Why Normalization Factors Become Unstable
A normalization factor is an estimate of relative technical depth: how much count opportunity a cell had compared with other cells. It is not a direct measurement of total RNA, and dividing by a factor cannot by itself distinguish technical depth from genuine biological differences in RNA content or composition. That distinction becomes especially important when a dataset contains few cells, very small libraries, or many zeros.
In this tutorial, let \(Y_{ig}\) denote the observed count for gene \(g\) in cell \(i\), and let \(L_i=\sum_g Y_{ig}\) be the cell's library size over the included genes. A size factor \(f_i\) is a positive relative scaling estimate. For example, dividing counts by \(f_i\) places cells on a common relative scale. The factor's absolute scale is arbitrary: multiplying every factor by the same constant does not change their relative values.
The main warning sign in a low-library cell is that each count has high leverage on its total. If one count is added, the relative change in library size is
Thus, one additional count changes a library of 10 counts by 10%, but changes a library of 1,000 counts by 0.1%. The same sparsity also leaves fewer genes that can anchor a ratio-based estimate. With only a few cells, it is harder to detect an outlying factor, distinguish a technical effect from a different cell composition, or estimate extra model parameters reliably.
Assumptions to Check Before Choosing a Method
All the approaches discussed here rely on a reference notion of what should be comparable across cells. In broad terms, they assume that enough of the measured signal reflects a shared biological baseline, so differences in counts can partly be attributed to technical depth. If most genes truly change together, or if cells differ strongly in RNA composition, library size alone may not represent a technical nuisance. Normalizing such a difference away can erase biology.
- Total-count scaling assumes that library totals are useful proxies for relative depth, rather than being dominated by a few highly abundant genes or genuine global RNA differences.
- Ratio methods need a sufficient set of genes that are expressed across cells and whose relative expression provides a stable reference. Zeros and condition-specific expression reduce the usable set.
- Pooling and deconvolution rely on pooled counts being more informative than individual sparse counts, with enough informative genes and overlapping pools to separate cell-level factors.
- NB regression requires a suitable count mean model and enough information to estimate its parameters. It can accommodate count variability, but a flexible model does not create information absent from a tiny or nearly all-zero dataset.
These are working assumptions, not guarantees. Factors should be treated as estimates with uncertainty, and a method should not be chosen solely because it returns a number for every cell.
Worked Example: A One-Count Change in Total-Count Scaling
Worked Example: A One-Count Change in Total-Count Scaling
Consider three hypothetical cells with counts for three genes. The third gene is not detected in any cell.
| Cell | Gene 1 | Gene 2 | Gene 3 | Library size \(L_i\) |
|---|---|---|---|---|
| A | 8 | 2 | 0 | 10 |
| B | 16 | 4 | 0 | 20 |
| C | 80 | 20 | 0 | 100 |
For illustration, define a relative total-count factor as library size divided by the median library size. The median is 20 counts, so
Now suppose one additional count is recorded for cell A. Its library size becomes 11, while the median remains 20. Its factor becomes \(11/20=0.550\), a relative increase of \((0.550-0.500)/0.500=0.100\), or 10%. The corresponding increase in cell C's library would change its total from 100 to 101, only a 1% change. This sensitivity is a consequence of the small denominator, not a computational error.
The factors are useful only if library size reasonably represents technical depth. Here all detected counts happen to follow the same 4-to-1 gene ratio, which is consistent with a simple common-scale difference. In a real dataset, a large total driven by one abundant transcript could instead reflect composition. These three cells are also too few to establish that the scaling assumption holds.
Ratio Methods When Few Genes Are Shared
Ratio methods compare each cell's counts for a gene with a reference value for that gene, then summarize the gene-wise ratios across genes. A common reference is the geometric mean across cells. A gene with a zero in one or more cells has geometric mean zero and cannot supply an ordinary ratio for all cells; implementations therefore use eligibility rules or other zero-handling strategies. In highly sparse data, many genes may be excluded, leaving a small and potentially unrepresentative set.
Worked Example: Ratios Supported by Only Two Shared Genes
In a hypothetical three-cell dataset, the counts are:
| Gene | Cell A | Cell B | Cell C |
|---|---|---|---|
| Gene 1 | 4 | 8 | 12 |
| Gene 2 | 2 | 4 | 6 |
| Gene 3 | 0 | 1 | 0 |
| Gene 4 | 0 | 0 | 1 |
Only Genes 1 and 2 are positive in all three cells. For Gene 1, the geometric mean is \((4\cdot8\cdot12)^{1/3}=7.268\) counts, giving ratios \(4/7.268=0.550\), \(8/7.268=1.101\), and \(12/7.268=1.651\). Gene 2 has geometric mean \((2\cdot4\cdot6)^{1/3}=3.634\) counts and gives the same three ratios. The median ratio across these two eligible genes is therefore approximately \(0.550,\ 1.101,\ 1.651\) for Cells A, B, and C.
The numerical agreement is reassuring within this constructed example, but the estimate rests on only two genes, and both have exactly the same pattern. It offers little protection against a gene-specific anomaly. Genes 3 and 4 cannot contribute to the standard all-positive calculation because they contain zeros. Reporting these factors without also reporting how many genes supported them would conceal a key limitation.
In practice, check the number and identity of genes contributing to each cell's factor, and test whether factors change materially under reasonable eligibility rules. If only a handful of genes remain, a ratio estimate may be too fragile to use as the primary normalization, even if the calculation succeeds.
Pooling: More Counts, With a Different Estimation Target
Pooling combines counts across cells before estimating relative depth, so a gene that is zero in one cell may be observed in a pool. Deconvolution uses multiple overlapping pools to infer cell-level factors from these pooled measurements. Pooling does not make the original cells less sparse; it borrows information across them. A single aggregate pool can support a factor for the aggregate, but cannot by itself identify each cell's separate factor.
Worked Example: Recovering Three Factors From Overlapping Pools
Suppose three hypothetical cells have unknown relative factors \(f_1,f_2,f_3\). For one stable gene, assume an idealized expected count of \(10f_i\) in cell \(i\). Three overlapping pools produce counts of 30 for cells 1 and 2 together, 40 for cells 1 and 3, and 50 for cells 2 and 3. Under this idealized setup, the pooled equations are
Dividing each equation by 10 gives \(f_1+f_2=3\), \(f_1+f_3=4\), and \(f_2+f_3=5\). Adding the first two equations and subtracting the third gives \(2f_1=3+4-5=2\), so \(f_1=1\). Then \(f_2=3-1=2\), and \(f_3=4-1=3\). The overlapping pools have supplied enough distinct equations to recover all three factors in this simplified example.
Real data are noisy, and deconvolution uses many genes and pools rather than treating one gene's counts as exact expectations. The example shows the source of the information gain, not a promise of exact recovery. If there are too few cells to construct informative overlapping pools, too few expressed genes, or factors that are highly variable across pools, pooling may remain unstable. Pool construction and factor uncertainty should be checked rather than assumed away.
When Regression Is Preferable—and When It Is Not
Negative-binomial (NB) regression offers a model-based way to relate gene counts to depth. A typical gene-wise mean specification is
Here \(\mu_{ig}\) is the expected count, \(\phi_g\) is a dispersion parameter, \(\alpha_g\) is a gene-specific intercept, and \(\boldsymbol{x}_i\) contains modeled cell covariates. The term \(\log(L_i)\) is an offset: its coefficient is fixed at one, rather than estimated as an ordinary predictor coefficient. This approach can account for count overdispersion and covariates, but it still depends on a meaningful depth measure and a model that is estimable from the available data.
With very few cells, a gene may have only zeros or one nonzero observation. Such a gene contributes little information about its mean or dispersion, and adding covariates can use up scarce information. A fitted model may return estimates while those estimates remain highly uncertain or sensitive to individual cells. Check convergence warnings, fitted means, residual patterns, and sensitivity to removing one cell. If almost all genes are uninformative, regression is not automatically a remedy for sparsity.
Regression is most attractive when there is adequate information to fit the model and the analyst needs to represent covariates explicitly. Simple scaling is more transparent when library size is a credible depth proxy and the data are not dominated by composition effects. Pooling can help when individual counts are too sparse for reliable ratios but there are enough cells and overlapping pools to share information. Ratio methods are compelling when many stable genes are shared across cells. None of these methods can identify technical depth separately from a genuine global biological shift without additional assumptions or information.
A Practical Troubleshooting Sequence
Summarize \(L_i\), the fraction of zeros, and the number of detected genes per cell. Flag cells for which a small number of counts dominate the total or a one-count change would materially alter the factor.
For ratio methods, record the number of eligible genes. For pooling, examine whether the pool equations are informative and whether factors vary across pools. For regression, check convergence and sensitivity to individual cells.
Estimate factors with methods whose assumptions are defensible for the data. Compare their relative ordering and magnitude; substantial disagreement is a diagnostic signal to investigate, not a vote that one method must win.
Check whether one or a few highly expressed genes dominate totals, whether eligible reference genes are concentrated in particular cell types, and whether differences could reflect true biology. Reconsider the estimand or restrict claims if no stable reference is available.
A useful diagnostic is sensitivity analysis: vary reasonable gene filters, pool configurations, or cell subsets and see whether the resulting factors materially change. Large changes indicate that the data do not strongly determine the normalization. Conversely, agreement between methods does not prove that a shared biological shift is absent, because methods can share the same underlying assumptions.
Common Mistakes and Limits
- Assuming every returned factor is reliable. A method can produce a factor from very little information. Report support, uncertainty where available, and sensitivity to reasonable choices.
- Treating a zero as evidence of no expression. A zero observed count can arise from sparse sampling; it can also reflect genuine absence. Zero handling must not be mistaken for proof of either explanation.
- Using a single pooled total as cell-level information. One aggregate does not identify separate cell factors. Cell-level recovery needs an informative design of overlapping pools or another source of information.
- Choosing the method with the most complex model by default. Extra flexibility can increase estimation variance when the number of cells and nonzero observations is small.
- Calling normalized counts corrected biological expression. A factor adjusts a chosen count scale under assumptions; it does not remove all technical effects or establish that remaining differences are biological.
The practical goal is not to force every sparse dataset into a single normalization recipe. It is to establish whether the available data support a stable relative scale, choose the least fragile method consistent with the assumptions, and carry uncertainty and limitations into downstream interpretation.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A cell has a library size of 25 counts. What is the relative change in its library size if one count is added? Why does that matter for a factor based on total counts?
- A ratio method has only three genes positive in every cell. What should be checked before treating its factors as reliable?
- Why can one aggregate pool not identify separate factors for four cells? What feature of deconvolution addresses this problem?
- In an NB depth model, why is a fixed offset not the same as estimating a depth coefficient?
- Two methods agree on the relative ordering of cells. Does this establish that their factors represent technical depth rather than biology? Explain.
Solutions. (1) The relative change is \(1/25=0.040\), or 4%; a single count can therefore noticeably change the factor. (2) Check whether the genes are a representative, stable reference set and whether factors change under reasonable gene filters; three genes provide little protection from a gene-specific anomaly. (3) One aggregate gives information about a combined quantity, not four separate unknowns; multiple overlapping pools provide distinct equations that can constrain individual factors. (4) The offset contributes \(\log(L_i)\) with coefficient fixed at one, while an estimated coefficient is free to change the depth relationship. (5) No. Agreement is reassuring about method sensitivity but cannot distinguish a shared technical effect from a shared biological shift without a suitable reference assumption or additional information.