Aggregating Cells Into Donor-Level Count Tables

Why Aggregate Cell Labels?

A cell-level data set may contain thousands of rows, but a differential-abundance comparison usually asks whether biological samples differ in their cell-type composition. The relevant replication is generally the donor or biological sample, not each cell: cells from one donor share biological and technical context. Counting cells by cell type within each sample creates a data structure in which samples can be compared without pretending that every cell is an independent replicate.

Aggregation does not discard the cell-level labels; it summarizes them for a particular analysis. The summary is useful only if the aggregation unit, cell-type categories, and denominator are explicit. In particular, a donor who contributes samples under multiple conditions must not have those samples collapsed together when the comparison concerns condition.

The term “donor-level” is appropriate when each donor contributes one sample to the comparison. With repeated samples, the rows should instead identify the donor and the sample occasion or condition. The donor identifier remains essential metadata because repeated rows from one donor are not automatically independent.

Constructing the Count Table

Begin with one record per retained cell, including a sample identifier and a cell-type label. The sample identifier must map each cell to the biological sample from which it came. Define the cell types before counting, and use the same category definitions across samples. If the analysis includes an “Other” category, for example, it must be defined consistently rather than used as an informal receptacle for labels that differ across samples.

For each sample \(i\) and cell type \(k\), count the records satisfying both conditions. Equivalently, if \(z_c\) is the assigned cell type for cell \(c\), and \(s_c\) is its sample, then

where \(I(\cdot)\) equals 1 when the condition is true and 0 otherwise. This definition makes clear that every retained cell contributes to exactly one table entry if its sample and cell-type labels are both valid and each cell has exactly one assigned type. The row total \(N_i\) is the number of retained cells from sample \(i\) represented in the table.

Metadata should be stored at the same row level as the counts. Typical fields include donor ID, condition, time point, batch, and any sample-level covariates relevant to the planned comparison. Metadata values should describe the sample, not be inferred from cell-level repetitions without checking that they are consistent. Keep a stable sample ID even when donor ID and condition together appear unique; a stable ID makes joins and auditing safer.

Worked Construction and Checks

The following short justification explains why these counts summarize cell labels without changing the sample as the comparison unit. For a fixed sample, each cell belongs to one category, so the indicators across categories sum to 1 for that cell. Summing first across cells and then across categories gives

Thus the table partitions the retained cells in sample \(i\), and its row total is the number of cells assigned to one of the listed categories. Aggregation yields one count vector per biological sample. It does not make cells from that sample independent biological replicates.

Metadata, Denominators, and Unequal Cell Yield

The total \(N_i\) is the number of cells included in the table for sample \(i\), not necessarily the number originally captured, sequenced, or collected. It depends on filtering, quality control, annotation, and the chosen set of cell types. Therefore, record the inclusion rules and retain the totals. A table without totals can obscure whether a count of 20 arose from 25 observed cells or 2,000.

The denominator for an abundance summary must correspond to the stated estimand. If the question concerns a cell type as a share of all retained cells in a sample, its proportion is \(C_{ik}/N_i\). If the analysis intentionally excludes specified categories from the denominator, define the resulting denominator and report it. A count table allows these choices to be made transparently; it does not determine which choice is scientifically appropriate.

Unequal totals are not automatically errors: samples can yield different numbers of usable cells. But they affect the precision of sample-level summaries, and extreme differences can make comparisons fragile. Do not solve unequal yield by pooling all cells in a condition and treating the resulting large total as if it represented many independent donors. The independent sample count remains the number of biological units available for comparison.

Diagnostics and Common Failure Modes

Several checks catch aggregation mistakes before analysis:

• Row-total check: verify \(N_i=\sum_k C_{ik}\) and compare \(N_i\) with the number of eligible cell records for that sample.
• Category check: confirm that the table has the same cell-type columns for every sample. A missing category should usually appear as a zero count, not as a missing row or absent column.
• Identifier check: each sample ID should map to one donor, condition, and set of sample-level metadata. Investigate duplicated IDs or conflicting metadata rather than silently selecting one record.
• Label check: look for spelling variants, unexpected labels, or cells assigned to multiple categories. Such problems can inflate totals or split one biological category across columns.
• Yield check: inspect totals across samples for zeros and unusually small or large values. A zero total means no retained cells are represented, so sample proportions are undefined; an unexpectedly small total may make the estimated composition unstable.

A frequent mistake is to count each cell as an observation in a group comparison. The reasoning is that thousands of cells provide thousands of replicates. The correction is that those cells are nested within a smaller number of donors or samples; they share sample-level conditions and are not independent replicates of those conditions. The count table makes the biological replication visible by placing counts in rows indexed by samples, but the subsequent analysis must still account for pairing, repeated sampling, or other dependence when present.

Aggregation is also not a substitute for checking the study design. If condition is perfectly confounded with processing batch, a table cannot separate those effects. If one donor contributes samples under several conditions, collapsing by donor loses the comparison, while treating the rows as independent ignores within-donor dependence. These design features must be preserved in metadata and handled by an appropriate analysis.

What Aggregation Does—and Does Not—Provide

The table is therefore a data-preparation step, not a complete inferential method. Its main advantage over pooled counts is preserving one count vector per biological sample. Its advantage over proportions alone is retaining the integer counts and the total from which each proportion was calculated. The later choice among statistical models depends on the sampling structure, the target comparison, and features such as overdispersion; aggregation alone does not resolve those issues.

The earlier material on donors as independent units in abundance tests motivates keeping biological replication explicit. The practical consequence here is simple: build the table at the unit that the comparison requires, retain donor identity even when a donor has multiple rows, and never let the number of cells stand in for the number of donors.