Why Pool Design Matters
Pooling combines counts from several cells before estimating normalization factors. It can make genes detectable in the aggregate even when they are absent from most individual-cell profiles. But a pool set has two jobs: it must provide informative aggregate measurements, and its membership patterns must retain enough information to distinguish the cells contributing to those measurements. Larger pools can help with the first job while making the second harder.
The aim here is to reason about the construction of the pools themselves. The central design choices are pool size, how often each cell appears, and how pool memberships overlap. The exact later calculation that turns pooled measurements into cell-level factors is a separate step.
Pools are overlapping when a cell appears in more than one pool. Overlap is useful because different combinations of cells create distinct equations about their contributions. Merely repeating the same pool membership does not create a new membership pattern: it may provide another noisy measurement, but it does not by itself add a new independent equation.
A simple model clarifies what the pool sums can convey. Suppose the expected count for gene \(g\) in cell \(i\) is \(q_g s_i\), where \(q_g\) represents the gene's baseline abundance and \(s_i\) is the cell-specific exposure or factor. By additivity of counts and expectations,
Thus, under this model, a pool's expected exposure is the sum of the cell exposures in that pool. This interpretation requires a common gene-abundance profile \(q_g\), up to the cell-specific exposure factors. If cell types or states differ substantially in their gene-abundance profiles, an aggregate can reflect that biological mixture as well as exposure. Pool construction should therefore account for known groups rather than silently treating composition differences as technical depth differences.
Pool Size and Gene Detection
The detection benefit of pooling can be quantified under a simple count model. Suppose a particular gene has independent Poisson counts across the cells in a pool, with mean \(\lambda_i\) in cell \(i\). The pooled count is also Poisson, with mean equal to the sum of the cell means. A gene is detected in the pool if its pooled count is at least one.
If all cells have the same per-cell mean \(\lambda\) and the pool contains \(m\) cells, this becomes \(1-e^{-m\lambda}\). The probability rises with pool size, especially when the gene is rare. The calculation assumes the Poisson model and independence for this particular probability expression; biological overdispersion or dependence can change the detection probability. The broader point—that summing nonnegative counts makes a pooled gene count more likely to be nonzero—does not require Poisson counts.
Worked Example: How Pool Size Changes Detection
Suppose a hypothetical low-abundance gene has a Poisson mean of \(0.08\) counts per cell, and consider pools of 4 and 12 cells. Assume equal means across cells and independent counts for this calculation.
For a pool of 4 cells, the mean pooled count is \(4(0.08)=0.32\). Its detection probability is \(1-e^{-0.32}=1-0.7261=0.2739\), or about 27.4%.
For a pool of 12 cells, the mean is \(12(0.08)=0.96\). The detection probability is \(1-e^{-0.96}=1-0.3829=0.6171\), or about 61.7%.
The larger pool is more likely to yield a nonzero count for this gene. This does not mean that every pool should be as large as possible: a set made only of very large pools may not contain enough distinct membership information to separate the cells' contributions.
Pool size also affects what can be learned from the design. When many cells are combined, the measured sum can be informative about their total exposure but does not directly say how that total is divided among them. Overlap helps by placing cells into different combinations. For a fixed number of pools, however, increasing pool size can make membership patterns more similar and can reduce the design's ability to distinguish individual contributions.
Membership Patterns and Recoverability
In the idealized exposure model, define the pool exposure \(t_p=\sum_i A_{pi}s_i\). In vector form, the pool exposures satisfy \(t=As\), where \(s\) is the vector of cell factors. If the pool exposures were known exactly, a necessary design condition for uniquely determining an unrestricted vector of \(n\) cell factors is that \(A\) have column rank \(n\). This requires at least \(n\) pools, and the membership patterns must be sufficiently distinct. It is not a guarantee of accurate estimation from noisy gene counts: near-redundant columns can make the problem unstable even when the rank is technically full.
This rank check concerns the idealized sums, not every detail of pooled-ratio estimation. In actual data, pool exposures are not observed directly; they are inferred from gene counts and reference information. Sparse counts, uncertain reference values, and differences in biological composition add further uncertainty. Full rank is therefore a basic design check, not a promise that the resulting factors will be precise.
Worked Example: Checking a Four-Cell Pool Design
Consider four hypothetical cells and the four pools \(S_1=\{1,2\}\), \(S_2=\{2,3\}\), \(S_3=\{3,4\}\), and \(S_4=\{1,2,3\}\). The membership matrix is
To check whether these rows provide four independent constraints, subtract row 1 from row 4. Expanding the resulting determinant along the first column leaves a \(3\times3\) determinant:
Since the determinant is nonzero, \(A\) has rank 4. The design has overlapping memberships and, in the idealized sum model, can distinguish four cell factors if the four pool exposures are known. For example, if the hypothetical cell factors are \(1,\ 2,\ 1.5,\ 0.5\), the pool sums are \(3,\ 3.5,\ 2,\ 4.5\), respectively. These sums illustrate the measurements implied by the design; they are not themselves a complete procedure for estimating factors from observed counts.
A design can fail this check even when every cell appears in several pools. For instance, if two cells always occur together, their columns in \(A\) are identical. Their individual contributions cannot be separated using pool sums alone. Conversely, a cell appearing in only one pool may be distinguishable in some designs, but its estimate can be especially sensitive to noise in that one measurement. Balanced coverage—each cell appearing a similar number of times—is a useful practical goal, alongside full rank.
Constructing Ratios from Pooled Counts
Pool membership also determines which counts are combined to construct pooled gene ratios. If target gene \(g\) and reference gene \(r\) are both counted in pool \(p\), the observed pooled ratio is \(Y_{gp}/Y_{rp}\), provided the reference count is positive. Under the shared-profile model, expected target and reference counts scale with the same pool exposure. Their ratio can therefore provide information about relative abundance that is less directly tied to the pool's total exposure. A ratio of observed counts is noisy and is not generally equal to a ratio of expectations; the model explains its purpose, not an exact identity for every observed pool.
Worked Example: Ratios in Overlapping Pools
Suppose a hypothetical five-cell dataset is organized into pools \(S_1=\{1,2,3\}\) and \(S_2=\{3,4,5\}\), so cell 3 is shared. The observed aggregate counts for a target gene and a reference gene are as follows.
| Pool | Target count | Reference count | Target/reference ratio |
|---|---|---|---|
| \(\{1,2,3\}\) | 9 | 45 | 9/45 = 0.200 |
| \(\{3,4,5\}\) | 6 | 24 | 6/24 = 0.250 |
For the first pool, the ratio is \(9/45=0.200\); for the second, it is \(6/24=0.250\). Both denominators are positive, so both observed ratios are defined. The different values can arise from count noise or from differences in the cells' gene composition, as well as from other departures from the shared-profile model. Because the pools overlap at cell 3, their measurements are not independent: they share that cell's counts. Any later analysis that treats the ratios as independent would need justification.
Practical Design Checks and Alternatives
A useful pool set balances three features: detection, coverage, and distinctness. Pools must be large enough to make sparse genes more informative, but varied enough in membership to support cell-level separation. Before using a design, inspect the membership matrix and the proposed pool counts.
- Check coverage. Count how often each cell appears. Very uneven coverage can leave some cell factors much less informed than others.
- Check distinctness. Look for identical membership patterns or cells that always occur together. Compute the rank of \(A\), and consider whether its columns are nearly redundant.
- Check pool sizes. A range of pool sizes can create useful variation in summed exposures, but extremely large pools can blur cell-specific information.
- Check biological composition. If possible, construct pools within relevant groups or balance group membership across pools. Otherwise, changing cell-type composition can be confused with changing exposure.
- Check reference counts. A target/reference ratio is undefined when the pooled reference count is zero. Very small denominators also make ratios highly variable.
A single large pool is an alternative when the only goal is to obtain a more detectable aggregate profile, but it cannot distinguish individual-cell factors. Disjoint pools avoid shared-cell dependence, yet each cell contributes to only one aggregate and the design may not provide enough independent constraints. Overlapping pools offer richer membership patterns, at the cost of correlated measurements and greater design complexity. These options are not interchangeable: choose the design for the information required, not solely for the largest possible pooled count.
Common Mistake and Limitations
A common mistake is to reason, “A gene is more detectable in a larger pool, so larger pools must make individual-cell factors easier to recover.” The first clause can be correct while the conclusion is not. Detection concerns whether an aggregate count is nonzero; recovery concerns whether the pool memberships distinguish the cells. A large pool that always contains the same cells can improve detection while adding little or no new information about how the total is divided among them.
Pooling also cannot repair a wrong normalization target or eliminate biological differences that violate the shared-profile assumption. Rank does not diagnose all problems: even a full-rank design can be unstable if counts are extremely sparse, ratios have small denominators, or the incidence patterns are nearly redundant. Pool design supports estimation; it does not remove the need to assess model fit and uncertainty.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Under the equal-mean Poisson model with \(\lambda=0.05\) per cell, what is the detection probability for a pool of 10 cells?
- Why do two pools with identical memberships fail to provide two distinct membership patterns?
- A pool has target count 12 and reference count 0. What is its observed target/reference ratio, and what design or data issue does this reveal?
- What does full column rank of the incidence matrix establish, and what does it not establish?
- Why can overlapping pool ratios be statistically dependent?
Solutions. 1. The pooled mean is \(10(0.05)=0.5\), so the detection probability is \(1-e^{-0.5}=1-0.6065=0.3935\), about 39.35%. 2. Their incidence-matrix rows are identical, so the second row adds no new independent membership equation. 3. The ratio is undefined because division by zero is undefined; the reference is absent in that pool, so the ratio cannot be used without a suitable alternative treatment. 4. Full column rank means the idealized exact pool sums uniquely determine an unrestricted vector of cell factors; it does not guarantee accurate estimates from noisy counts or validate the shared-profile model. 5. The pools share one or more cells, and the same cell counts contribute to both aggregates, so their measurements can be correlated.