Why the Unit of a Count Matters
A count matrix can look like a table of ordinary numbers: rows represent genes, columns represent cells, and each entry is a nonnegative integer. But the numbers do not all count the same thing. Depending on the assay and the processing procedure, an entry may count distinct UMI-defined molecule proxies, reads, or fragments assigned to a gene. The distinction is essential: a value of 20 has no complete interpretation until its counting unit is known.
This tutorial focuses on reading matrix entries in the context of assay design. A UMI-based value is generally a count of distinct retained UMI groups associated with a gene and cell; a non-UMI value is derived from assigned sequencing reads or fragments. Neither value is automatically a direct measurement of the absolute number of RNA molecules originally present in a cell.
A UMI is a sequence tag attached during library preparation to help distinguish captured molecules. After reads are grouped and deduplicated according to the workflow, several reads supporting the same gene, cell, and UMI generally contribute one to the UMI count—not one count for every read. In a non-UMI assay, there is no such UMI-based molecule deduplication. Multiple reads or fragments may derive from the same original RNA molecule, so the resulting entry is read- or fragment-derived.
“UMI count” is a practical counting unit, not a guarantee that every original molecule in the cell was detected exactly once. Molecules can be lost before sequencing, and a molecule may fail to yield a usable UMI observation. UMI collisions, sequencing errors, and processing choices can also affect the final count. The count is therefore best understood as a deduplicated molecular proxy under a specified assay and workflow.
Reading an Entry in Its Assay Context
To interpret a value, establish four things: which assay produced the matrix, whether it uses UMIs, what the pipeline counts after processing, and which gene and cell the entry indexes. For paired-end data, for instance, a pipeline may count a read pair as one fragment rather than count the two read ends separately. The word “read count” alone may not resolve that distinction.
The earlier material on UMI deduplication and read counts in non-UMI plate-based assays provides the assay-specific background. Here, the practical rule is to state the unit explicitly whenever describing a value: “18 UMI groups for this gene in this cell,” or “120 assigned fragments,” rather than simply “18 counts” or “120 counts.”
Worked Example: Several Reads Support One UMI Count
Suppose a hypothetical droplet-based UMI workflow has assigned 47 reads to one gene in one cell. After the workflow groups reads by cell, gene, and corrected UMI sequence, the reads support 18 distinct retained UMI groups. What is the matrix entry if the matrix records UMI counts?
The unit is a distinct retained UMI group. Therefore the entry is 18, not 47. The 47 reads provide evidence for the 18 groups; they are not 47 separate UMI-counted observations.
The average number of supporting reads per retained UMI group is \(47/18 \approx 2.611\) reads per UMI group. This describes the sequencing support in this example; it does not change the matrix entry from 18 to 47, nor does it establish that exactly 18 RNA molecules were originally present in the cell.
Worked Example: Read Ends Versus Paired-End Fragments
Consider a hypothetical non-UMI library. A gene has 120 assigned paired-end fragments, and each fragment was sequenced from both ends. If the matrix’s documented counting convention is one count per fragment, what entry should be reported?
There are 120 fragments, so the entry is 120 fragments. Those fragments comprise \(120 \times 2 = 240\) read ends. Reporting 240 would be appropriate only if the matrix were explicitly defined to count read ends; it would not follow from a fragment-counting convention.
The arithmetic shows why the unit label matters: the same sequencing data can yield 120 fragment counts or 240 read-end counts under different conventions. Neither number can be identified as the matrix entry without knowing the convention used to construct the matrix.
Why the Units Differ
In a UMI workflow, reads with the same relevant cell, gene, and UMI information are grouped so that repeated sequencing observations of a tagged molecule do not each count separately. In a read-based workflow without UMIs, a gene’s entry instead reflects the number of assigned sequencing observations under the pipeline’s read or fragment convention. A transcript may generate multiple such observations, and their number can depend on amplification and sequencing as well as on the RNA captured.
A simple accounting identity clarifies the distinction. Suppose the detected molecules in a hypothetical UMI workflow are indexed by \(j\), and \(r_j\) is the number of reads supporting detected molecule \(j\). The total supporting reads are the sum of the \(r_j\)'s, while the UMI count is the number of distinct detected UMI groups. If there are \(m\) such groups, then:
Here \(R\) is the total reads supporting the groups and \(U\) is the number of distinct groups. The ratio \(R/U\) is the average number of supporting reads per group when \(U>0\). It need not equal one, and it may vary across genes, cells, libraries, or sequencing runs. Thus, read totals and UMI counts are related through the observations in a workflow, but they are not interchangeable units.
This identity is an accounting description, not a universal model of how read counts arise. In particular, it does not establish a fixed conversion factor from reads to molecules. The number of reads per detected molecule can vary, and read-based workflows may not identify which reads came from the same original molecule.
Comparisons Across Assays
A raw comparison is interpretable only when the compared values have units and measurement processes that support that comparison. For example, dividing a read-derived count by a UMI count produces a numerical ratio, but that ratio is not automatically a biological fold change. Its numerator and denominator count different things.
Worked Example: Why a Sixfold Raw Difference Is Not a Sixfold Expression Difference
Suppose two hypothetical assays examine a similar biological sample. For one gene, a UMI-based matrix reports 40 distinct UMI groups in a cell, while a non-UMI matrix reports 240 assigned fragments in a cell. A reader divides the entries and obtains \(240/40 = 6\). What does this calculation establish?
It establishes only that the displayed read-derived value is six times the displayed UMI-derived value numerically. The numerator is 240 fragments and the denominator is 40 UMI groups, so the ratio has units of fragments per UMI group. It does not establish that expression is six times greater in the non-UMI assay.
The assays can differ in capture, library preparation, amplification, sequencing depth, and counting procedure. The 240 fragments are not known to represent 240 distinct molecules, while the 40 UMI groups are not a complete census of all original RNA molecules. Without a justified measurement model and appropriate harmonization, the raw ratio cannot be interpreted as a ratio of biological expression levels.
The same caution applies even when two matrices both report UMI counts or both report read-derived counts. Matching labels make the units more alike, but do not by themselves ensure comparable measurements. Assays can differ in capture efficiency, sequencing depth, gene assignment, cell selection, and other processing choices. A cross-assay comparison needs to account for the design and processing differences relevant to the scientific question; the unit label alone is not sufficient.
| Matrix entry type | What the entry commonly counts | What it does not establish by itself |
|---|---|---|
| UMI-based | Distinct retained UMI groups for a gene and cell | The exact total number of original RNA molecules in that cell |
| Read-based | Assigned reads or fragments under the pipeline’s convention | The number of distinct original RNA molecules that produced them |
| Either type | Counts produced under a defined assay and processing workflow | Direct comparability with a value from a different workflow or counting unit |
Diagnostics and Common Failure Modes
When the unit is unclear, do not infer it from the appearance of the matrix. Integer values are consistent with both molecule-proxy counts and read-derived counts. Instead, check the assay description and the documentation for the matrix-generation pipeline. Look for whether UMIs were used and deduplicated, and whether paired-end reads were counted as individual ends or as fragments.
Warning signs of an interpretation problem include descriptions that call every matrix entry a “molecule” without establishing UMI-based counting, comparisons that divide read counts by UMI counts and label the result an expression fold change, or reports that omit whether the entries count read ends or fragments. If metadata and documentation disagree or leave the unit unspecified, the safe conclusion is that the counting unit is unresolved—not that one convention can be guessed from the values.
A common mistake is to reason, “This gene has 240 reads and another has 40 UMIs, so the first has six times as much RNA.” The arithmetic \(240/40=6\) is correct, but the interpretation is not: reads and UMI groups are different units, and the counts arise through different measurement processes. State the two units, report the raw comparison only as a numerical comparison of those entries, and avoid converting it into a biological expression ratio without a justified method.
Check Your Understanding
Answer each question using the stated counting convention. A solution follows each problem.
- A UMI-based matrix entry is 32 for a gene in a cell. What is the direct interpretation, and what should not be claimed? Solution: The entry represents 32 distinct retained UMI groups assigned to that gene and cell under the workflow. It should not be described as a guaranteed exact count of all original RNA molecules in the cell.
- A fragment-counting pipeline assigns 75 paired-end fragments to a gene. How many counts does the matrix contain under that convention, and how many read ends are represented? Solution: The matrix contains 75 fragment counts. The fragments represent \(75 \times 2 = 150\) read ends.
- One assay reports 50 UMI groups for a gene and another reports 200 assigned reads. Is the second value evidence of fourfold higher expression? Solution: No. The arithmetic ratio is \(200/50=4\), but it compares reads with UMI groups. The result is not, by itself, a fourfold expression difference.
- Two matrices both report read-derived counts. Does that alone prove that their entries are directly comparable? Solution: No. The label does not establish that the assays, read-versus-fragment conventions, sequencing depths, or processing procedures are comparable. Those details must be checked before interpreting differences.