Experimental Units in Single-Cell Genetic Studies

Why the Unit of Replication Matters

A single-cell dataset can contain hundreds of thousands of measurements while representing only a modest number of people. That distinction matters in genetic analyses because the amount of data is not the same as the amount of independent biological replication. A study may measure many cells from each donor, but inherited genotype is a property of the donor: the cells from one person do not provide new, independently sampled genotypes.

The first task is therefore to identify what was independently sampled or assigned, what was measured, and what was repeated. These are related but not interchangeable units. Naming them clearly helps explain where genetic information enters the analysis and which sources of variation a result can reflect.

Definition: An experimental unit is the smallest unit independently assigned to an experimental condition. In an observational genetic study, the relevant biological replicate is typically an independently sampled donor. An observational unit is the unit on which a measurement is recorded; in single-cell data, it may be a cell. A technical replicate is a repeated measurement or processing of the same biological material, rather than a new independent biological sample.

These definitions depend on the design. For an observational study of inherited genetic effects in a cell type, the donor is generally the independent unit for genotype-related comparisons: genotype is measured on, or inherited by, that donor. Cells from the donor are nested observations that help describe the donor’s cells. If an experiment instead independently randomizes wells or cultures to a treatment, the well or culture may be the experimental unit for that treatment effect. The unit should be identified from how samples were obtained or assigned, not from which data table has the most rows.

Donors, Cells, and Technical Replicates

Suppose researchers collect a tissue sample from each donor, identify a cell type, and measure gene expression in many cells. The donor-level biological variation can include differences in genetic background, age, exposures, and other donor characteristics. Cells from the same donor can also differ from one another because of cell state, stochastic expression, and local biological conditions. Finally, measurement and processing introduce technical variation, including sequencing noise, library-preparation differences, and run or batch effects.

A useful conceptual decomposition for one measurement from cell \(j\) of donor \(i\) is:

Y_{ij} = \mu + u_i + v_{ij} + e_{ij}

Here, \(Y_{ij}\) is the measured value, \(\mu\) is an overall mean, \(u_i\) represents a donor-specific deviation, \(v_{ij}\) represents cell-to-cell deviation within that donor, and \(e_{ij}\) represents measurement-level technical variation. This is a variance-accounting framework, not a claim that every analysis must use this exact model. In particular, a shared library or sequencing batch can affect multiple cells together and should not be treated as independent cell-specific noise.

For an inherited genetic association, the genotype value is shared across the cells of a donor. A conceptual mean model might therefore include a donor’s genotype as a predictor, while recognizing that cells from that donor share both genotype and donor-level characteristics. The genetic comparison is supported by variation in genotype across donors, not by counting each cell as another donor. The number of cells can still matter: it can improve the description of the cell-type measurements for each donor. It does not create additional independent genotype assignments.

Worked Example: Counting the Units in a Cell-Type Study

Consider a hypothetical study with four unrelated donors. Researchers collect 300 cells of a specified cell type from each donor. The donors’ genotypes at a variant are 0, 0, 1, and 2, where the values count copies of a designated allele. How many biological replicates and cell observations are present for a genetic comparison?

There are \(4\) biological replicates: the four independently sampled donors. The total number of cell observations is:

4 \text{ donors} \times 300 \text{ cells per donor} = 1{,}200 \text{ cells}.

The genotype values occur on four donors. In particular, the 600 cells from the two donors with genotype 0 do not represent 600 independently observed genotype-0 donors. They are cells nested within two donors. Thus, the dataset has 1,200 cell measurements but only four donor-level genotype observations. The cell count describes how many cells were measured; it does not replace the count of biological replicates.

How the Sources of Variation Relate

The decomposition above also clarifies why measurements from cells belonging to the same donor are generally related. Under a simple model, suppose donor deviations \(u_i\), cell deviations \(v_{ij}\), and technical errors \(e_{ij}\) have mean zero, are mutually independent, and have variances \(\tau^2\), \(\omega^2\), and \(\sigma^2\), respectively. These are simplifying assumptions for the derivation; actual data can require additional terms, such as batch effects.

For two distinct cells \(j\) and \(k\) from the same donor \(i\), the donor deviation is shared. Their covariance is:

\begin{aligned} \operatorname{Cov}(Y_{ij},Y_{ik}) &= \operatorname{Cov}(u_i+v_{ij}+e_{ij},\;u_i+v_{ik}+e_{ik}) \\ &= \operatorname{Var}(u_i) \\ &= \tau^2. \end{aligned}

The second equality follows because the cell-specific and technical terms for the two distinct measurements are independent under the stated model, leaving only the shared donor term. The variance of one measurement is \(\tau^2+\omega^2+\sigma^2\). Therefore, the within-donor correlation under this model is:

\operatorname{Corr}(Y_{ij},Y_{ik}) = \frac{\tau^2}{\tau^2+\omega^2+\sigma^2}.

This expression is not a universal description of every single-cell dataset; it follows from the specific assumptions just stated. Its practical lesson is that shared donor-level variation can make cells from one donor more alike than cells from different donors. More cells do not erase that shared structure. The later analysis of pseudoreplication and effective sample size develops the consequences for inference; here, the essential design question is simply to distinguish cells from independent donors.

Worked Example: Interpreting a Variance Decomposition

Suppose a hypothetical variance-components summary for a gene’s measured expression in one cell type reports donor variance \(0.36\), cell-level variance \(0.49\), and technical variance \(0.15\), all in squared normalized-expression units. Under the independent-components model above, find the total variance and the correlation between two distinct cells from the same donor.

The total variance is the sum of the three components:

$$ 0.36 + 0.49 + 0.15 = 1.00 $$

The shared donor component is the covariance between distinct cells from the same donor in this model. Thus:

\operatorname{Corr}(Y_{ij},Y_{ik}) = \frac{0.36}{1.00} = 0.36.

The implied within-donor correlation is \(0.36\), a dimensionless quantity. It does not mean that 36% of cells are “correlated,” nor does it estimate a genetic effect. It says that, under this variance model, the donor component accounts for 0.36 of the total measurement variance and induces positive similarity between two cells from the same donor. The cell and technical components still contribute to the observed spread.

Technical Variation and Design

Technical replication is useful for understanding or reducing measurement uncertainty, but it is not a substitute for biological replication. For example, splitting one donor’s sample into two library preparations can reveal how much results vary across preparations. It still gives one donor’s biological material and one donor-level genotype. Likewise, sequencing the same library on two runs may help assess run-to-run variability, but it does not add another biological sample.

A technical difference can also become a design problem when it is entangled with the genetic comparison. If all donors with one genotype are processed in one batch and all donors with another genotype in a different batch, a measured difference could reflect genotype, batch, or both. Merely recording the batch label cannot separate effects that the design has made inseparable. Whenever feasible, samples from different genotype groups should be distributed across processing batches, with donor identity retained throughout the workflow.

Worked Example: Separating Donor Replication From Repeated Processing

In a hypothetical study, eight donors are sampled: four have genotype 0 and four have genotype 1. Each donor’s sample is split between two library preparations, and 100 cells of the target type are measured from each preparation. Assume that every donor contributes to both preparations, so neither preparation is exclusive to one genotype group. How many biological donors, library preparations, and cell observations are in the design?

The biological replicate count is \(8\), because there are eight independently sampled donors. The number of library preparations is:

8 \text{ donors} \times 2 \text{ preparations per donor} = 16 \text{ preparations}.

The total cell-observation count is:

8 \text{ donors} \times 2 \text{ preparations per donor} \times 100 \text{ cells per preparation} = 1{,}600 \text{ cell observations}.

These totals answer different questions: \(8\) donors provide the biological replication, \(16\) preparations provide repeated technical processing, and \(1{,}600\) is the number of cell observations. The balanced splitting also means both genotype groups are represented in both preparations, rather than being assigned to separate preparation groups. This does not by itself guarantee that all technical biases are absent, but it avoids the simplest form of genotype–preparation confounding.

A Practical Classification

Before interpreting a cell-type genetic result, describe the design in terms of its levels. A concise account should state who or what was independently sampled, which measurements were made on each unit, and which processing steps were repeated. It should also make clear whether genotype varies between donors or was assigned to some other independently randomized unit.

Donor: usually the biological replicate and genotype-bearing unit in an observational study of inherited genetic effects.
Cell: a nested observational unit that provides information about cellular measurements within a donor.
Library, assay, or sequencing run: a technical level that can contribute measurement variation; it is not automatically an independent biological replicate.
Batch: a processing condition that may affect many observations together and may be confounded with genotype if the design is not balanced.

The word “usually” matters. The correct unit follows the sampling and assignment structure. If an intervention was independently assigned to cultures, for example, its experimental unit may be a culture rather than the donor from whom cells originated. For an inherited variant measured in donors, however, additional cells from the same person do not supply additional independent genotype values.

Common Mistakes

Wrong reasoning: “There are 20,000 cells, so the genetic study has 20,000 independent replicates.” Correction: Count the independently sampled donors for a donor-level genetic comparison. Cells are nested measurements, even when there are many of them.
Wrong reasoning: “Two libraries from each donor double the biological sample size.” Correction: They are repeated technical preparations of the same donor’s material. They can inform technical variability but do not create new donors.
Wrong reasoning: “All observed variation is either genetic or random noise.” Correction: Distinguish donor-level differences, cell-to-cell heterogeneity, and technical variation; also consider shared effects such as processing batch.
Wrong reasoning: “Adjusting for batch always solves batch confounding.” Correction: If genotype and batch are perfectly aligned, their separate effects cannot be distinguished from the observed data alone. A design that distributes genotype groups across batches is needed to make that comparison possible.

Key takeaway: In cell-type-resolved genetic studies, distinguish the number of donors from the number of cells and technical replicates. Donors generally supply the independent biological genotype information; cells and technical repeats describe measurements at nested levels and contribute their own sources of variation.

Check Your Understanding

For each question, identify the relevant unit or explain the design implication.

A study measures 500 cells from each of 12 donors. How many cell observations and how many donor-level biological replicates are there?
Why does measuring more cells from a donor not create more independent genotype observations?
Name one source of variation that can differ among donors, one that can differ among cells from the same donor, and one technical source of variation.
A sample from each of six donors is split into three library preparations. What do the 18 preparations represent, and what remains the biological replicate count?
All donors in one genotype group are processed in batch A, and all donors in the other group in batch B. Why can the genetic and batch differences not be cleanly separated by this design?