Why the Source of Variation Matters
A gene’s observed count in a single cell is not a direct readout of a perfectly measured biological quantity. Even if the cell’s underlying molecular state were held fixed, only some molecules may be captured and recorded. At the same time, cells from the same sample can genuinely differ in the amount of that gene’s RNA. The first source creates molecule-sampling variability; the second is cell-to-cell biological variation.
These sources answer different questions. Repeating a measurement on the same biological material asks how much the recorded count can fluctuate when the underlying cell state is held fixed. Measuring different cells asks how much the cells differ in their biological states, as well as how much their counts fluctuate during measurement. Counts remaining variable after cell and contamination quality control is not, by itself, evidence that the variation is biological.
The distinction is useful throughout single-cell analysis: it affects what a count can support about an individual cell, how replicated measurements should be interpreted, and whether observed heterogeneity might reflect biology or measurement. It is a conceptual distinction, not a claim that every experiment can cleanly separate the sources.
A Simple Model of a Measurement
The symbol \(\theta_i\) is deliberately general. It could describe a cell’s underlying abundance for a gene, but it is not itself the observed count. The observed count depends both on that state and on the process by which molecules are captured and recorded. The relationship need not be the same across cells if, for example, capture efficiency or sequencing depth differs.
A simple illustration treats each of \(n\) molecules as having the same probability \(q\) of being captured and recorded, independently of the others. If \(Y\) is the number recorded, then \(Y\) is a sum of \(n\) independent zero-or-one capture indicators. Under these assumptions, \(Y\) has a binomial distribution, with expected count \(nq\) and variance \(nq(1-q)\). This is an illustrative sampling model, not a universal model for an observed single-cell count. In particular, real count generation can involve additional processes and constraints.
Here \(I_k=1\) if molecule \(k\) is recorded and \(I_k=0\) otherwise. The mean follows by adding the expected values of the indicators. Because the indicators are independent, the variance of their sum is the sum of their variances; each has variance \(q(1-q)\). Thus, even with the cell’s molecular state and capture probability fixed, repeated measurements need not produce identical counts.
Worked Example: Repeated Measurements at a Fixed Capture Probability
Suppose, hypothetically, that a fixed piece of biological material contains \(100\) eligible molecules for a gene, and each is independently recorded with probability \(q=0.20\). Treat the number of eligible molecules and the capture probability as fixed across repetitions. What mean and standard deviation of counts does the binomial model predict?
The model assumptions are that each molecule has the same capture probability, capture events are independent, and the material being measured does not change between repetitions. Under those assumptions,
The model therefore predicts an average of \(20\) recorded molecules and a standard deviation of \(4\) across repeated measurements. A count of \(24\) on one run would not imply that the underlying material necessarily changed: it is four counts above the model mean, a difference that can arise from sampling. The model does not say that every count must be close to \(20\), nor does it establish that the assumptions hold in an actual assay.
Repeated Measurements and Repeated Cells
A repeated measurement is informative about sampling only insofar as the biological material is effectively held fixed and the repetitions are independent in the relevant way. In practice, RNA measurements are destructive, so a literal second measurement of an unchanged cell may not be available. Repeated runs on aliquots of the same material can approximate the comparison, but aliquoting, library preparation, and sequencing can introduce additional differences. Shared material or shared processing can also make repetitions dependent.
Repeated cells answer a different question. They are distinct biological units, so their underlying states may differ. Even if all cells came from one sample and were handled similarly, they need not have identical expression. Conversely, counts from distinct cells can differ because of cell-specific capture or depth. A spread of counts across cells therefore combines potential biological differences with measurement variability; it cannot identify either source by itself.
Worked Example: Comparing Within-Material and Between-Cell Spread
Imagine a hypothetical experiment with three cells. Each cell’s material is measured twice in independent technical runs, and the gene counts are:
| Cell | Measurement 1 | Measurement 2 | Cell’s mean count |
|---|---|---|---|
| A | 9 | 11 | 10 |
| B | 18 | 22 | 20 |
| C | 29 | 31 | 30 |
Assume, for this illustration, that each pair of measurements concerns the same cell material, the technical runs are independent, and the runs use comparable procedures. The measurement differences within cells are \(2\), \(4\), and \(2\) counts. The sample variance for two measurements \(a,b\) is \((a-b)^2/2\), so the three within-cell sample variances are \(2\), \(8\), and \(2\) count-squared, respectively. These are noisy estimates because each is based on only two measurements.
The cell means are \(10\), \(20\), and \(30\) counts. Their sample mean is
Their sample variance is
The cell means are clearly separated in these data, whereas the paired measurements within each cell are closer together. This pattern is consistent with substantial differences between the cells’ underlying states relative to the observed repeat-to-repeat differences. It does not prove that the separation is biological: systematic differences in capture or processing between cells could also contribute. Nor do two repeats per cell establish a precise technical variance.
Averaging independent repeated measurements can make a fixed material’s measured count more stable. If each repeat has variance \(v\), and the repeats are independent with the same expected value, then their average has variance \(v/r\) for \(r\) repeats. The result follows because the variance of a sum of independent measurements is the sum of their variances, and dividing the sum by \(r\) divides its variance by \(r^2\). Shared sources of variation can violate independence, so this reduction should not be assumed automatically.
Worked Example: Why Unequal Depth Can Mimic Cell Differences
Consider two hypothetical cells with the same underlying amount of a gene’s RNA and the same capture probability, \(q=0.10\). Suppose a simplified model gives cell A \(n=100\) eligible molecule opportunities and cell B \(n=200\). If the opportunities are independent, each with capture probability \(0.10\), their expected recorded counts are:
The corresponding binomial variances are \(100(0.10)(0.90)=9\) and \(200(0.10)(0.90)=18\) count-squared. Thus the model predicts different raw-count means even though the assumed underlying amount and capture probability are the same; the number of opportunities differs. The hypothetical model’s assumptions are explicit, but real sequencing depth is not always equivalent to a known number of independent opportunities. The example demonstrates why a larger raw count in one cell cannot automatically be interpreted as greater biological expression. Measurement depth and other technical conditions must be considered.
What Repeated Data Can and Cannot Tell You
The most direct comparison uses both kinds of repetition: multiple measurements of the same material and measurements from multiple cells. Repeated measurements provide information about variability under a fixed biological state, subject to the technical differences between runs. Repeated cells provide information about observed heterogeneity among cells. Separating those contributions requires a design and assumptions that connect the repeats to the sources of variation; simply computing one variance across all observed counts does not do so.
| Data being compared | Main question | Important qualification |
|---|---|---|
| Repeated measurements of the same material | How much do recorded counts change when the underlying material is intended to be fixed? | Run-to-run technical changes and dependence can also affect the differences. |
| Different cells measured once | How different are the observed counts across cells? | Differences can reflect biology, sampling, depth, or other technical factors. |
| Different cells, each measured repeatedly | How do within-cell measurement differences compare with between-cell differences? | Reliable separation needs enough repetitions and appropriate comparability assumptions. |
This distinction is related to, but not identical with, the idea of technical and biological variability in an experimental design. A technical replicate repeats a measurement on the same biological material; a biological replicate is a distinct biological unit. Repeated cells from one sample may reveal cell-to-cell heterogeneity, but they do not automatically substitute for independent biological samples when the scientific claim concerns variation across donors or experiments.
Diagnostics, Pitfalls, and Limitations
Before interpreting count spread, examine how the counts were generated and what was actually repeated. Ask whether repeated measurements used the same material, whether they shared a library or processing step, and whether cells had comparable sequencing depth and quality. Also consider whether the cells represent the same relevant biological population: a mixture of cell states can generate between-cell differences that are real but not variation within a single state.
- Calling all spread biological. A wide distribution of counts across cells does not isolate cell-to-cell biology. Sampling and technical differences can also widen it.
- Calling all spread sampling noise. Similar processing does not make distinct cells biologically identical. Genuine cell-state differences may be present.
- Treating technical repeats as new cells. Repeated runs on the same material do not increase the number of biological units. Counting them as independent cells exaggerates the amount of independent biological information.
- Comparing raw counts across unequal depths. A larger count can reflect more measurement opportunity rather than greater underlying abundance. A raw-count difference alone does not distinguish these explanations.
- Assuming the simple capture model is exact. The binomial illustration requires independent, equal-probability capture trials and a fixed number of opportunities. Real data can depart from these assumptions, so the formula is not a universal description of count variability.
A further limitation is that technical replication is often unavailable for individual cells. When only one count is observed per cell, repeated cells do not reveal what that same cell would have produced under another independent measurement. In that setting, interpretation depends on design information, relevant measurements of technical variation, and explicit modeling assumptions. The observed counts alone cannot label each fluctuation as biological or technical.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In a fixed-material binomial capture model with \(n=50\) and \(q=0.20\), calculate the expected count and variance.
- Why does observing counts of 8 and 12 in two repeated measurements of the same material not by itself show that its biological state changed?
- Three distinct cells have observed counts 5, 10, and 15. Can the differences be attributed entirely to biological variation? Explain.
- What is the main distinction between measuring one cell’s material repeatedly and measuring three different cells once each?
Solutions. (1) The expected count is \(nq=50(0.20)=10\). The variance is \(nq(1-q)=50(0.20)(0.80)=8\) count-squared. (2) Under fixed material, independent capture can produce different counts; the difference of four counts may be sampling variability, though technical changes between runs are also possible. (3) No. The counts may reflect biological differences, molecule sampling, unequal measurement conditions, or a combination; three single observations do not separate the sources. (4) Repeating one material targets variation in recorded counts while its biological state is intended to be fixed. Measuring different cells includes possible differences in their biological states, as well as measurement variability.