Why Counts Need a Sample-Specific Scale
Sequencing produces counts: for each sample and genomic feature, the number of reads assigned to that feature. These counts are not expression measurements on a common scale by themselves. A sample with more usable reads can have larger counts for many features simply because it was sequenced more deeply. Comparing its raw counts with those from a shallower sample can therefore confuse technical sampling exposure with differences in expression.
A library size factor is a sample-level scaling quantity used to account for that exposure. It does not change the observed counts, and it does not by itself make counts measures of absolute RNA molecules per cell. Rather, it provides a common relative scale for comparing count profiles, provided the scaling assumptions are reasonable.
The total count \(L_j\) and the size factor \(s_j\) are related ideas, but they are not interchangeable by definition. The total is directly calculated from the observed counts. A size factor is intended to represent relative exposure and may be estimated using information beyond the total count. For a simple illustration, total-count scaling uses \(L_j\) itself, or a constant multiple of it, as the factor. More robust ways to estimate factors are separate topics.
Because a factor is relative, its absolute numerical scale is arbitrary. Replacing every \(s_j\) by \(c s_j\) for the same positive constant \(c\) multiplies every normalized count by \(1/c\), but leaves ratios between samples unchanged. For convenience, a set of relative factors may be rescaled to have geometric mean one. This choice sets a reference scale; it does not add information or alter the relative comparisons.
A Simple Count Model and Its Implication
A useful model separates a sample-wide exposure effect from feature-specific abundance. Suppose the expected count for feature \(i\) in sample \(j\) is
where \(s_j\) is the sample's relative sequencing exposure and \(q_{ij}\) is the feature-specific expected abundance on the reference scale. The model assumes that exposure acts multiplicatively and equally across features: doubling \(s_j\), with \(q_{ij}\) held fixed, doubles the expected count for every feature. This is a simplifying assumption, not a guarantee about real data.
For illustration, counts may be modeled with a Poisson distribution, whose mean and variance are both \(\mu_{ij}=s_jq_{ij}\). In many sequencing applications, counts are more variable than that model permits; a negative-binomial model allows variance greater than the mean. The essential scaling point is the same in either case: the expected count includes a sample-specific exposure term. Normalization targets that term; it does not remove biological variability or make every feature equally expressed.
The justification for dividing by the factor follows directly from the model. If \(s_j\) were known, then the expected normalized count is
Thus, if two samples have the same feature-specific expected abundance \(q_{ij}=q_{ik}\), but different exposures, their raw expected counts differ while their normalized expected counts agree:
This result requires that the factors correctly describe the exposure difference and that the feature-specific abundance being compared has the stated relationship across samples. With estimated rather than known factors, normalized counts also inherit estimation error.
Worked Example: Correcting a Twofold Depth Difference
Consider two hypothetical libraries and two features. Suppose the raw counts are 100 and 900 in sample A, and 200 and 1,800 in sample B. For this illustration, assume sample B had twice the sequencing exposure of sample A, with no change in the underlying abundance profile. Use factors \(s_A=1\) and \(s_B=2\).
For the first feature, the normalized counts are \(100/1=100\) in sample A and \(200/2=100\) in sample B. For the second feature, they are \(900/1=900\) and \(1{,}800/2=900\). The totals are 1,000 and 2,000, respectively, so the raw profiles differ in scale, while the normalized profiles agree.
The calculation supports the intended interpretation only because the example stipulates a common twofold exposure difference and an unchanged profile. It does not show that every pair of libraries should be normalized by its total count, nor that a factor can reveal absolute expression per cell.
Factors, Totals, and the Assumptions Behind Scaling
A scalar factor is appropriate when a sample-wide difference in exposure is a plausible explanation for a broadly proportional change in counts. Relevant exposure differences can arise from the number of reads retained after quality control, sequencing depth, and the fraction of reads that can be assigned to the features being analyzed. Samples also need comparable feature definitions and processing: if one sample's counts include a different set of features or a systematically different class of reads, a single factor may not make the profiles comparable.
Total-count scaling is easy to calculate but makes a strong implicit assumption: the total counts provide a suitable measure of relative exposure. This can fail when the composition of the library changes substantially. For instance, a few highly abundant features can contribute a large share of one sample's total. The total can then change because of those features, rather than because all features experienced a common sequencing-depth multiplier.
Worked Example: When the Total Count Is a Misleading Proxy
Suppose two hypothetical samples have counts for a highly abundant feature and for all other features combined as follows: sample A has 100 and 900; sample B has 1,100 and 900. The totals are \(L_A=100+900=1{,}000\) and \(L_B=1{,}100+900=2{,}000\).
If total-count factors are set to \(s_A=1\) and \(s_B=2\), the normalized counts for the combined other features are \(900/1=900\) in A and \(900/2=450\) in B. The highly abundant feature is \(100/1=100\) in A and \(1{,}100/2=550\) in B.
If the other features were in fact unchanged and the increase in the highly abundant feature were biological, total-count scaling has divided the unchanged count of 900 by a factor of two. It therefore makes the unchanged features appear lower on the normalized scale. The example does not prove which biological explanation is true; it shows that the total alone cannot distinguish a global exposure change from a compositional change. A factor-estimation approach needs assumptions or additional information that make this distinction credible.
A second important distinction is between a sequencing-scale adjustment and an absolute measurement. Without suitable external calibration, counts generally describe relative representation in the sequenced library. If the total amount of RNA per cell changes between samples, a scaling factor estimated from the count profiles alone does not, by itself, identify that change. Claims about absolute abundance require information that anchors the measurements to an appropriate external scale.
Worked Normalization and a Count-Model Offset
Normalized counts are helpful for describing and plotting profiles, but many inferential analyses use the raw integer counts and include the size factor in the statistical model. This preserves the count-data scale rather than treating a noninteger rescaled count as if it were an observed count.
Under the mean model \(E(K_{ij})=s_jq_{ij}\), taking logarithms for positive means gives
The term \(\log(s_j)\) is an offset: a known or estimated sample-specific quantity added to the feature's model for the log expected count. It adjusts for exposure while allowing the model to describe differences in \(q_{ij}\). The logarithm here is of the mean, not a claim that the observed count itself follows a normal distribution.
Worked Example: Rescaling Counts and Setting an Offset
Suppose three hypothetical samples have size factors \(s_1=0.8\), \(s_2=1.0\), and \(s_3=1.2\), and a feature has raw counts 40, 50, and 72. Treat these factors as given for this calculation; they represent relative exposure on a common scale.
The normalized counts are \(40/0.8=50\), \(50/1.0=50\), and \(72/1.2=60\). The third sample has a higher normalized count for this feature, even after adjusting for its larger factor. That descriptive difference is not, by itself, a statistical test or evidence of a particular biological cause.
For a count model, the corresponding offsets are \(\log(0.8)\approx-0.2231\), \(\log(1)=0\), and \(\log(1.2)\approx0.1823\). For sample 3, for example, the mean relation is \(E(K_{i3})=1.2q_{i3}\), or \(\log E(K_{i3})=0.1823+\log(q_{i3})\). The factor is accounted for in the model rather than by rounding \(72/1.2\) to an integer and treating 60 as the observed count.
Checking Whether the Adjustment Is Plausible
A size factor is a model-based adjustment, not a certificate that samples are comparable. Before interpreting normalized profiles, check whether a single sample-wide multiplier is plausible and whether the factors are supported by the data or by the technical information used to obtain them.
- Check sample totals and usable reads. Large differences may reflect depth, but totals alone do not establish that explanation. Confirm that counts were produced using the same feature set and comparable filtering.
- Inspect normalized profiles. If many features remain broadly shifted in one sample after scaling, the factor may be inadequate, the samples may differ in composition, or another technical effect may be present.
- Look for factor extremes. An unusually large or small factor can flag a low-depth sample, a sample with unusual composition, or a processing problem. It merits investigation rather than automatic acceptance or rejection.
- Check biological design and technical information. If a group of samples differs in both biological condition and processing, a scalar depth adjustment cannot by itself separate those effects.
- Assess whether a few features dominate. If they account for a large share of a library, the total may be driven by those features and may not represent a common exposure change across the profile.
These checks diagnose potential problems; they do not identify a unique cause from counts alone. When factors depend on assumptions about which features are stable, widespread changes or strong compositional shifts can make the scale difficult to estimate. External controls can provide an anchor in suitable experiments, but their validity depends on how they were added and what biological quantity they are intended to calibrate.
Alternatives and Limits
| Approach | What it does | Main limitation |
|---|---|---|
| Compare raw counts | Uses the observed integer counts without scaling. | Confounds differences in sequencing exposure with differences in feature abundance when depths vary. |
| Scale by total counts | Uses each sample's sum of counts as a proxy for exposure. | Can be distorted when library composition changes, especially when a few features dominate. |
| Use estimated or externally anchored size factors | Scales samples using a relative exposure estimate based on assumptions or calibration information. | Results depend on the validity of the estimation assumptions or external reference. |
| Use an offset in a count model | Accounts for sample exposure in the expected count while modeling raw counts. | Does not correct a poor factor estimate or other sources of bias. |
These approaches address different practical needs. Raw counts are the observations on which count models are based; normalized counts are convenient for scale-adjusted description; and offsets incorporate exposure into a model. None can recover a biological quantity that the experimental design and calibration do not identify. A size factor also cannot fix sample swaps, poor mapping, feature-specific biases, or confounding between condition and technical processing.
Common Mistake
Wrong reasoning: “Sample B has twice the total counts of sample A, so every feature in B must have twice the expression; dividing all its counts by two gives the true expression.” This conclusion treats the observed total as a direct measure of exposure and interprets relative counts as absolute expression.
Correction: A factor can remove a proportional sample-wide exposure difference when that model is appropriate. The total may instead be large because the library composition changed, and normalized counts remain relative quantities unless an external calibration supports an absolute interpretation. Check the assumptions before choosing or interpreting a factor.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Two samples have counts of 30 and 60 for the same feature, and their known exposure factors are 1 and 2. What are the normalized counts, and what does their agreement show?
- Why can total-count scaling be misleading if one highly abundant feature changes substantially?
- A size factor is 1.5. What is the count-model offset, and how does it enter the log mean model?
- After normalization, many features still appear shifted in one sample. Name two plausible explanations and explain why the factor alone cannot resolve them.
- Does dividing counts by size factors establish an absolute number of RNA molecules per cell? Explain.
Solutions. (1) The normalized counts are \(30/1=30\) and \(60/2=30\). They agree after the stipulated twofold exposure adjustment; this is consistent with equal expected feature abundance under the model, not proof of equal absolute expression. (2) The changing feature can drive the total, so the total may reflect composition rather than a common exposure multiplier; dividing every feature by that total-based factor can distort unchanged features. (3) The offset is \(\log(1.5)\approx0.4055\), and the log mean is \(\log E(K_{ij})=0.4055+\log(q_{ij})\) for that sample. (4) The factor may be inadequate because of compositional change, or another technical or biological shift may remain; counts alone need not identify which explanation applies. (5) No. Normalization gives a relative scale unless suitable external calibration provides an anchor for absolute abundance.