Why a Normalization Target Comes First
Expression measurements are not direct readings of biological abundance. A measured signal can change because the amount of a transcript changed, but also because of differences in sample input, labeling, imaging, sequencing depth, or other technical features. Normalization adjusts measurements to a chosen reference scale so that specified comparisons are more interpretable. The target matters: a procedure that makes samples comparable in one sense may remove real biological differences in another.
This tutorial focuses on the meaning of that target rather than on a particular algorithm. Before applying any normalization, it is useful to state which comparisons are intended: differences among genes within one sample, differences in a gene across samples, or some form of absolute abundance per cell or per unit of material. These are distinct goals, and ordinary expression measurements do not automatically support all of them.
A Measurement Model and the Meaning of Scale
Let \(A_{gj}\) denote the biological abundance of feature \(g\) in sample \(j\), and let \(Y_{gj}\) denote its measured signal. A simple model for positive measurements is
where \(c_j\) is a sample-specific technical scale factor and \(q_g\) is a feature-specific response factor. For example, \(c_j\) could represent a common multiplicative difference in signal intensity, while \(q_g\) could represent a feature's differing measurement response. Real data also contain noise and can depart from this simple model; it is a useful starting point, not a complete description of every platform.
On a logarithmic scale, multiplication becomes addition:
This makes the distinction clear. A sample-wide technical shift adds the same amount to every feature on the log scale, whereas biological abundance can vary by feature and sample. If the same feature is measured comparably across two samples, its response factor \(q_g\) cancels in a ratio. But sample scale and biology remain entangled unless there is a reference or assumption that separates them.
Thus the observed fold change is not necessarily the biological fold change. A normalization target supplies a way to estimate or define the technical component, typically by assuming that some reference features, control measurements, or overall distribution should be comparable. That assumption must be plausible for the study question.
Worked Example: A Technical Gain Can Mimic a Biological Change
Suppose one gene has hypothetical abundance \(A_{g1}=50\) units in sample 1 and \(A_{g2}=50\) units in sample 2. The feature response is \(q_g=1\), but the sample scale factors are \(c_1=1\) and \(c_2=2\).
The measured values are \(Y_{g1}=1(1)(50)=50\) and \(Y_{g2}=2(1)(50)=100\). The raw ratio is \(100/50=2\), which looks like a two-fold increase. Under the stated model, however, the biological ratio is \(50/50=1\). Correcting by the known scale factors gives \(Y_{g1}/c_1=50\) and \(Y_{g2}/c_2=50\).
This calculation is valid because the example supplies the technical factors and assumes a common multiplicative effect. In an actual experiment, those factors are not known merely from observing the two values. Without a reference or additional information, the data alone cannot establish whether the difference came from \(c_j\), \(A_{gj}\), or both.
Within-Sample and Between-Sample Comparisons
Within-sample normalization uses relationships among measurements from the same sample. It can be appropriate when the scientific comparison is relative within that sample, or when a known within-sample technical relationship needs correction. For example, in a two-channel measurement, comparing the two channel signals for the same feature can remove a common feature response factor. This does not by itself make different samples comparable, nor does it create a measure of absolute abundance.
Between-sample normalization instead aligns samples to a common scale. It may aim to remove a sample-wide technical multiplier, or it may define a common distributional reference. Such adjustment can support comparisons of a given feature across samples when the reference assumptions hold. It does not necessarily make the numerical values comparable across different features, because \(q_g\) can vary by feature.
The two operations answer different questions. A within-sample ratio can be stable despite a sample-wide multiplier, because that multiplier cancels. But it may not preserve information about total abundance in the sample. Between-sample scaling may support relative comparisons across samples, but its estimate can be misleading if the reference features have changed biologically in a coordinated way.
Worked Example: Interpreting a Within-Sample Ratio
Imagine a hypothetical two-channel assay for two genes in one sample. Gene A has signals 400 in channel 1 and 200 in channel 2; gene B has signals 100 and 100. Define the within-sample channel ratio as channel 1 divided by channel 2.
For gene A, the ratio is \(400/200=2\), and its log base 2 ratio is \(\log_2(2)=1\). For gene B, the ratio is \(100/100=1\), and its log base 2 ratio is \(\log_2(1)=0\). The within-sample comparison says that the channel 1 signal is twice the channel 2 signal for gene A, while the channel signals are equal for gene B.
Suppose, as a model check, that a technical multiplier of 3 affected both channels equally. Then the signals would be 1200 and 600 for gene A, and 300 and 300 for gene B. The ratios remain 2 and 1, since \(3x/3y=x/y\). This cancellation depends on a common multiplier applying to both channels. It does not show that channel 1 abundance is absolutely twice channel 2 abundance unless channel-specific technical effects and the biological interpretation of the channels are also addressed.
Choosing a Reference Defines the Estimand
A normalized value is best understood as a measurement relative to its reference, not as an unqualified quantity. A reference might be a set of control features, an external standard, a sample-wide summary, or a specified amount of starting material. Each choice corresponds to a different comparison and requires different assumptions.
- Stable endogenous references: These are measured features assumed not to change biologically across the samples of interest. They can provide an internal scale if that stability is credible for the relevant tissue, treatment, and conditions.
- External controls: Added standards can help anchor measurements to a known input or processing scale. Their usefulness depends on being added and measured in a way that tracks the technical variation relevant to the target measurements.
- Global or distributional references: A sample-wide target can be useful when most features are not changing in a coordinated way, or when the intended estimand is explicitly relative to the overall profile. A widespread biological shift can violate this interpretation.
These targets should not be treated as interchangeable. For example, expression per total measured RNA and expression per cell are different quantities. If the total RNA amount per cell changes, a scale based on equal total RNA input can conceal that change. Similarly, a relative measure can show that one gene forms a larger share of the measured pool without showing that there are more copies of that gene per cell.
Worked Example: Equal Proportions Do Not Mean Equal Copies
Consider two hypothetical samples with measured totals of 20,000 and 40,000 count units. A gene has 100 units in the first sample and 200 in the second. Its share of the total is \(100/20{,}000=0.005\) in sample 1 and \(200/40{,}000=0.005\) in sample 2.
On a total-count relative scale, the gene has the same value in both samples: \(0.005\), or 0.5% of the measured total. The raw measured amount doubled, but so did the total. If the totals reflect only different technical sampling depth, the equal relative values are an appropriate correction for that difference.
If instead the totals reflect a real two-fold increase in total RNA per cell and the gene's copies per cell also doubled, the relative scale would hide that biological increase. The arithmetic is identical in either scenario; the interpretation differs because the reference does. Without a per-cell anchor or another suitable measurement, the relative data alone do not distinguish these explanations.
A Short Identifiability Argument
The central limitation can be shown directly. Under the simple model \(Y_{gj}=c_jq_gA_{gj}\), suppose all observed measurements in sample \(j\) are multiplied by a positive constant \(k\). The same observed values can be represented by replacing \(c_j\) with \(kc_j\) and leaving the abundances unchanged, or by leaving \(c_j\) unchanged and multiplying every \(A_{gj}\) by \(k\). In either case, the product \(c_jA_{gj}\), and therefore \(Y_{gj}\), is unchanged.
Consequently, a common technical scale and a coordinated biological shift cannot be separated from these measurements alone. Normalization resolves this ambiguity by adopting a reference convention or assumption; it does not prove that the convention is biologically true. This is why the normalization target is part of the scientific interpretation, not merely a computational setting.
Diagnostics, Alternatives, and Limitations
A normalization choice should be checked against the reference assumption it relies on. Useful checks include comparing sample-wide distributions before and after adjustment, inspecting whether the estimated adjustments are unexpectedly large, and examining known controls or features expected to be stable. These checks can reveal unusual samples, residual technical structure, or a target that is poorly matched to the data. They cannot establish biological invariance solely from the same expression measurements used to estimate the adjustment.
Warning signs include a reference set that differs systematically by treatment, strong sample-wide shifts aligned with a biological condition, and normalization factors that track an experimental group rather than plausible technical variation. When treatment and processing batch are completely confounded, for example, no normalization based only on those measurements can reliably determine which source caused the difference. A numerical adjustment may still be computed, but it does not recover information the design did not provide.
| Approach | What it can support | Main concern |
|---|---|---|
| Within-sample ratios | Relative contrasts among channels or features within a sample, under the relevant common-effect assumptions | Can discard sample-wide abundance information and need not align different samples |
| Between-sample scaling | Comparisons across samples relative to a chosen reference scale | Can remove real global changes if the reference assumes they are absent |
| External or validated controls | An anchor tied to controls or a specified input scale | Controls may fail to track the technical effects or biological quantity of interest |
No reference is universally best. The appropriate choice depends on the scientific estimand, the measurement process, and the study design. Normalization also cannot correct every measurement problem: it does not automatically remove feature-specific bias, repair poor-quality data, or resolve confounding. Different platforms and preprocessing stages can require different procedures, so the particular adjustment should be chosen with the measurement model in view.
Common Mistake
A common error is to say, “After normalization, the values are the true expression levels,” or to interpret a normalized ratio as an absolute fold change without qualification. This reasoning treats the adjustment as if it had identified technical variation from the data alone. The correction is to state the scale explicitly: for example, “the gene's measured share of the sample-wide total was higher,” or “the signal was two-fold higher relative to the stated reference.” An absolute biological interpretation requires an appropriate anchor and assumptions that justify it.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If every signal in one sample is twice as large because of a common technical multiplier, what happens to within-sample ratios? What additional comparison is needed to align samples?
- Two samples have gene counts 80 and 160 and totals 16,000 and 32,000. Calculate each gene's total-relative value. What can and cannot be concluded?
- Why can a global biological increase in abundance be indistinguishable from a sample-wide technical multiplier?
- A normalized value is higher in a treated group after scaling to a set of reference genes. State the key assumption needed to interpret it as a change relative to those genes.
Solutions. (1) A common multiplier cancels in a ratio: \(2x/2y=x/y\). Between-sample normalization is needed to align sample scales, using a justified reference. (2) The values are \(80/16{,}000=0.005\) and \(160/32{,}000=0.005\), so both are 0.5%. They are equal relative to the totals; this does not establish equal or different copies per cell. (3) Both mechanisms multiply the observed measurements by the same factor, so the observations alone do not identify which mechanism produced it. (4) The reference genes must be sufficiently stable in biological abundance across the groups, and the adjustment must address the relevant technical differences.