Why Replicate Labels Matter
A single-cell experiment may produce thousands of cell profiles, several libraries, and multiple sequencing runs. Those counts describe different levels of observation and processing; none automatically tells us how many independent biological units the study represents. A study with many measured cells can still have little biological replication if those cells came from only a few donors.
The relevant distinction is not simply whether two measurements are numerically different. It is whether they arise from distinct biological units relevant to the question, or are repeated measurements of the same biological material. That distinction affects how much new information a measurement adds about biological variation and what population a conclusion could reasonably address.
In many human single-cell studies, donors are the biological units of greatest interest because conclusions are intended to describe variation across people. Distinct donors then provide biological replication. By contrast, separately preparing two libraries from the same cell suspension usually creates technical replicates: the preparation is repeated, but the donor is not. A different study may target variation across independently collected tissue specimens rather than across donors. Its specimen-level replication count can differ from its donor-level count.
What Counts at Each Stage?
The following distinctions are useful, but they must be interpreted in light of how material was collected and what the study is meant to represent.
| Study material or measurement | Typical role | Question to ask |
|---|---|---|
| Distinct donors | Biological replicates for donor-population questions | Were these different people or animals sampled independently for the target population? |
| Distinct specimens from one donor | Distinct biological material, but not new donor replicates | Does the target concern tissues or locations, and are they independent sampling units for that target? |
| Separate libraries from one specimen or cell suspension | Usually technical replicates | Was the same biological material divided and processed more than once? |
| Repeated sequencing of one library, such as across lanes | Technical measurements | Was the prepared library measured again, rather than new biological material collected? |
| Cells from a donor-derived specimen | Cell-level observations within biological material | Does the question concern individual cells, or variation across donors? |
A specimen is not automatically a biological replicate merely because it has a separate tube or barcode. Two tissue pieces from one donor contain distinct material and may be informative about within-donor tissue variation. They do not, however, provide two independent donors when the target is a population of people. Similarly, separately cultured wells may be biological replicates if each was independently initiated or independently assigned under the design; wells seeded from the same culture can instead be repeated measurements of a shared source. The provenance of the material matters.
Worked Example: Counting Replicates Across Donors, Libraries, and Cells
In a hypothetical study, tissue from four donors is dissociated once per donor. Each donor’s cell suspension is split into two libraries, and each library yields 800 passing cell profiles. The investigators want to describe how expression varies across donors.
There are four biological replicates for the donor-level question: the four distinct donors. There are eight libraries and 6,400 cell profiles, but neither count changes the number of independently sampled donors. The two libraries from a donor are repeated processing of that donor’s material, and the cells provide detailed measurements of that material.
For a donor-level analysis, the relevant biological replicate count is therefore \(n=4\), not \(n=8\) or \(n=6{,}400\). The extra libraries may help assess library-preparation variability, and the cells may describe within-specimen heterogeneity. Neither supplies additional independent donors. This conclusion assumes the four donors were independently sampled from the population relevant to the question; if they were not, the scope of generalization must be narrower.
Why Technical Repeats Do Not Replace Biological Replication
A simple model shows what technical repeats can and cannot accomplish. Suppose \(Y_{ij}\) is a measured summary from biological unit \(i\) on technical repeat \(j\). Write \(Y_{ij}=\theta_i+\varepsilon_{ij}\), where \(\theta_i\) is the underlying value for biological unit \(i\), and \(\varepsilon_{ij}\) is technical measurement error. Assume biological units are independently sampled, with \(\operatorname{Var}(\theta_i)=\tau^2\); technical errors have mean zero, variance \(\sigma_e^2\), and are independent across repeats and independent of \(\theta_i\).
If \(r\) technical measurements are made on the same biological unit, their average is \(\bar{Y}_i=\theta_i+\bar{\varepsilon}_i\). Independence and zero-mean error give \(\operatorname{Var}(\bar{\varepsilon}_i)=\sigma_e^2/r\). Thus:
Averaging technical repeats reduces the technical-error contribution, but the biological variation \(\tau^2\) remains. With \(m\) independently sampled biological units, the variance of the average of their technical-repeat averages is \((\tau^2+\sigma_e^2/r)/m\), under the same assumptions. Increasing \(r\) reduces only the \(\sigma_e^2/r\) term; increasing \(m\) reduces both terms in the variance of the overall average. This is the key reason technical replication cannot substitute for sampling additional biological units.
Worked Example: Quantifying the Contribution of Technical Repeats
Suppose, hypothetically, that a gene-expression summary has biological variance \(\tau^2=9\) expression-units squared across independently sampled units, and technical measurement error has variance \(\sigma_e^2=4\) expression-units squared. Compare one technical measurement with the average of four technical measurements on each unit.
With one measurement, the variance across measured biological units is \(9+4=13\) expression-units squared. With four independent technical measurements per unit, the variance of each unit’s technical average is \(9+4/4=10\) expression-units squared. The corresponding standard deviations are \(\sqrt{13}=3.606\) and \(\sqrt{10}=3.162\) expression units, respectively.
The repeat measurements improve precision by reducing the technical contribution from 4 to 1 expression-units squared. They do not remove the biological variance of 9 expression-units squared. If the study still includes only four independently sampled biological units, it still has four biological replicates, even if each is measured four times.
Replicate Counts Depend on the Biological Question
The same collection of material can support different descriptions, but the replicate count must be named alongside the target. If the target is variation across people, distinct donors are the relevant independent units. If the target is variation among tissue locations, multiple separately collected locations may be relevant biological samples, while their shared donor identity still matters when describing donor-population evidence. If the target is repeatability of library preparation, libraries prepared from the same material can be the technical repeats of interest.
This distinction also matters in experiments with paired conditions. If each donor contributes a control specimen and a treated specimen, the design has two condition measurements per donor, not twice as many independently sampled donors. The specimens are biologically distinct in condition, but they share a donor. Counting them as unrelated people would discard the pairing and misdescribe the source of replication.
Worked Example: Paired Conditions With Duplicate Libraries
A hypothetical experiment collects control and treated specimens from each of six donors. Each specimen is split into two libraries, and each library yields 500 cell profiles. The scientific question concerns the difference between conditions across donors.
There are six donors, each contributing a pair of condition specimens: 12 specimens in total. Each of the 12 specimens has two libraries, so the experiment has 24 libraries and 12,000 cell profiles. For a comparison of conditions across donors, the biological replication is six donor pairs. The two libraries per specimen are technical repeats; they do not add donor pairs.
A suitable description is “six paired donors, with duplicate libraries per condition specimen.” Reporting only “24 replicates” would confuse library count with biological replication. This classification assumes each donor’s control and treated specimens are genuinely paired as described and that the donors represent the population to which the intended conclusion refers.
Diagnostics, Alternatives, and Limitations
Before interpreting a replicate count, trace each observation back through the experimental record. A useful audit records the donor or source, specimen, cell suspension, library, and sequencing run for every measurement. Then ask which units were sampled independently, which were split from shared material, and which were measured repeatedly. If the study design cannot answer these questions, the replication structure is uncertain and should not be inferred from the number of rows in an expression matrix.
- Look for repeated identifiers. Many cell barcodes linked to the same donor indicate many cell observations, not many donors.
- Check material provenance. Separate library identifiers may trace to aliquots of one suspension rather than independently collected specimens.
- Check condition allocation. If all treated material was processed in one batch and all control material in another, a condition difference may be inseparable from a batch difference.
- Report counts at each level. State numbers of donors, specimens, libraries, and cells separately, and make clear which count is the biological replication for the question.
Several practical approaches address technical repeats, but they answer different questions. Averaging technical replicates can reduce measurement noise when repeat errors are reasonably independent and measurements are on a comparable scale. Keeping the repeats separate can help describe technical variability or identify a failed preparation, but it does not make them independent biological units. Collecting more independent donors or independently assigned biological units adds biological replication, although it costs more and may be constrained by recruitment or experimental resources.
No labeling rule resolves every design. Replicates can share sources, such as specimens from the same donor, or share processing batches. Technical errors may be correlated, in which case averaging \(r\) repeats need not reduce their variance to \(\sigma_e^2/r\). The simple model above assumes independent errors and a common variance; it is an explanation of the distinction, not a guarantee that every experiment follows that model. Accurate classification therefore requires the study’s sampling, allocation, and processing history.
Check Your Understanding
For each question, identify the relevant replication and compare your reasoning with the solutions.
- A study uses three donors, prepares two libraries from each donor’s cell suspension, and measures 1,000 cells per library. How many donor-level biological replicates are there?
- Four tissue regions are collected from each of five donors. For a question about differences across donors, how many donor-level biological replicates are there, and why are the 20 regions not 20 independent donors?
- A sequencing library is run across two lanes. Are the lanes new biological replicates? Explain.
- In the variance model \(Y_{ij}=\theta_i+\varepsilon_{ij}\), what does increasing the number \(r\) of technical repeats reduce, and what does it leave unchanged?
- Each of eight donors contributes one control and one treated specimen, and each specimen is processed into three libraries. State the donor-level replication and the library count.
Solutions. (1) There are three donor-level biological replicates. The six libraries and 6,000 cells are additional measurements of material from those donors. (2) There are five donor-level biological replicates. The 20 regions are distinct specimens but are grouped within five donors, so they do not represent 20 independently sampled people. (3) No. Two lanes measure the same prepared library, so they are technical repeats of sequencing. (4) Under the stated independent-error assumptions, increasing \(r\) reduces the technical component \(\sigma_e^2/r\); it leaves the between-unit component \(\tau^2\) unchanged. (5) There are eight paired donors for the donor-level condition comparison. There are \(8\times2\times3=48\) libraries, but the library count does not change the eight donor pairs.