Biological Gene-Off States And Non-Detection

Two Processes Can Produce the Same Observed Zero

A zero in a single-cell count matrix means that no molecules of a gene were recorded for that cell. It does not, by itself, say why. The gene may have been biologically off during the relevant period, so no transcript was available to capture. Alternatively, transcripts may have been present but missed because the gene was low-abundance or because capture and sequencing sampled only a small fraction of the available molecules. These mechanisms can produce the same observed value, \(Y=0\), while implying different biology.

The distinction matters when interpreting cell states, estimating the fraction of cells in which a gene is active, or deciding whether a zero indicates a meaningful absence. Earlier material on dropout, sampling zeros, and capture variation establishes that limited sampling can generate many zeros. Here the focus is the additional question: what evidence, especially from repeated observations, can help separate sampling non-detection from a biological on-off process?

The time window in the definition is important. Gene activity can fluctuate, and transcription may occur in bursts. A gene that is silent at one instant may not be permanently inactive, while a short assay window may make intermittent activity look like an off state. Thus, “off” is always relative to a biological timescale and a measurement design.

A Simple Model for the Two Sources of Zeros

Let \(Y_{ij}\) denote the observed count for gene \(g\) in cell \(i\) under measurement \(j\). For a simple model, let \(Z_{ij}=0\) denote a biological off state and \(Z_{ij}=1\) denote an on state. Write \(\pi=P(Z_{ij}=0)\) for the off-state probability in the population being studied. Conditional on being on, let \(q_{ij}=P(Y_{ij}=0\mid Z_{ij}=1)\) be the probability of sampling no molecules. For example, if the on-state count follows a Poisson distribution with captured mean \(\mu_{ij}\), then \(q_{ij}=e^{-\mu_{ij}}\).

This expression adds the probability of being biologically off to the probability of being on and nevertheless recording zero. It assumes that an off state produces a zero count and that, when the gene is on, the zero probability is described by the specified sampling model. It also makes clear why one observed zero cannot reveal its cause: the same total zero probability can arise from different combinations of \(\pi\) and \(q_{ij}\).

In the Poisson example, \(\mu_{ij}\) is the expected captured count, not necessarily the number of transcripts in the cell. It depends on abundance and measurement efficiency. A small \(\mu_{ij}\) makes \(q_{ij}\) large even when the gene is on. If the on-state count is modeled with a negative binomial distribution instead, \(q_{ij}\) is its conditional probability of zero; the same decomposition applies.

Why Repeated Measurements Can Help

A repeated-measure design can provide information beyond a single count if it measures the same biological unit more than once and if the biological state persists across those measurements. A persistent off state predicts repeated zeros. By contrast, if a gene remains on and each measurement independently samples molecules, a zero in one measurement does not force a zero in the next. The distinction concerns the dependence between repeated outcomes, not simply the total number of zeros in a data set.

The following justification shows the pattern under deliberately simple assumptions. Consider two measurements of the same unit. Suppose the biological state is shared and persistent: the gene is off for both measurements with probability \(\pi\), and otherwise it is on for both. When on, assume each measurement has probability \(q\) of a sampling zero, independently conditional on the on state. Then the probability of two zeros is

If instead the state were independently redrawn for each measurement, each measurement would have zero probability \(p_0=\pi+(1-\pi)q\), and the joint probability of two zeros would be \(p_0^2\). Their difference is

The inequality follows because \(0\leq \pi\leq1\) and \(0\leq q\leq1\). Under these assumptions, shared persistence increases the chance of seeing two zeros relative to independently redrawn states. This result is a justification for looking for repeated-zero dependence; it is not proof that any observed cluster of zeros is biological. Shared changes in capture efficiency, shared cell quality, or other sources of dependence can also produce repeated zeros.

What Counts as a Useful Repetition?

In ordinary single-cell RNA sequencing, measuring the same intact cell twice is generally not straightforward: the assay consumes or alters the material being measured. Repeated cells from the same condition are not automatically repeated measurements of the same biological unit. They may help estimate how often zeros occur across a population, but population-level repetition alone does not reveal whether a particular cell was persistently off or repeatedly missed by sampling.

Potentially informative designs include repeated measurements of the same cell with a suitable non-destructive assay, repeated sampling of a shared sample or split material where the interpretation is justified, or measurements across time when the unit and biological state can be tracked. A design should make clear what is repeated: the same cell, a matched group of cells, a sample split, or simply the same experimental condition. Each supports different conclusions.

Assumptions, Diagnostics, and Limits

A repeated-zero interpretation depends on assumptions that should be made explicit before drawing a biological conclusion:

• State persistence: the biological on/off state must remain stable over the interval between measurements. If the gene switches rapidly, repeated measurements may not share a state.
• Appropriate sampling model: the conditional zero probability must reflect abundance and capture for the units being compared. A single average rate can be misleading when cells have different expression levels or efficiencies.
• Conditional independence: after conditioning on state and relevant measurement conditions, the sampling events must be sufficiently independent for the calculation used. Shared technical effects can violate this assumption.
• Comparable measurements: assay conditions, timing, and definitions of a zero should be aligned. Changes in protocol or exposure can change the sampling-zero probability.
• Representative repeated units: the set of units measured repeatedly must support the intended inference. Selective loss or tracking failure can distort the observed pattern.

Useful diagnostics include comparing observed repeated-zero frequencies with those predicted by an on-state sampling model, checking whether zero frequency changes with expected abundance or capture-related measurements, and examining whether paired outcomes are more concordant than the model predicts. These are diagnostic comparisons, not definitive classifications. Large residual disagreement may indicate a missing biological state, an inadequate count model, unmodeled heterogeneity, or technical dependence.

The main alternatives answer different questions. A sampling-only count model attributes zeros to its specified count distribution and capture conditions. A biological on-off model adds a latent state that can persist or vary across observations. A model that adds an extra zero-generating component can describe observed zero frequencies, but a better description of counts does not establish that the component corresponds to biological silence. Repeated-measure designs can add evidence about persistence, while single observations mostly inform the overall zero frequency and its relationship to measured covariates. Model fitting and formal comparisons require assumptions beyond the patterns described here.

Common Mistake: Treating a Zero as Proof of Silence

A common but incorrect argument is: “The count is zero, so the gene was not expressed in that cell.” The count records detected molecules, not the complete transcriptional history of the cell. At low captured abundance, a gene can be on and still produce a zero with high probability. The correction is to state what is observed (“no molecules were detected”) and then assess whether the sampling model, repeated-measure pattern, and biological design provide evidence for an off state.

The reverse overstatement is also a mistake: an excess of repeated zeros does not automatically prove a biological off state. A shared technical limitation, unmeasured variation in abundance, or a misspecified conditional count distribution could create similar patterns. The strongest interpretation is conditional: a pattern supports a persistent biological state relative to specified alternatives and under stated assumptions.