Reading the MA Plot in DESeq2
An MA plot in DESeq2 is a scatter plot of log2 fold change against the mean of normalized counts across all samples. The y axis is “M” for minus, since a log ratio is a difference of logs. The x axis is “A” for average and is drawn on a log scale. Each point is one gene. Points colored by the significance threshold are those whose adjusted p-value falls below alpha. The plot answers two questions at a glance: how large are the effects you detected, and at what expression level do they live. If the cloud is centered on zero and symmetric, and it narrows as counts increase, your normalization and dispersion fit are behaving. If it is off-center, fanned, or lopsided, the plot is telling you about a problem upstream of the statistics.
What the axes contain
The y axis is the log2 fold change for one coefficient, not a global measure of change. plotMA pulls it from the log2FoldChange column of a DESeqResults object, which corresponds to whichever contrast you asked results() for. Check with resultsNames(dds) before you interpret anything. A plot of the wrong coefficient looks perfectly reasonable and means nothing.
The x axis is baseMean, the average of counts divided by size factors, across every sample in the object. It is not group-specific. A gene expressed at 400 counts in treated samples and 0 in controls sits near 200 and not 400. This matters because the width of the cloud is a function of counting noise. Genes with low baseMean have few reads, so their fold-change estimates are poorly constrained and scatter widely. That funnel shape is the expected signature of Poisson-plus-overdispersion sampling. It is not a bug.
Shrink the fold changes before you plot
The maximum likelihood fold changes that come straight out of results() are noisy at low counts. A gene with 3 reads in one group and 0 in the other can report a log2 fold change of 4 or more with no real evidence behind it. Those genes dominate the left edge of the plot and make the effect-size distribution look far more dramatic than it is.
We always plot shrunken estimates. The empirical Bayes shrinkage in lfcShrink pulls poorly-supported fold changes toward zero while leaving well-measured large effects nearly untouched. The apeglm estimator is our default because it uses a heavy-tailed prior and so shrinks noise hard without flattening genuine large effects.
library(DESeq2); library(apeglm)
dds <- DESeqDataSetFromMatrix(counts, colData = meta, design = ~ batch + condition)
dds <- DESeq(dds)
resultsNames(dds)
# [1] "Intercept" "batch_B_vs_A" "condition_treated_vs_control"
res_mle <- results(dds, name = "condition_treated_vs_control", alpha = 0.05)
res_ape <- lfcShrink(dds, coef = "condition_treated_vs_control", type = "apeglm")
par(mfrow = c(1, 2))
plotMA(res_mle, ylim = c(-5, 5), alpha = 0.05, main = "MLE")
plotMA(res_ape, ylim = c(-5, 5), alpha = 0.05, main = "apeglm")
Two details about that call. First, alpha in plotMA defaults to 0.1 and only controls coloring, so set it to whatever you used in results() or the colors will not match your gene list. Second, lfcShrink with type = "apeglm" keeps the p-values and adjusted p-values from the underlying Wald test, so significance calls do not change when you switch. Use type = "ashr" instead if you need shrinkage for an arbitrary contrast specified through the contrast argument, which apeglm does not support in the same way.
Reading the shapes
A healthy plot has three features. The bulk of genes sits on the horizontal line at zero. The cloud is symmetric above and below that line. Its width shrinks monotonically from left to right, and the colored points on the left are pushed to larger absolute fold changes than the colored points on the right, because that is what it takes to clear significance at low counts.
Now the deviations. A cloud whose center sits above or below zero across the whole x range points at normalization. DESeq2 uses median-of-ratios size factors, which assume most genes are not changing. Break that assumption, through a global shift in transcriptional output or a heavy contamination of one library with rRNA or globin, and the whole cloud shifts. Inspect sizeFactors(dds) and compare it to library size. If a sample’s size factor and column sum disagree badly, find out why before trusting a single fold change.
A cloud that tilts, wide at high expression rather than narrow, usually means a technical covariate is tracking expression level. GC content and transcript length effects behave this way. You can absorb them with sample-specific and gene-specific offsets through normalizationFactors(dds), estimated by a package such as cqn or EDASeq.
Hollow triangles at the top and bottom edge are genes whose fold change falls outside your ylim, drawn pointing in the direction they exceed. If you see many of them, widen ylim rather than assume they are all real. After apeglm shrinkage, most disappear.
Grey points that you expected to be colored are usually genes with padj = NA. There are two causes. Independent filtering removes low-count genes to gain power, and results() sets their adjusted p-values to NA. Check metadata(res)$filterThreshold. Separately, a gene with an extreme Cook’s distance in one sample gets its raw p-value set to NA when you have at least three replicates per group, unless the count is replaced. Both mechanisms are visible in the plot as unlabeled points sitting well off the zero line.
When only a handful of genes come out significant
This is the most common complaint, and the MA plot tells you which of two situations you are in. If the cloud is wide at all expression levels and almost nothing is colored, your within-group variability is large relative to the between-group difference. Look at plotDispEsts(dds) and at a PCA of vst(dds) transformed counts. A batch or subject effect that is not in the design will inflate dispersion and flatten your power. Adding the covariate to the design formula is the fix, if the design is not confounded.
If the cloud is tight and centered but the colored points are few and small, the effect genuinely may be small. Small effect sizes are normal in human tissue comparisons with modest sample sizes, and studies that go on to nominate candidate markers typically combine differential expression with other layers rather than relying on a single contrast 1. Multi-omics work of this kind routinely pairs transcript-level changes with protein or metabolite measurements to separate signal from noise 2. Integration across layers is also how these studies arrive at prognostically useful signatures rather than long unfiltered gene lists 3.
Customizing the plot
DESeq2::plotMA is a thin wrapper and gets cramped fast. For anything you intend to show someone else, build it yourself from the results data frame.
library(ggplot2)
df <- as.data.frame(res_ape)
df$sig <- !is.na(df$padj) & df$padj < 0.05
ggplot(df, aes(baseMean, log2FoldChange, color = sig)) +
geom_point(size = 0.5, alpha = 0.5) +
scale_x_log10() +
geom_hline(yintercept = 0, linewidth = 0.3) +
scale_color_manual(values = c(`FALSE` = "grey60", `TRUE` = "#0b5394")) +
coord_cartesian(ylim = c(-5, 5)) + theme_bw()
Add ggrepel::geom_text_repel on the top genes by padj for labels. If you prefer to click on points in the base plot, the DESeq2 vignette’s idx <- identify(res$baseMean, res$log2FoldChange) still works and returns row indices you can pass back into rownames(res)[idx].
One interpretive limit is worth stating plainly. An MA plot describes a contrast between groups of samples. It does not tell you what any single gene means for a person, and a differential expression result is not a clinical finding. Cross-tissue and cross-cohort work shows how much context a transcript needs before it carries any interpretation 4. If a result bears on your health, take it to a clinician.
Questions people also ask
Does the MA plot replace a volcano plot? They show different things. The MA plot puts effect size against expression level, which is what you need to judge normalization and count-dependent noise. A volcano plot puts effect size against significance, which is better for picking genes to follow up. We look at both.
Should I use apeglm or ashr? Use apeglm when your comparison is a single coefficient in resultsNames(dds). Use ashr when you need a custom contrast or a linear combination of coefficients. Avoid type = "normal" for new work, since its prior shrinks large true effects more than the alternatives.
Why are my low-count genes fanning out so far? Because a fold change estimated from a handful of reads has almost no information in it. That is the expected behavior of the maximum likelihood estimate, and it is the reason to plot shrunken fold changes instead.
What baseMean cutoff should I apply? Prefilter only lightly, for example keep <- rowSums(counts(dds) >= 10) >= smallestGroupSize, and let independent filtering inside results() handle the rest. Hard cutoffs chosen after looking at the MA plot bias your gene list.
Can I make an MA plot for proteomics or metabolomics data? Yes, and it is a useful diagnostic there too, though you plot log intensity rather than normalized counts on the x axis and use a model appropriate to the assay. Integrated studies commonly apply the same visual check to each layer before combining them 5.
Footnotes
-
Naiyar Iqbal, Pradeep Kumar. Integrated COVID-19 Predictor: Differential expression analysis to reveal potential biomarkers and prediction of coronavirus using RNA-Seq profile data. Computers in Biology and Medicine, 2022. https://doi.org/10.1016/j.compbiomed.2022.105684 ↩
-
Mattias Backman, Florian Flenkenthaler, Andreas Blutke, et al. Multi-Omics Insights into Functional Alterations of the Liver in Insulin-Deficient Diabetes Mellitus. SSRN Electronic Journal, 2018. https://doi.org/10.2139/ssrn.3307901 ↩
-
Joseph Pierre Aboumsallem, Canxia Shi, Sanne De Wit, et al. Multi-omics analyses identify molecular signatures with prognostic values in different heart failure aetiologies. Journal of Molecular and Cellular Cardiology, 2023. https://doi.org/10.1016/j.yjmcc.2022.12.001 ↩
-
Luca Massimino, Alberto Barchi, Francesco Vito Mandarino, et al. A multi-omic analysis reveals the esophageal dysbiosis as the predominant trait of eosinophilic esophagitis. Journal of Translational Medicine, 2023. https://doi.org/10.1186/s12967-023-03898-x ↩
-
Chi-Sheng Wu, Hsin-Pai Li, Chia-Hsun Hsieh, et al. Integrated multi-omics analyses of oral squamous cell carcinoma reveal precision patient stratification and personalized treatment strategies. Cancer Letters, 2025. https://doi.org/10.1016/j.canlet.2025.217482 ↩
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.