Why a Global RNA Shift Is Hard to See
Endogenous size factors use the measured genes themselves to estimate differences in exposure across cells. This is useful when most genes provide a stable reference, but it creates a fundamental ambiguity: if the RNA content of a cell changes broadly, endogenous counts can rise or fall together in much the same way as counts would change with sequencing depth. A size factor may then treat a biological shift as a technical one.
The key diagnostic question is not simply whether one group has larger libraries. It is whether the difference is better explained by sequencing exposure, a broad change in RNA per cell, or both. Cell-level summaries can identify patterns worth investigating, but a total-count difference alone cannot distinguish these explanations. External controls and experimental context can provide information that endogenous counts do not contain.
A simplified model makes the ambiguity explicit. Let \(Y_{ig}\) be the observed count for gene \(g\) in cell \(i\), and let \(\mu_g\) be its baseline abundance. Suppose, for illustration, that
Here \(q_i\) combines effective capture and sequencing exposure, while \(r_i\) represents a global biological shift. This is a simplified mean model, not a claim that every gene must respond identically or that observed counts are free of sampling noise. The same expected count is obtained from many pairs \((q_i,r_i)\) with the same product. For any positive constant \(c\), replacing \(q_i\) by \(cq_i\) and \(r_i\) by \(r_i/c\) leaves \(q_i r_i\) unchanged. Endogenous counts alone therefore cannot identify the two components without additional assumptions or information.
What Endogenous Summaries Can and Cannot Tell You
Useful cell-level summaries include total endogenous counts, the number of detected genes, and distributions of counts across genes. Compare these summaries across the relevant experimental conditions, and, where possible, within batches or other design strata. Look at distributions across cells rather than relying only on a single group average: a shift in the median, spread, or proportion of unusually high-count cells may reveal patterns hidden by an average.
These summaries are diagnostic clues, not direct measurements of total RNA per cell. A larger endogenous library can result from greater sequencing exposure, greater capture efficiency, more RNA, or a combination. The number of detected genes also depends on sequencing depth and on how RNA abundance is distributed across genes. Differences in cell quality, cell size, cell type, and technical processing can alter these summaries. Consequently, a pattern in endogenous counts alone can raise a hypothesis about a global RNA shift, but it cannot establish that explanation.
Worked Example: Endogenous Counts Confound Exposure and RNA Content
Suppose two hypothetical cells have total endogenous counts of 1,000 and 2,000. Assume for this example that their gene profiles have the same relative composition and that the counts are summarized without sampling error. If the baseline total abundance is \(M\), the expected endogenous total is \(q_i r_i M\). The observed totals therefore have a ratio of
This ratio is compatible with twice the sequencing exposure and unchanged RNA content: \(q_2=2q_1\), \(r_2=r_1\). It is also compatible with unchanged exposure and twice the RNA content: \(q_2=q_1\), \(r_2=2r_1\). More generally, it is compatible with any combination satisfying \(q_2r_2=2q_1r_1\).
An endogenous normalization procedure that assumes the broad gene profile is a stable reference may estimate a factor of about 2 for the second cell. That factor describes the observed multiplicative difference; it does not, by itself, tell us whether the difference arose from exposure or RNA content. The data given are therefore insufficient to diagnose the cause.
Using External Controls to Add Information
An external control can help break the ambiguity if it is introduced in a way that makes its expected amount comparable across the cells being compared. For example, a known amount of spike-in RNA may be added per known cell before lysis. Let \(Z_i\) be the observed spike-in count and \(c\) the known spike-in amount. Under the simplifying assumptions that endogenous RNA and spike-ins experience the same effective exposure \(q_i\), and that the spike-in amount is constant per cell, the expected counts are
where \(Y_{i\cdot}\) is the total endogenous count and \(M\) is the baseline total endogenous abundance. The ratio of these expected counts is
The common exposure term cancels. Thus, under these assumptions, a change in the endogenous-to-spike-in ratio is evidence of a change in endogenous RNA relative to the fixed external amount, rather than merely a change in sequencing exposure. In actual data, the ratio of observed counts is noisy, especially when counts are small; it is a diagnostic quantity, not an exact measurement of the biological multiplier.
Worked Example: A Shift Relative to a Fixed Spike-In Amount
In a hypothetical experiment, a fixed amount of spike-in RNA is added per cell before lysis. One cell has 1,000 endogenous counts and 100 spike-in counts; another has 2,000 endogenous counts and 100 spike-in counts. The observed endogenous-to-spike-in ratios are
The ratio has doubled. If spike-ins and endogenous RNA share effective exposure, and the spike-in amount and processing are comparable, this pattern supports a twofold increase in endogenous RNA relative to the fixed spike-in amount. It is not enough to say that the second cell was sequenced twice as deeply: under the model, a twofold exposure increase would also increase the expected spike-in count.
This interpretation depends on how the control was used. If spike-ins were added per unit of total RNA rather than per cell, their amount would not provide a fixed per-cell reference for this question. If spike-ins and endogenous RNA have different capture behavior, or if control addition varies across samples, the cancellation argument may fail. The control’s preparation and timing are part of the evidence, not administrative details.
Worked Example: A Larger Library Explained by Exposure
Consider a different hypothetical comparison. The first cell has 800 endogenous counts and 80 spike-in counts. The second has 1,600 endogenous counts and 160 spike-in counts. The endogenous-to-spike-in ratios are
Both count types doubled, while their ratio stayed constant. Under the shared-exposure and fixed-spike-in assumptions, this pattern is consistent with a twofold increase in effective exposure and provides no evidence, from these ratios, for a change in endogenous RNA relative to the spike-in. It does not prove that RNA content is identical: noise, differences in control behavior, or other unmeasured factors could obscure a change. It does show why an increase in endogenous library size alone is not decisive.
Combine Cell-Level Patterns with Experimental Context
A convincing diagnosis draws on evidence with different sources of information. First, inspect cell-level distributions of endogenous totals, detected genes, and endogenous-to-control ratios, stratified by condition and technical batch. Second, check how the external control was added and whether it plausibly shares the relevant technical exposure with endogenous RNA. Third, ask whether independent measurements or the experimental design support a per-cell RNA change. Examples include a separate assay of RNA amount per known cell, a validated cell-size measurement, or a perturbation with a well-supported mechanism that changes global transcription.
Worked Example: Triangulating a Global Shift Across Cells
Suppose a hypothetical, batch-matched experiment has four cells in each of two conditions. The median endogenous totals are 1,000 counts in the reference condition and 2,000 counts after treatment. The median spike-in counts are 100 in each condition. For a simple summary using these medians, the ratio of median endogenous counts to median spike-in counts is
The ratio of these summaries is \(20/10=2\). Suppose, in addition, that the cell-level endogenous-to-spike-in ratios are generally higher in treated cells, detected-gene counts also tend to be higher, and an independent assay reports greater RNA amount per cell after treatment. The spike-in ratio and independent measurement support a broad RNA-content change; the endogenous summaries show that it is visible across cells. The detected-gene pattern is consistent with the diagnosis but is not, on its own, independent confirmation because it is also sensitive to exposure.
The reasonable conclusion is that a global RNA-content shift is supported by several converging observations under the stated control assumptions. The data do not establish that every gene changed by exactly twofold, nor do they prove that no technical differences remain. If this shift is a real treatment response, normalizing it away would change the scientific target: the analysis would focus on expression relative to the cell-wide RNA scale, not on the total RNA response per cell.
Assumptions, Diagnostics, and Failure Modes
The simple control-based reasoning relies on assumptions that should be checked explicitly. A fixed control amount must be added per comparable number of cells, and control and endogenous molecules must experience sufficiently similar processing for exposure to cancel in their ratio. Cell recovery, lysis, and sample handling should not systematically differ by condition in ways that alter one source more than the other. The treatment groups should also be compared within, or adjusted for, relevant technical batches; otherwise, a batch-specific shift can masquerade as a condition effect.
Look for inconsistent evidence. For example, a total-count increase without a change in endogenous-to-spike-in ratios is compatible with increased exposure. A ratio increase that appears only in one poorly performing batch deserves investigation. A few extreme cells may drive a group mean, so inspect cell-level distributions and robust summaries. Conversely, a stable group median does not rule out a shift confined to a subpopulation. Sparse counts and sampling variability can make individual ratios unstable, particularly when their denominators are small.
External controls are not automatically a perfect reference. Their addition can vary, they may not behave like endogenous transcripts, and their counts can be too sparse for precise cell-level ratios. An orthogonal RNA assay also has its own measurement assumptions. Agreement among methods strengthens a diagnosis; disagreement is a reason to investigate what each method measures, not to select whichever result is preferred.
Choosing the Interpretation, Not Just the Factor
Endogenous size factors, total-count scaling, and external controls answer related but not identical questions. Endogenous factors estimate relative exposure under assumptions about the stability of the endogenous reference. Total-count scaling uses each cell’s endogenous total and can therefore absorb a global shift into the scale. A suitably designed external control can provide a reference outside the endogenous transcript pool, but its interpretation depends on its addition and processing. Cell-level summaries help reveal patterns; they do not independently identify their cause.
The choice is ultimately about the target of the analysis. If the scientific question concerns expression composition relative to the cell’s total endogenous RNA, removing a global scale difference may be appropriate. If the question concerns RNA amount per cell, treating that difference as a nuisance can erase a meaningful response. State which target is intended and preserve the evidence needed to distinguish biological change from technical exposure.
Inspect endogenous totals, detected genes, and their distributions across cells; do not infer RNA content from a total-count difference alone.
Confirm whether the external control amount is fixed per cell and whether it is expected to share effective exposure with endogenous RNA.
Assess control ratios alongside experimental context and any orthogonal measurements of RNA amount or cell size.
Decide whether the analysis should preserve a global per-cell RNA response or express gene abundance relative to a cell-wide scale.
Check Your Understanding
Answer each question using the assumptions stated in the question. Solutions follow the questions.
- Why can endogenous counts alone fail to distinguish a twofold increase in sequencing exposure from a twofold global RNA-content increase?
- In one cell, the endogenous total is 900 and the spike-in count is 90. In another, the corresponding counts are 1,800 and 180. What happens to the endogenous-to-spike-in ratio, and what explanation is supported under the shared-exposure assumptions?
- Why is a higher number of detected genes not sufficient evidence of a global RNA-content shift?
- What is a key design requirement if spike-ins are to help assess RNA content per cell?
- If a global RNA shift is a biological treatment response, what could happen if normalization removes that shift?
Solutions. (1) The expected endogenous count depends on the product \(q_i r_i\). Either component can double while producing the same expected counts, so endogenous observations do not separate them without additional information. (2) Both ratios equal \(900/90=10\) and \(1{,}800/180=10\). The unchanged ratio is consistent with increased exposure rather than a detected increase in endogenous RNA relative to the control; it does not prove that RNA content is unchanged. (3) Detection depends on exposure and on the distribution of abundance across genes as well as RNA content, so more detected genes can result from deeper measurement. (4) The control amount should be fixed per known, comparable number of cells, and it should experience sufficiently similar processing to endogenous RNA for exposure to cancel in the ratio. (5) The analysis may erase a real per-cell RNA response and instead report expression relative to the cell-wide RNA scale.