different.

Three departures from that shape matter. A vertical offset, where the whole band sits above or below zero, means the normalization assumption failed. Median-of-ratios normalization in DESeq2 and TMM in edgeR both assume most genes are unchanged. Break that assumption, for example with a global transcriptional shift or a dominant highly-expressed transcript soaking up library depth, and every fold change inherits the bias. A curved trend in M across A points at a composition or GC-content effect that a single size factor cannot absorb. Strong asymmetry, with many more up-regulated points than down, is sometimes real and sometimes the same normalization problem in a different guise.

Compositional effects deserve specific attention because sequencing data is compositional by construction. Your library size is fixed by the sequencer, so an increase in one transcript’s share forces decreases elsewhere. Fold changes are therefore relative to a reference frame set by your normalization, not absolute abundances, and treating them as absolute is a documented source of spurious results.3

The second thing to read is where the signal lives along A. Significant hits crowded entirely into the low-expression left edge are usually low-count genes that squeaked through, and they deserve suspicion. Hits spread across the dynamic range, including genes with baseMean in the thousands, are more likely to be reproducible. This is the question the MA plot answers that the volcano plot cannot, because the volcano plot has discarded the abundance axis.

One practical point on shrinkage. plotMA on a raw DESeq2 result draws unshrunken fold changes, which produces enormous spikes at the left edge. Run lfcShrink(dds, coef=2, type="apeglm") first. The apeglm estimator pulls low-information fold changes toward zero while leaving well-measured large effects intact, which makes the plot reflect the evidence and makes ranking by effect size meaningful.

res   <- results(dds, contrast = c("group", "treated", "control"))
resLFC <- lfcShrink(dds, coef = 2, type = "apeglm")
plotMA(resLFC, ylim = c(-4, 4), alpha = 0.05)

Reading a volcano plot

The volcano plot plots logFC on x and −log10(p) on y, so significance increases upward and effect size increases outward in both directions. The characteristic V shape follows from the test statistic. For a fixed variance, larger fold changes give smaller p-values, so the point cloud rises along two arms away from the origin.

We plot adjusted p-values, not raw ones, and we say which on the axis label. With 15,000 to 25,000 tested genes, an uncorrected p-value of 0.05 gives you hundreds of false positives by construction, and the fact that omics data are wide and shallow is the core statistical problem of the field.4 Benjamini-Hochberg control at 0.05 or 0.10 is the standard. Note that DESeq2 sets padj to NA for genes filtered by independent filtering, and those rows silently vanish from a naive ggplot call.

Three patterns in a volcano plot are worth naming. Points high on the y-axis with logFC near zero are small but very precisely measured effects, common with large sample sizes and often biologically uninteresting. Points far out on x with low y are large apparent effects with weak evidence, usually low-count genes or a single outlier sample driving the contrast. A flat volcano, with nothing rising, means either no effect or insufficient power. A volcano that is asymmetric and shifted sideways sends you back to the MA plot, because that is a normalization signature.

Both plots share one hard limitation: they display exactly one contrast. With three or more groups you need one plot per pairwise comparison, and reviewing them independently invites inconsistent conclusions. Interactive tools that link the same points across comparisons and across omics layers address this directly, and they matter more as you add proteomics and metabolomics to transcriptomics.1 The alternative is a dedicated multi-omics pipeline that keeps the contrasts organized for you.5

How we would use them together

Our sequence is fixed. First, plot MA on shrunken fold changes and confirm the band is centered and flat. If it is not, fix normalization before reading any gene names. Second, plot the volcano with adjusted p-values and identify the significant set. Third, go back to the MA plot and check that the significant set is not confined to the low-abundance edge. Fourth, take a handful of top hits and plot their normalized counts per sample, because both plot types hide whether a hit is driven by one sample.

Only then does enrichment analysis make sense. Gene set methods amplify whatever bias is in your ranking, so a normalization artifact that survives to the ranked list will emerge as a confident-looking pathway result. Multi-omics work makes the same point in reverse: when a signal is real, it tends to appear consistently across layers, and cross-layer agreement is stronger evidence than any single volcano plot.6

Neither plot is a clinical instrument. Differential expression from one person’s tissue or blood at one or two timepoints is a measurement. Mapping it to anything about your health requires a clinician who knows your history.

Questions people also ask

What does MA plot mean? M is the log ratio between two conditions, defined as log2(A) − log2(B). A is the average log expression, (log2(A) + log2(B)) / 2. The plot is M on the y-axis against A on the x-axis. The letters have nothing to do with anyone’s initials.

What is the difference between an MA plot and a volcano plot? Both use log2 fold change, but the MA plot pairs it with mean expression and the volcano plot pairs it with significance. The MA plot is a normalization and dynamic-range diagnostic. The volcano plot is a results summary that identifies your significant set.

What is the full name of the MA plot? It is commonly written out as a mean-difference plot, or a log-ratio versus mean-average plot. In limma the equivalent function is called plotMD, for mean-difference.

How do I interpret RNA-seq results? Confirm library sizes and sample clustering before testing, then check the MA plot for normalization bias, then read the volcano plot using adjusted p-values. Rank by shrunken effect size rather than by p-value alone, and inspect per-sample counts for your top hits. The general principle is that wide, shallow data demand multiple-testing control and independent confirmation.4

How do I interpret gene set enrichment analysis? GSEA tests whether a predefined gene set is concentrated at the top or bottom of your full ranked gene list, rather than testing an arbitrary significance cutoff. Report the normalized enrichment score alongside the false discovery rate, and treat enrichment as a hypothesis about coordinated behavior. Pathway databases overlap heavily, so several related sets frequently light up from the same handful of genes.2

Footnotes

  1. Jakob Willforss, Valentina Siino, Fredrik Levander. OmicLoupe: facilitating biological discovery by interactive exploration of multiple omic datasets and statistical comparisons. BMC Bioinformatics, 2021. https://doi.org/10.1186/s12859-021-04043-5 ↩ ↩2

  2. Murodzhon Akhmedov, Axel Martinelli, Roger Geiger, et al. Omics Playground: a comprehensive self-service platform for visualization, analytics and exploration of Big Omics Data. NAR Genomics and Bioinformatics, 2019. https://doi.org/10.1093/nargab/lqz019 ↩ ↩2

  3. Thomas P Quinn, Ionas Erb, Greg Gloor, et al. A field guide for the compositional analysis of any-omics data. GigaScience, 2019. https://doi.org/10.1093/gigascience/giz107 ↩

  4. Daniela Dunkler, Fátima Sánchez-Cabo, Georg Heinze. Statistical Analysis Principles for Omics Data. Methods in Molecular Biology, 2011. https://doi.org/10.1007/978-1-61779-027-0_5 ↩ ↩2

  5. Laura Madrid-Márquez, Cristina Rubio-Escudero, Beatriz Pontes, et al. MOMIC: A Multi-Omics Pipeline for Data Analysis, Integration and Interpretation. Applied Sciences, 2022. https://doi.org/10.3390/app12083987 ↩

  6. Marta Ballanti, Lorenzo Antonetti, Maria Mavilio, et al. Decreased circulating IPA levels identify subjects with metabolic comorbidities: A multi-omics study. Pharmacological Research, 2024. https://doi.org/10.1016/j.phrs.2024.107207 ↩

Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.