Relative Abundance and Absolute Cell Number
A cell type can make up a larger share of a sample while becoming less numerous. This is not a paradox: a share depends both on the number of cells of interest and on the number of all other cells in its denominator. In single-cell RNA sequencing, the data commonly provide a profile of captured cells. That profile can support comparisons of relative abundance, but it does not, by itself, reveal how many cells were present in the original tissue.
The distinction matters whenever an analysis is interpreted biologically. A higher observed proportion may reflect an increase in the cell type of interest, a decrease in other types, or both. An absolute claim requires an abundance scale tied to a defined biological unit, not just the relative composition of captured cells.
The unit and denominator must be explicit. “Thirty percent of cells” is incomplete unless it is clear which cells are counted: all cells in a tissue, all cells passing a sorting gate, or all cells captured and assigned a cell type. These quantities need not be the same. In particular, the proportion among sequenced cells is not automatically the proportion in the original tissue. The earlier material on capture bias explains why cell types can differ in their probability of being recovered.
How a Proportion Can Rise While a Count Falls
The definition \(p_k=A_k/T\) makes the denominator’s role direct. If both \(A_k\) and \(T\) change between conditions, the direction of the change in \(p_k\) need not match the direction of the change in \(A_k\). The following hypothetical counts illustrate this.
Worked Example: A Higher Share but Fewer Cells
Suppose a defined tissue unit contains 20 cells of a target type and 80 other cells under condition A. Under condition B, it contains 15 target cells and 35 other cells. Assume for this example that these are complete counts from the same-sized biological unit.
Under condition A, the total is \(20+80=100\) cells, so the target proportion is \(20/100=0.20\), or 20%. Under condition B, the total is \(15+35=50\) cells, so the target proportion is \(15/50=0.30\), or 30%.
The target type’s share rises by 10 percentage points even though its count falls by 5 cells, a 25% decrease relative to its count in condition A. The share rises because the total number of cells falls more sharply than the target count. It would be correct to report a higher target proportion in B and a lower absolute target count per defined unit. It would be incorrect to infer an increase in target cell number from the proportion alone.
More generally, suppose \(p_{k,1}=A_{k,1}/T_1\) and \(p_{k,2}=A_{k,2}/T_2\). A rise in proportion means \(A_{k,2}/T_2>A_{k,1}/T_1\). This inequality does not imply \(A_{k,2}>A_{k,1}\), because the totals \(T_1\) and \(T_2\) may differ. Absolute count comparisons therefore require information about the totals, or another measurement on an absolute scale.
Why Relative Data Do Not Determine the Scale
A vector of proportions records composition but discards the overall scale. If every cell count is multiplied by the same positive constant, the proportions remain unchanged. This is a simple but important identifiability result: composition alone cannot distinguish, for example, 20 target cells among 100 total cells from 200 target cells among 1,000 total cells.
The justification follows by substitution: the common multiplier appears in both the numerator and denominator and cancels. No statistical procedure applied only to the same proportion vector can recover the missing multiplier without additional assumptions or data. A model may impose assumptions that relate relative composition to absolute numbers, but those assumptions do not turn the observed proportions alone into direct absolute measurements.
Worked Example: Identical Proportions, Different Counts
Consider two hypothetical samples with the same composition: 30% target cells and 70% other cells. In sample A, the total is 100 cells; in sample B, the total is 500 cells.
For sample A, the target count is \(0.30\times100=30\) cells, and the other-cell count is \(0.70\times100=70\). For sample B, the target count is \(0.30\times500=150\) cells, and the other-cell count is \(0.70\times500=350\).
The observed proportions are identical, while the target count is five times as large in sample B. If only the proportions were available, these absolute counts could not be distinguished. This is why a relative-abundance comparison cannot, on its own, establish that a cell type’s absolute number is stable across samples.
What External Measurements Can Support an Absolute Claim?
To estimate an absolute count, a relative proportion needs to be paired with information that anchors the total or directly measures the target count. The measurements must refer to the same biological sampling unit, include a compatible set of cell types, and be made in units that can be related to the proportion. Examples include a validated total-cell count for the sampled tissue, a calibrated flow-cytometry count using counting beads, or a quantitative imaging or stereological estimate of cells in a defined volume or tissue section. Each method has its own sampling and measurement assumptions; none is automatically a gold standard in every setting.
If \(\hat p_k\) estimates the target fraction in the same population counted by an external method, and \(\hat T\) estimates the total number of those cells in the defined unit, then a natural estimate is \(\hat A_k=\hat p_k\hat T\). This product is meaningful only if the fraction and total refer to compatible units and populations. For example, multiplying the fraction among viable cells by a total that includes dead cells does not estimate the target count among viable cells.
Worked Example: Combining a Proportion With a Total Count
Suppose, hypothetically, a representative assay estimates that a target type is 0.30 of all cells in a defined tissue sample. A separate calibrated measurement estimates 50,000 total cells in that same sample and cell population. For the calculation, assume the proportion estimate is representative of those counted cells and the total measurement is correctly calibrated.
The estimated target count is the estimated proportion multiplied by the estimated total:
The supported interpretation is an estimated 15,000 target cells in that defined sample, conditional on the compatibility and measurement assumptions. It is not automatically an estimate of cells per milligram, per organ, or per individual: those claims require the sampling unit and scaling measurements to support them. Nor does the calculation remove uncertainty in either input. If both \(\hat p_k\) and \(\hat T\) are estimated, their uncertainty should be reflected in uncertainty for the product.
Capture Counts and Tissue Proportions Are Not Interchangeable
A sequenced-cell proportion is usually calculated among cells that were captured, passed quality control, and received an annotation. Let \(c_k\) denote a type-specific capture probability in a simplified setting with \(A_k\) original cells of type \(k\). The expected captured count for that type is \(c_kA_k\), when each original cell is independently captured with probability \(c_k\). If counts are large and the capture process is suitably stable, the captured share is often approximated by \(c_kA_k/\sum_j c_jA_j\). In a deterministic expected-count model, that ratio is the share of the expected counts. In general, however, the expected value of a random captured share need not equal the ratio of expected captured counts.
Worked Example: Why a Random Captured Share Can Differ
Consider a hypothetical sample with one cell of type 1 and one cell of type 2. Type 1 is captured with probability 1, while type 2 is captured with probability 0.5. Condition on at least one cell being captured, which is guaranteed here because type 1 is always captured.
With probability 0.5, both cells are captured and type 1’s share is \(1/2\). With probability 0.5, only type 1 is captured and its share is 1. Therefore the expected captured share of type 1 is
The expected captured counts are 1 for type 1 and 0.5 for type 2. Their normalized ratio is \(1/(1+0.5)=2/3\), not 0.75. This small example shows why the random share’s expectation cannot generally be obtained by normalizing expected counts. For larger counts, the ratio of expected counts can be a useful approximation under suitable conditions, but it is not an exact identity for every capture process.
Even a well-estimated captured proportion does not by itself solve this problem. If capture probabilities differ by cell type, the captured composition may differ from the tissue composition. Recovering tissue proportions requires either a credible measurement model for the relevant capture probabilities or an external method that measures composition more directly. A total-cell measurement also needs to match the population represented by the proportion; multiplying an uncorrected captured share by a tissue total does not generally yield a valid tissue count.
Assumptions, Diagnostics, and Limits of Interpretation
Before translating a relative result into an absolute one, check the assumptions behind that translation. At minimum, the target cell type must be defined consistently across samples; the proportion must refer to a clear and compatible denominator; and the external total or direct count must correspond to the same sample unit and cell population. If observed captured proportions are used as estimates of tissue proportions, the capture and processing steps must not introduce unaccounted cell-type-specific differences, or those differences must be addressed with defensible calibration.
Useful diagnostics are practical rather than decisive proofs of validity. Confirm that total counts were measured for each relevant sample, in the same tissue compartment and under comparable protocols. Check whether cell recovery, viability, tissue dissociation, or quality-control filtering could selectively alter cell types. Compare the estimated totals with recorded input material and any independent count available. Large discrepancies across methods or conditions are a warning to investigate differences in units, sampling, or recovery, not a reason to choose whichever number supports a preferred conclusion.
Relative-abundance models remain appropriate when the scientific question concerns composition, such as whether a cell type’s share differs across conditions. Absolute counting or a calibrated proportion-times-total estimate addresses a different estimand: the number of cells in a defined unit. These approaches are complementary, not competing versions of the same measurement. A relative analysis may be statistically strong and still not establish an absolute increase or decrease.
An absolute estimate also has limits. It may describe the sampled tissue rather than the whole organ or organism. If the external count is itself estimated, its uncertainty, sampling error, and calibration error matter. When \(\hat p_k\) and \(\hat T\) are treated as random estimates, the variance of their product depends on both variances and their covariance; treating one or both as exact can understate uncertainty. More fundamentally, a count difference between conditions is an association unless the study design and assumptions support a causal interpretation.
Common Mistake
Wrong reasoning: “The target type went from 20% to 30%, so there must be more target cells in the second condition.”
Correction: The proportion’s denominator may have changed. In the worked example, the target count fell from 20 to 15 while the total fell from 100 to 50, producing a higher share. State the measured result as a relative change unless a compatible absolute measurement supports a count claim.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A target type changes from 12 cells among 60 total cells to 10 cells among 30 total cells. What happens to its proportion and absolute count?
- Two samples both have a target proportion of 0.25. Can their target counts be inferred to be equal? Explain.
- A researcher multiplies the proportion among captured, quality-controlled cells by a total count that includes all cells in the tissue, including cells that were not eligible for capture. What compatibility issue should be considered?
- In the two-cell capture example, what is the expected share of type 1, and why is it not equal to the ratio of expected counts?
Solutions. (1) The first proportion is \(12/60=0.20\); the second is \(10/30=0.3333\). The proportion rises by about 13.33 percentage points, while the absolute count falls by 2 cells. (2) No. For example, a proportion of 0.25 corresponds to 25 target cells in 100 total cells or 250 in 1,000; the proportion does not identify the total scale. (3) The proportion and total may refer to different cell populations. The product is not justified unless the fraction represents the same population counted by the total, or a valid adjustment makes them compatible. (4) The expected share is \(0.5(1/2)+0.5(1)=0.75\). The ratio of expected counts is \(1/(1+0.5)=2/3\); expectation of a random ratio is not generally the ratio of expectations.