Reading An RNA-seq MA Plot

What an MA Plot Shows

An RNA-seq MA plot places each gene at a point whose horizontal coordinate represents average expression and whose vertical coordinate represents the estimated log fold change between two conditions. It is a compact way to see whether estimated effects vary with expression, whether many genes are shifted in one direction, and which genes have large estimated effects. It is a diagnostic display, not by itself a test of differential expression.

For clarity, take condition 0 as the reference and condition 1 as the comparison. Define the fold change as expression in condition 1 divided by expression in condition 0. In the classical MA definition, \(M\) is the log base 2 ratio, and \(A\) is the mean of the two log base 2 expression levels. For RNA-seq, these expression levels should be based on normalized counts or model-adjusted estimates, not raw counts when library sizes differ.

RNA-seq software does not always use exactly this \(A\) coordinate. Some plots use the log of a mean normalized count on the horizontal axis, while others use the average of log expression levels. Read the axis label or software description before comparing plots. In either case, the horizontal coordinate is an expression-level summary, and the vertical coordinate is usually a log fold change. The vertical sign depends on the stated contrast: reversing the numerator and denominator reverses every sign.

The zero line is a useful visual reference. Points above it have positive estimated log fold changes in the stated direction; points below it have negative estimates. Distance from zero describes the estimated effect size on the log scale, not the certainty of the estimate. For example, \(M=1\) means a twofold increase in condition 1 relative to condition 0, while \(M=-1\) means expression is half as high in condition 1.

Connection to the Count Model

The plotted log fold change often comes from a negative binomial generalized linear model (GLM), rather than a direct ratio of two raw group means. For gene \(i\) in sample \(j\), a typical model is

Here \(Y_{ij}\) is the observed count, \(s_j\) is a sample-specific size factor, \(\mathbf{x}_j\) is the design vector, and \(\alpha_i\) is the gene’s dispersion. A common negative binomial parameterization has \(\operatorname{Var}(Y_{ij})=\mu_{ij}+\alpha_i\mu_{ij}^2\). The model assumes the counts are appropriately represented by this mean-variance relationship, the design captures the relevant comparisons, and samples are independent conditional on the modeled covariates. The size factors must adequately account for differences in sequencing depth and composition.

If the condition coefficient is estimated on the natural-log scale, the log base 2 fold change is the coefficient divided by \(\ln(2)\). If the model uses a log base 2 parameterization, the coefficient is already on the MA plot’s usual vertical scale. Model-based estimates can differ from simple ratios of observed means because the model accounts for sample-level offsets and the design.

Why the Axes Work Together

The coordinates separate overall abundance from relative change. This follows directly from their definitions. Write \(a=\log_2(q_{i0})\) and \(b=\log_2(q_{i1})\), so \(A=(a+b)/2\) and \(M=b-a\). Solving these two equations gives

Consequently, increasing both condition expression levels by the same multiplicative factor moves a point to the right but leaves \(M\) unchanged. Changing the ratio between conditions changes \(M\). This is why the plot can reveal whether fold-change estimates are systematically different at low versus high expression.

Under a null condition effect, the expected log fold change is zero, so a well-calibrated set of estimates should generally be centered near the horizontal zero line. This does not mean that every point must lie close to zero: sampling variation, real condition effects, and model uncertainty all contribute to the scatter.

Reading Low-Expression Variability

A frequent MA-plot pattern is a wider vertical spread at low average expression. This is plausible for count data: when the mean count is small, ordinary sampling variation can be large relative to the mean. In the negative binomial model, the variance-to-mean-squared ratio is \( \operatorname{Var}(Y_{ij})/\mu_{ij}^2=1/\mu_{ij}+\alpha_i \). The \(1/\mu_{ij}\) component is larger when the mean is small. Thus low-count genes can have less stable fold-change estimates, even when their underlying effects are similar to those of more highly expressed genes.

This pattern is not proof that every low-expression estimate is unreliable, nor does a high-expression estimate automatically have small uncertainty. Replicate number, dispersion, design, and the observed data also matter. The MA plot shows the relationship between estimated effect and expression; it does not show a confidence interval for each effect unless uncertainty is explicitly added.

Broad Shifts and Large Effects

A broad vertical displacement can indicate that many estimated effects point in the same direction. It may reflect a real widespread biological response, an issue with normalization, or both. Many RNA-seq normalization procedures rely on an assumption that most genes are not changing in the same direction by the same amount, or that increases and decreases are sufficiently balanced. If that assumption is unsuitable, a global shift can be difficult to distinguish from a sample-wide technical difference using the MA plot alone.

Large absolute values of \(M\) correspond to large fold changes: \(M=2\) corresponds to fourfold higher expression in condition 1, and \(M=-2\) corresponds to one-quarter as much expression in condition 1. A large effect is not automatically precise or statistically compelling. In particular, an extreme estimate at low expression may be based on limited count information. Formal inference requires the model and its uncertainty, not just visual distance from zero.

Diagnostics, Comparisons, and Limitations

Read the MA plot as a pattern, then check whether that pattern is compatible with the analysis. Look for a center near zero, a change in spread across expression levels, broad one-direction shifts, and isolated large effects. Check the axis definitions, the contrast direction, and whether the plotted effects are raw model estimates or estimates altered by a separate processing step. A mismatch between the plotted values and the stated contrast can reverse the interpretation of every sign.

The plot complements rather than replaces other diagnostics. A dispersion-mean plot examines how estimated variability relates to abundance; an MA plot examines estimated condition effects against abundance. Residual diagnostics assess model fit from a different angle. A table of coefficients and uncertainty or test results is needed for formal gene-level inference. No single one of these displays answers all three questions: how large is an effect, how uncertain is it, and how well does the model represent the data?

Common warning signs include a systematic offset from zero, a strong trend in the middle of the plot, and unusually wide scatter at low expression. These are prompts to check the contrast, normalization, design, and model fit, not automatic proof of a particular failure. An isolated extreme point deserves attention, but diagnosing its count-level cause requires examining the underlying observations and model; its height alone is not enough.

The main limitation is that an MA plot compresses many details into two coordinates. It does not show sample-to-sample variability, uncertainty intervals, or the reason for a global shift. Different software may also use different horizontal summaries. Interpret the plot in context, and do not compare coordinates across analyses until the axis definitions and contrast directions are aligned.