Library Size Factors as Multiplicative Exposure Corrections

Why Cells Need an Exposure Correction

A raw count reflects more than the relative abundance of a gene. It can also depend on how much material was captured from a cell and how deeply that cell was sequenced. If one cell has twice the effective exposure of another, it may produce roughly twice as many reads for every gene even when their underlying relative expression profiles are the same. Comparing those raw counts directly can therefore confound exposure with expression.

A size factor represents this cell-specific multiplicative exposure. Dividing each count in a cell by its size factor puts the counts on a common relative scale, provided the factor adequately represents the exposure difference across genes. This is a correction of scale, not a transformation that turns observed counts into exact measurements of molecules.

Definition: Let $K_{gj}$ be the observed count for gene $g$ in cell $j$, and let $s_j>0$ be that cell’s size factor. A multiplicative exposure model has mean $$ E[K_{gj}]=s_j\lambda_{gj}, $$ where $\lambda_{gj}$ is the gene’s mean count on the chosen reference exposure scale. The factor-scaled count is $K_{gj}/s_j$. If the model is appropriate, its expected value is $\lambda_{gj}$.

The factor is shared across genes within a cell. That shared-factor assumption is what makes a single cell-level correction useful: it says that exposure changes the expected counts for all genes proportionally. It does not say that genes have the same expression, or that the observed counts themselves are multiplied by an exact deterministic constant. Counts remain random, and the factor adjusts their expected scale.

Assumptions and Meaning of the Corrected Scale

The multiplicative interpretation requires several conditions. First, factors must be positive, since division by zero is undefined and negative exposure has no meaningful interpretation. Second, the relevant exposure difference must act approximately proportionally across the genes being compared. Third, the factors must refer to a consistent reference scale across cells. If any of these conditions fails, dividing by the factors may leave systematic cell differences or distort gene comparisons.

A simple count model makes the assumptions explicit. For example, suppose counts are conditionally Poisson:

K_{gj}\mid s_j,\lambda_{gj}\sim\operatorname{Poisson}(s_j\lambda_{gj}).

Here the Poisson assumption describes the count variation around the specified mean; the exposure correction itself follows from the mean being a product of $s_j$ and $\lambda_{gj}$. Other count distributions can also have a mean of this form. The correction does not require the Poisson variance assumption, but the variance matters for uncertainty and for downstream analyses.

The $\lambda_{gj}$ values are relative to the chosen exposure scale. If every factor is multiplied by the same positive constant, the corrected values are divided by that constant. Thus a factor scale must be chosen or standardized to give normalized values a particular numerical reference. Comparisons among cells are often unchanged by this common rescaling, but the numerical units of the corrected counts do change.

Worked Example: Correcting a Two-and-a-Half-Fold Exposure Difference

Suppose two hypothetical cells have observed counts of 24 and 60 for the same gene. Their supplied size factors are $s_1=0.8$ and $s_2=2.0$. Assume these factors adequately represent a shared multiplicative exposure difference for this gene.

The exposure ratio is $2.0/0.8=2.5$. The second cell has 2.5 times the factor of the first. Divide each observed count by its own factor:

\frac{24}{0.8}=30,\qquad \frac{60}{2.0}=30.

Both factor-scaled counts are 30 on this reference scale. The raw counts differed by a factor of $60/24=2.5$, matching the factor ratio, so their difference is consistent with exposure alone under the stated model. This agreement does not prove the cells have identical biology; it shows how the supplied correction removes the specified multiplicative difference in this example.

Why Division Works—and What It Does Not Do

The expectation of a scaled count follows directly from linearity of expectation. For a fixed positive factor $s_j$,

E\left[\frac{K_{gj}}{s_j}\right] =\frac{E[K_{gj}]}{s_j} =\frac{s_j\lambda_{gj}}{s_j} =\lambda_{gj}.

So when the mean model is correct and the factor is treated as known, division removes the factor from the expected value. The reasoning is about the mean, not about each realization: a particular observed count divided by its factor can still be above or below $\lambda_{gj}$ because of sampling variation.

For the Poisson model, the variance also shows why equal expected values do not imply equal precision. Since $\operatorname{Var}(K_{gj})=s_j\lambda_{gj}$,

\operatorname{Var}\left(\frac{K_{gj}}{s_j}\right) =\frac{\operatorname{Var}(K_{gj})}{s_j^2} =\frac{\lambda_{gj}}{s_j}.

For example, at the same underlying $\lambda_{gj}$, a cell with a larger factor has a smaller variance for its factor-scaled count under this Poisson model. More generally, count distributions may be overdispersed relative to Poisson, and estimated factors add uncertainty of their own. Factor-scaled values are therefore not automatically equally precise measurements.

Worked Example: Correcting Counts Across Two Genes

Consider two hypothetical cells with factors $s_A=0.5$ and $s_B=1.5$. For gene P, their counts are 20 and 60; for gene Q, their counts are 5 and 30. Suppose the shared-factor assumption is reasonable for these genes.

The factor ratio is $1.5/0.5=3$. For gene P, divide each count by its cell’s factor:

\frac{20}{0.5}=40,\qquad \frac{60}{1.5}=40.

The corrected values for gene P agree; its threefold raw-count difference is exactly the threefold factor difference. For gene Q, the corresponding calculation is

\frac{5}{0.5}=10,\qquad \frac{30}{1.5}=20.

After correction, gene Q has a factor-scaled count of 20 in cell B and 10 in cell A. Its raw counts differ sixfold, and the threefold exposure correction leaves a twofold difference: $20/10=2$. Under the model, that remaining difference is consistent with a difference in relative expression, though the small counts would make sampling variation important in a real analysis.

This example also illustrates a diagnostic idea. If many genes show similar raw-count ratios matching the factor ratio, a shared exposure explanation is plausible. If many corrected genes still shift in the same direction, the factor may be inadequate, or the cells may differ biologically in a broad way.

Factor Scale and Biological Interpretation

Factors are meaningful through their relative values, but their absolute scale is a convention unless an external calibration defines it. Suppose counts are 40 and 80, and the respective factors are 1 and 2. The scaled values are $40/1=40$ and $80/2=40$. If both factors are multiplied by 4, the new factors are 4 and 8, and the scaled values become $40/4=10$ and $80/8=10$. The common cell comparison is unchanged, while the numerical reference scale changes fourfold.

A more serious interpretive issue arises when biology itself changes globally. A size factor corrects the exposure represented by that factor; it cannot by itself distinguish technical exposure from a biological change that affects counts broadly. For instance, suppose both cells have a raw average count of 100 for a marker, but cell B’s factor is twice cell A’s. The scaled values are $100/s_A$ for A and $100/(2s_A)$ for B, so B’s scaled value is half A’s. Equal raw counts do not yield equal corrected values when the factors differ.

Conversely, if B’s raw average were 200 while its factor were twice A’s, then the scaled averages would agree: $100/s_A=200/(2s_A)$. The arithmetic is straightforward, but the biological interpretation depends on what the factor represents and what comparison the analysis intends. Without suitable external calibration, factor-scaled counts should not be described as absolute molecules per cell.

Worked Example: A Common Rescaling Changes Units, Not Ratios

A hypothetical gene has counts 40 and 80 in cells A and B, with factors 1 and 2. The corrected values are $40/1=40$ and $80/2=40$, so their ratio is $40/40=1$.

Now multiply both factors by $c=4$, giving factors 4 and 8. The corrected values become $40/4=10$ and $80/8=10$, again with ratio $10/10=1$. In general, if $s'_j=cs_j$ for every cell, then

\frac{K_{gj}}{s'_j} =\frac{1}{c}\frac{K_{gj}}{s_j}.

Every corrected value is rescaled by the same $1/c$, so comparisons such as ratios between cells are preserved. The values’ absolute numerical scale is not preserved. This is why a normalized value must be interpreted together with the convention used to set the factor scale.

Diagnostics, Alternatives, and Limitations

A factor correction is only as credible as the exposure model and factor values. Useful checks include comparing factors with total observed counts, examining whether corrected values retain strong cell-wide trends with the factors, and checking whether factor-scaled profiles behave plausibly across genes and cells. These checks can reveal problems, but none alone proves that a factor isolates technical exposure from biology.

Residual cell-wide shifts: if many genes remain systematically higher in the same cells after division, the factors may not capture the relevant exposure difference, or a broad biological shift may be present.
Gene-specific mismatch: if technical effects vary by gene, a single cell-level factor cannot correct them all. Systematic discrepancies across gene classes are a warning against assuming one shared multiplier.
Sparse counts: many zeros remain zero after division. Scaling changes nonzero values; it does not recover unobserved molecules or resolve why a zero occurred.
Uncertain factors: treating estimated factors as known understates uncertainty. This can matter especially when factors are unstable or based on little information.

Leaving counts raw avoids choosing a factor but retains exposure differences in the values being compared. Dividing counts gives an intuitive relative-scale representation, but the result is generally non-integer and its variance is altered by scaling. A model can instead represent exposure in its mean structure while retaining the original integer counts; that approach avoids treating scaled counts as if they were raw observations, but requires an appropriate count model and does not eliminate the need for valid factors. External calibration can help anchor exposure to an absolute scale when its assumptions are met, but it does not make every biological interpretation automatic.

The main limitation is conceptual: an endogenous size factor is not a direct measurement of technical exposure alone. If biological changes systematically affect the counts used to characterize exposure, the correction may remove part of the biology of interest or create misleading relative comparisons. A normalized value is therefore best described as a count adjusted to a stated reference scale under a multiplicative exposure assumption.

Common Mistakes

Reversing the correction. If a cell’s factor is larger, its count is divided by a larger number, not multiplied by it.
Claiming exact recovery. Division removes the factor from the expected value under the model; it does not make every noisy observed count equal to its underlying mean.
Calling scaled counts absolute expression. They are relative to the factor scale unless an appropriate calibration supports an absolute interpretation.
Assuming every gene shares the same exposure effect. One factor per cell cannot correct gene-specific distortions that violate the shared-multiplier model.
Equating equal raw counts with equal normalized values. If factors differ, equal raw counts generally produce different scaled values.

Key takeaway: A size factor is a positive cell-specific multiplier in the expected-count model. Dividing by it removes that multiplier from the expected value when the shared exposure assumption holds; the result remains a noisy, relative-scale quantity whose interpretation depends on factor construction and biological context.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A gene has counts 12 and 36 in two cells with factors 0.6 and 1.8. What are the factor-scaled counts, and what does their agreement show?
Under $E[K_{gj}]=s_j\lambda_{gj}$, derive the expected value of $K_{gj}/s_j$ for a fixed positive factor.
If every factor is multiplied by 3, what happens to each scaled count and to ratios of scaled counts between cells?
Two cells have equal raw counts for a gene, but the second cell’s factor is twice the first’s. How do their scaled values compare?
Name one reason that a shared cell-level factor might fail to make cells comparable.

Solutions. (1) The scaled counts are $12/0.6=20$ and $36/1.8=20$. Their agreement is consistent with the raw threefold difference being explained by the threefold factor difference under the model; it does not prove identical biological expression. (2) For fixed $s_j>0$, $E[K_{gj}/s_j]=E[K_{gj}]/s_j=(s_j\lambda_{gj})/s_j=\lambda_{gj}$. (3) Each scaled count is divided by 3, since $K_{gj}/(3s_j)=(1/3)(K_{gj}/s_j)$; ratios between cells are unchanged because the common multiplier cancels. (4) If the raw count is $K$ in both cells and the first factor is $s$, the scaled values are $K/s$ and $K/(2s)$, so the second is half the first. (5) For example, exposure may affect different genes unequally, or broad biological changes may be confounded with the factor. In either case, one shared multiplier may not represent the relevant difference.