Diagnostic Plots for Expression Normalization

Why inspect normalized expression?

Normalization aims to make expression measurements more comparable by reducing technical differences while retaining meaningful biological variation. A normalization method can produce values on a convenient scale without making every relevant feature of the data comparable. Diagnostics help reveal patterns that merit investigation: shifted sample distributions, intensity-dependent differences between samples, or variability that still changes with average expression.

Let \(y_{gj}\) denote the normalized, transformed expression value for gene \(g\) in sample \(j\). For sequencing data, \(y_{gj}\) might be log-CPM; for array data, it might be a normalized log-scale intensity. The plots below assume that the values being compared are on the same defined expression scale. The samples should also have undergone the intended preprocessing, and the gene set used for comparisons should be stated. Different filtering, missingness, or transformations can change a plot’s appearance.

No single plot establishes that normalization is correct. Distributional alignment is a useful check, not a universal target: a real biological difference in overall expression composition can make two properly processed samples look different. Conversely, a normalization procedure may force distributions to align while obscuring a genuine global shift. Interpret visual patterns alongside the study question and the assumptions of the normalization method.

Sample-wise boxplots and density curves

A boxplot summarizes the distribution of expression values across genes within each sample. Compare sample medians, interquartile ranges (IQRs), whiskers, and outlying values. Large differences can flag a sample-scale discrepancy, unusual spread, or a small group of extreme measurements. A density curve displays the distribution in more detail and can reveal differences in skewness, shoulders, or multiple modes that a boxplot compresses.

These comparisons require a common scale and a comparable set of genes. A difference in medians or curve shapes is a prompt to ask why, not an automatic instruction to force the distributions to match. In particular, if many genes truly shift together, global alignment can be biologically inappropriate. Density curves also depend on smoothing choices, so small wiggles should not be treated as distinct biological populations without further evidence.

MA plots: differences across expression intensity

An MA plot compares two samples gene by gene. For samples \(j\) and \(k\), define the average expression \(A_g\) and difference \(M_g\) for gene \(g\) by

Thus, \(A_g\) is the average log-expression and \(M_g\) is the log-expression difference, in the same log-scale units. If comparable samples have been normalized successfully and most genes are not expected to shift systematically, the center of \(M_g\) should be near zero across much of the intensity range, without a persistent trend in \(A_g\). This expectation depends on the biological comparison and the normalization assumptions; it is not a requirement that every gene have \(M_g=0\).

A global offset and an intensity-dependent bias are different patterns. To see why, suppose the difference between two samples can be represented approximately as

where \(c\) is a constant sample offset, \(f(\theta_g)\) varies with the gene’s underlying expression intensity, \(d\) is the size of an intensity-dependent component, and \(\varepsilon_g\) is gene-specific variation. A nonzero \(c\) shifts the plot vertically. If \(d\ne 0\), the expected difference can change with intensity, producing a trend or curvature. Centering the overall differences can remove a constant offset from a summary, but it cannot by itself remove a trend across intensity.

Mean–variance plots

A mean–variance plot asks whether variability across samples depends on a gene’s average expression. For each gene \(g\), calculate the sample mean and sample variance of its normalized expression values across \(n\) samples:

Plot \(s_g^2\) against \(\bar{y}_g\), often using a logarithmic vertical axis if variances span a wide range. A systematic increase or decrease in variance with mean indicates that the transformed values may retain a mean-dependent scale pattern. A transformation such as log-CPM or a variance-stabilizing transformation can reduce such dependence, but the plot checks what remains in the data actually analyzed.

The variance here is across the samples included in the calculation. It can combine technical variation with genuine differences among biological samples. If samples belong to distinct biological groups, their mean differences can contribute to a gene’s across-sample variance. Consequently, an upward trend is not automatically a normalization failure, and a flat trend is not proof that all sources of variation have been handled. State which samples and genes went into the calculation, and interpret the plot in that context.

Combining plots and checking failure modes

The plots answer related but distinct questions. Boxplots and density curves compare whole-sample distributions. An MA plot examines pairwise differences as a function of average intensity. A mean–variance plot examines how across-sample variability relates to gene-level average expression. Agreement among diagnostics can strengthen a concern, while disagreement can help localize it. For example, similar sample medians do not rule out an MA trend: positive differences at low intensity and negative differences at high intensity can partly cancel in an overall summary.

A practical review should first confirm that plots use the intended normalized scale and a clearly defined, comparable gene set. Then inspect whether patterns are broad or driven by a few genes, whether they recur across relevant sample pairs, and whether they are plausible given the biology and preprocessing. Plot labels, axes, sample identities, and units matter: an MA plot with reversed sample order changes the sign of every \(M_g\), though not the presence of an intensity trend.

Common failures include treating identical distribution shapes as proof of valid normalization, treating any sample-wise difference as a technical defect, and interpreting a mean–variance trend without considering group composition. Very small gene sets, extensive missingness, and a few extreme values can also make summaries unstable or misleading. These plots are diagnostics, not formal tests and not substitutes for understanding how the normalization method behaves.