Why a Normalization Comparison Must Respect the Platform
Expression measurements can differ between samples for reasons that are not the biological differences of interest. Normalization addresses specified technical or compositional differences so that measurements can be compared on a defensible scale. It does not make different platforms interchangeable, and it cannot determine whether an observed shift is technical or biological without assumptions about the data and study.
Microarrays measure probe intensities, often summarized and transformed to a log scale before between-array normalization. RNA sequencing begins with nonnegative integer counts, whose sample totals and composition can differ. A shared-distribution method may suit a microarray setting under its assumptions; a count-based size factor is designed to adjust sequencing counts while retaining their count scale. The introductory material on normalization targets and expression scale distinguishes these measurement scales; the comparison here applies those choices to small hypothetical data sets.
Microarray Example: Quantile Normalization
Suppose three hypothetical arrays have log-scale summarized probe intensities for four probes. Quantile normalization sorts each array, averages the values at each rank across arrays, then assigns each probe the average for its rank in that array. Its key working assumption is that the overall distributions should be comparable across arrays; if genuine biology shifts many probes together, forcing equal distributions can remove part of that signal.
| Probe | Array 1 | Array 2 | Array 3 |
|---|---|---|---|
| P1 | 4 | 6 | 5 |
| P2 | 8 | 8 | 9 |
| P3 | 6 | 10 | 6 |
| P4 | 12 | 12 | 11 |
The arrays sorted from smallest to largest are \((4,6,8,12)\), \((6,8,10,12)\), and \((5,6,9,11)\). The mean at each rank is calculated across the three arrays:
Worked Example: Mapping Microarray Ranks
In Array 1, P1, P3, P2, and P4 have ranks 1, 2, 3, and 4, respectively. Their normalized values are therefore 5, 6.667, 9, and 11.667 in that probe order. In Array 2 the probes are already ordered P1, P2, P3, P4, with values 6, 8, 10, 12, so they receive 5, 6.667, 9, 11.667. In Array 3, P1, P3, P2, P4 have ranks 1, 2, 3, 4, respectively.
The resulting normalized matrix is:
| Probe | Array 1 | Array 2 | Array 3 |
|---|---|---|---|
| P1 | 5.000 | 5.000 | 5.000 |
| P2 | 9.000 | 6.667 | 9.000 |
| P3 | 6.667 | 9.000 | 6.667 |
| P4 | 11.667 | 11.667 | 11.667 |
Each array now has the same sorted distribution, \((5,6.667,9,11.667)\), but probes do not all have the same value across arrays. For example, P2 is 9 in Arrays 1 and 3 but 6.667 in Array 2 because its rank differs. The rank mapping, rather than matching each probe to a fixed value, is the defining operation.
A useful check is to sort each normalized column and verify that the columns agree, allowing for ties and rounding. Tied values require a consistent rank-handling convention; software implementations may average tied ranks. The procedure also assumes that all arrays contain the same set of features after filtering and summarization. It should be applied to appropriate summarized expression values, not to raw probe intensities before the relevant background correction and summarization steps.
RNA Sequencing Example: Size Factors for Counts
For RNA-seq, a size factor \(s_j\) adjusts the count scale for sample \(j\); a normalized count is \(K_{ij}/s_j\), where \(K_{ij}\) is the count for gene \(i\). A simple library-total factor reflects each sample's total count. The median-ratio approach estimates factors from gene-wise ratios to geometric means and is less sensitive to a small number of highly changed genes, provided a substantial set of genes supplies a stable reference. Neither factor makes normalized counts into a direct measure of absolute molecules per cell.
Consider five hypothetical genes measured in three samples. Assume all counts are positive, so geometric means are defined.
| Gene | S1 | S2 | S3 |
|---|---|---|---|
| A | 100 | 1000 | 100 |
| B | 80 | 80 | 80 |
| C | 40 | 40 | 40 |
| D | 20 | 20 | 20 |
| E | 10 | 10 | 10 |
Worked Example: Comparing Total-Count and Median-Ratio Factors
The library totals are 250 for S1, 1150 for S2, and 250 for S3. Scaling each total by the mean total, \((250+1150+250)/3=550\), gives total-count factors 0.4545, 2.091, and 0.4545. For example, S2's factor is \(1150/550=2.091\). The normalized count for gene B in S2 using the unrounded factor is \(80/(1150/550)=38.26\), compared with \(80/(250/550)=176.0\) in S1. This method has treated S2's larger total as a general increase across genes.
For the median-ratio method, first take each gene's geometric mean across the three samples. For gene A it is \((100\times1000\times100)^{1/3}=215.4\), approximately. The three ratios for A are approximately 0.4643, 4.642, and 0.4643. For each of B through E, the geometric mean equals that gene's count, so each sample's ratio is 1. The sample-wise medians of the five ratios are therefore 1, 1, and 1: the outlying ratio for A does not move the median. The geometric mean of these preliminary factors is also 1, so the final median-ratio factors remain 1, 1, and 1.
The median-ratio normalized counts are consequently the original counts in this small example. In particular, gene B remains 80 in all three samples, while gene A is 100, 1000, and 100. This is appropriate if most genes provide a stable reference and the increase for A is a genuine gene-specific change. The example does not establish that assumption from the counts alone.
Library-total scaling assumes that differences in total counts mainly reflect a general sample-scale difference. When a small number of transcripts account for much of one sample's library, that assumption can be poor. Median-ratio scaling instead relies on a stable central set of gene ratios and can fail when changes are widespread and strongly one-directional, or when zeros prevent the chosen geometric-mean calculation. Other count-scaling choices, including TMM and upper-quartile scaling, make related but distinct assumptions; they should not be substituted merely because their factors produce attractive plots.
Use Diagnostics to Compare, Not to Automate, the Choice
Diagnostics summarize consequences of a normalization and can expose problems, but they do not identify the correct method without context. Before selecting a matrix, confirm that samples have comparable feature definitions, that the input scale is appropriate, and that the procedure's assumptions are credible for the experimental question. For the microarray example, equalized sorted distributions are expected by construction. That check confirms the calculation, not the biological validity of the equal-distribution assumption.
Worked Example: Reading a Gene-Wise Diagnostic
Compare S2 with S1 in the RNA-seq example. With median-ratio factors of 1, gene B has normalized counts 80 and 80, so its log2 ratio is \(\log_2(80/80)=0\). Gene A has normalized counts 1000 and 100, giving \(\log_2(1000/100)=\log_2(10)=3.322\). A gene-wise MA plot places the average expression level on one axis and the log-ratio on the other; the four stable genes B through E lie at log-ratio 0, while A has log-ratio 3.322.
Under total-count factors, gene B has normalized counts 38.26 in S2 and 176.0 in S1. Its S2-to-S1 ratio is \(38.26/176.0=0.2174\), equivalently \(250/1150\), and its log2 ratio is approximately \(-2.202\). Thus total-count scaling makes every unchanged gene appear lower in S2 in this constructed example. Gene A's ratio under that scaling is \((1000/(1150/550))/(100/(250/550))=2.174\), with log2 ratio approximately 1.120, rather than 3.322. The diagnostic makes the consequence visible; the choice of scale still depends on whether the stable genes or total-count assumption is more plausible.
Useful plots include sample boxplots or density plots to inspect marginal distributions, and MA plots to inspect gene-wise log-ratios across average expression. For microarrays, quantile normalization should produce closely matching marginal distributions, so a continuing distribution difference may signal a mismatched input or implementation. For RNA-seq, unequal sample distributions after a defensible count scaling are not automatically a failure: genuine expression changes and count discreteness can affect them. A broad MA shift can signal residual sample-scale differences or an unsuitable reference assumption; strong curvature or a dependence on expression level can indicate that a single scale factor is inadequate.
Replicate structure, sample labels, library quality, and known technical differences also matter when interpreting plots. A plot cannot decide whether a pattern is biological or technical. Nor should samples be forced to look alike if that removes a genuine global response. The subject of batch effects requires separate design information and analysis; these normalization choices alone are not batch correction.
A Practical Comparison and Its Limits
| Feature | Microarray quantile normalization | RNA-seq size-factor scaling |
|---|---|---|
| Input scale | Typically summarized, transformed intensities | Nonnegative gene-level counts |
| Operation | Matches empirical distributions by rank | Divides counts by a sample-specific factor |
| Key assumption | Overall distributions should be comparable | A stable reference exists for estimating relative sample scale |
| Important check | Sorted normalized columns agree as intended | Sample factors and gene-wise ratios are plausible |
The methods are not interchangeable. Quantile normalization changes individual values according to their within-array ranks and uses a shared target distribution. Size-factor scaling preserves relative count relationships within each sample by dividing by one factor per sample, although the resulting normalized counts are no longer integers. Do not apply a microarray quantile transformation to raw sequencing counts as a generic fix, and do not interpret normalized microarray intensities as sequencing counts.
All these approaches have limitations. A small toy matrix cannot establish the assumptions in a real experiment. Normalization cannot repair poor sample quality, incorrect feature annotation, or a design in which technical and biological effects are inseparable. Estimates may also depend on filtering, zero handling, and the chosen reference population of genes. Preserve raw data and record the transformation, factors, and diagnostics so that downstream analyses use the correct scale.
Common Mistakes
- Confusing equal distributions with equal expression. Quantile normalization makes array distributions match by construction; it does not show that each probe has equal biological expression across arrays.
- Dividing all sequencing samples by their totals without checking composition. A dominant transcript can inflate one library total and make stable genes appear reduced after total-count scaling.
- Choosing the method that makes a preferred plot look neat. Diagnostics reveal consequences, but the assumptions and scientific question determine whether those consequences are acceptable.
- Comparing normalized numbers across platforms. Microarray intensity values and normalized sequencing counts have different measurement origins and units.
- Treating normalization as batch correction. A scale adjustment addresses its stated sample-scale target; it does not establish or remove every technical source of variation.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- Three arrays have values at rank 1 of 2, 5, and 8. What value does quantile normalization assign to a probe with rank 1 in the second array?
- In the sequencing example, why do the median-ratio factors equal 1 even though S2 has a much larger library total?
- Using total-count factors, calculate the S2-to-S1 normalized-count ratio for an unchanged gene with count 40 in both samples.
- Does matching the microarray distributions prove that no biological signal was removed? Explain.
Solutions. (1) The rank-1 target is \((2+5+8)/3=5\), so the second-array value becomes 5. (2) Four genes have ratio 1 in every sample; the median of five gene-wise ratios remains 1 despite gene A's large S2 ratio, so the factors are 1. (3) The unchanged gene's ratio is \((40/(1150/550))/(40/(250/550))=250/1150=0.2174\). (4) No. Matching distributions is an enforced property of the method; if genuine biology changes many probes together, that assumption can remove some of the biological shift.