Background Components in CITE-seq Protein Counts

Why Protein-Tag Counts Need a Background Model

In CITE-seq, an antibody-derived tag count is evidence that an antibody-associated oligonucleotide was captured and sequenced. It is not, by itself, proof that the antibody bound its intended protein target. Some counts can arise from nonspecific antibody binding, free or ambient tags, Fc-receptor binding, or other technical processes. The statistical task is therefore not simply to label every positive count as expression. It is to describe how observed counts can contain both target-related signal and background, and to use controls to assess the background component.

This distinction matters especially for low counts. A count of one may be compatible with ordinary background under the assay conditions, while a count far above the control pattern may provide stronger evidence of target-associated signal. Neither conclusion follows from the count alone: it depends on the background model, the quality of the controls, and the experimental context. The earlier material on centered log-ratio normalization concerns a transformation of protein-tag measurements; a transformation does not, by itself, identify which counts are signal and which are background.

Definition: For cell \(i\) and antibody-tag feature \(j\), let \(Y_{ij}\) be the observed nonnegative integer count. The signal-plus-background model represents it as \(Y_{ij}=S_{ij}+B_{ij}\), where \(S_{ij}\) is the count attributable to target-related binding and \(B_{ij}\) is the count attributable to nonspecific or technical background. The two components are latent: the assay observes their sum, not their separate values.

An Additive Count Model

A useful starting model treats the two components as independent Poisson random variables:

S_{ij}\sim \operatorname{Poisson}(\lambda_{S,ij}), \qquad B_{ij}\sim \operatorname{Poisson}(\lambda_{B,ij}), \qquad Y_{ij}=S_{ij}+B_{ij}.

Here \(\lambda_{S,ij}\) and \(\lambda_{B,ij}\) are expected counts for the signal and background components. Both are nonnegative and may vary across cells and features. This model assumes that, conditional on those rates, the two components are independent and that counts within each component follow a Poisson distribution. The Poisson assumption implies that each component's conditional variance equals its mean. It also assumes the rates adequately represent relevant variation; strong extra cell-to-cell variation, dependence, or batch structure can make this simple model inadequate.

The sum has a particularly useful form. Since the components are independent, the probability of an observed total \(y\) is the sum over all possible ways to divide it between them:

\begin{aligned} P(Y_{ij}=y) &=\sum_{s=0}^{y}P(S_{ij}=s)P(B_{ij}=y-s)\\ &=e^{-(\lambda_{S,ij}+\lambda_{B,ij})} \sum_{s=0}^{y} \frac{\lambda_{S,ij}^{s}\lambda_{B,ij}^{y-s}}{s!(y-s)!}\\ &=e^{-(\lambda_{S,ij}+\lambda_{B,ij})} \frac{(\lambda_{S,ij}+\lambda_{B,ij})^y}{y!}. \end{aligned}

The last equality follows from the binomial theorem. Thus \(Y_{ij}\sim \operatorname{Poisson}(\lambda_{S,ij}+\lambda_{B,ij})\), and

E(Y_{ij})=\lambda_{S,ij}+\lambda_{B,ij}, \qquad \operatorname{Var}(Y_{ij})=\lambda_{S,ij}+\lambda_{B,ij}.

This result explains both the value and the limit of the model. The observed count informs the total rate, but one count does not determine its two parts. Without additional information or assumptions, many pairs of rates have the same sum. Controls provide information about background; they do not make the latent division automatically observable.

What Controls Can and Cannot Tell Us

An isotype control is an antibody control chosen to lack the intended specific binding target while resembling relevant properties of the assay antibodies. Its counts can help reveal nonspecific binding patterns. A negative-control feature is a tag or feature expected not to have genuine target-related signal in the cells under study. Counts for such controls can provide another view of the background distribution. Both are most informative when processed alongside the target features under comparable staining, washing, loading, and sequencing conditions.

A control is not a perfect measurement of the background for every target feature. Isotype antibodies may differ from target antibodies in abundance, affinity for unintended binding sites, conjugation, or other properties. Negative controls may have their own feature-specific behavior. Consequently, a control can diagnose background without establishing that the background rate is identical across all antibodies or cells. A cell-specific estimate based on controls further assumes that the controls track the relevant cell-to-cell changes in background.

If a feature is a valid negative control, its target-related signal rate is treated as zero for the purpose of checking the background model. Its observed counts then inform the distribution of \(B_{ij}\). For target features, controls can motivate or calibrate an estimate of \(\lambda_{B,ij}\), but a conclusion about \(\lambda_{S,ij}\) still depends on the suitability of that calibration. In particular, observing a count above a control average is not a proof of specific binding.

Worked Example: Comparing Target Counts With an Isotype Control

Suppose six hypothetical cells have counts for one target tag of \(1, 1, 2, 2, 4, 8\), and counts for a matched isotype control of \(0, 1, 1, 2, 2, 3\). The target counts sum to \(18\), giving a mean of \(18/6=3.0\) counts per cell. The isotype counts sum to \(9\), giving a mean of \(9/6=1.5\) counts per cell.

The difference between these sample means is \(3.0-1.5=1.5\) counts per cell. It is a descriptive contrast, not an automatically valid estimate of the target signal rate. Interpreting it as an average signal would require the isotype to represent the target feature's background on the same scale and in the same cells. The largest target count, 8, is notably above the largest isotype count, 3, but six observations do not establish a universal threshold or show that every target count above 3 is specific.

The comparison is useful as a first diagnostic: the target distribution appears shifted upward relative to this control in these hypothetical data. Further assessment should examine whether that pattern is consistent across cells, whether the control is well matched, and whether cell or batch differences could explain it. A low target count such as 1 remains compatible with background; the observed value alone does not resolve its source.

Using a Background Tail Probability Carefully

If a background rate \(\lambda_B\) is specified, the Poisson model gives the probability of a background count at least as large as an observed value \(y\):

P(Y\geq y\mid \lambda_S=0,\lambda_B) =1-\sum_{k=0}^{y-1}e^{-\lambda_B}\frac{\lambda_B^k}{k!}.

This tail probability describes how unusual the count would be under the stated background-only model. It is not the probability that the count is background, nor the probability that the null model is true. Its interpretation depends on the background rate being appropriate and on the Poisson assumptions being reasonable. If the rate was estimated from a small set of controls, treating it as known also understates uncertainty.

Worked Example: Assessing a Count Under a Specified Background Rate

Suppose, hypothetically, that a relevant control supports a background rate of \(\lambda_B=1.5\) counts per cell for a particular feature, and one cell has \(Y=5\). Under the background-only model, \(Y\sim\operatorname{Poisson}(1.5)\). The probability of observing at least 5 counts is

\begin{aligned} P(Y\geq 5\mid \lambda_B=1.5) &=1-\sum_{k=0}^{4}e^{-1.5}\frac{1.5^k}{k!}\\ &=1-e^{-1.5}\left(1+1.5+\frac{1.5^2}{2!} +\frac{1.5^3}{3!}+\frac{1.5^4}{4!}\right)\\ &\approx 0.01858. \end{aligned}

Thus, if the rate and model were correct, about 1.858% of background-only counts would be 5 or greater. This is evidence that the count is unusual relative to that specified background distribution; it is not a 98.142% probability that the cell expresses the target. If many cells and features are assessed, the number of large counts expected by chance also depends on how many comparisons were made. A control rate estimated with uncertainty, or a background distribution with more variability than Poisson, could change the tail probability materially.

Cell-Level Controls and Background Estimation

Background can vary from cell to cell, so a single pooled control average may conceal important structure. One practical diagnostic is to compare target counts with one or more controls within each cell. For a simple illustration, if several controls are considered exchangeable on the count scale, their average in a cell can serve as a rough estimate of that cell's control level. This requires a strong assumption: the selected controls must respond similarly to the background affecting the target. If they do not, the average is only a descriptive reference.

Worked Example: A Cell-Specific Control Summary

Consider four hypothetical cells, each measured for one target tag and two isotype controls. The paired control counts are \((1,2)\), \((0,2)\), \((3,4)\), and \((1,1)\); the respective target counts are \(4, 1, 5,\) and \(7\). Using the mean of the two controls in each cell gives control summaries \(1.5, 1.0, 3.5,\) and \(1.0\) counts per control feature.

Subtracting each summary from its target count gives descriptive excesses of \(4-1.5=2.5\), \(1-1.0=0\), \(5-3.5=1.5\), and \(7-1.0=6.0\) counts. These are not observed integer signal counts; they are differences between an observed count and a control summary. In particular, an excess of zero does not establish that the target has no signal, and an excess of 6.0 does not establish six true target-derived molecules. The calculation is useful for locating cells whose target counts are high relative to their controls, provided the controls are comparable and the uncertainty in their small counts is kept in view.

Derivation and Practical Implications

The additive model also clarifies why direct subtraction is not a literal decomposition of each cell's observed count. Under the model, an observed total \(Y=y\) can arise from any split \(S=s\) and \(B=y-s\), for \(s=0,\ldots,y\). Even if the expected background rate were known, the count itself does not label which individual molecules came from signal. Subtracting an expected background from an observed count estimates an excess on average under the calibration assumptions; it does not recover the latent \(S\) exactly.

For example, if a control suggests an expected background of 1.5 counts and the target observation is 1, the arithmetic excess is \(1-1.5=-0.5\). A negative value cannot be a physical count of signal. It indicates that an excess estimate can fluctuate below zero when the count is noisy. Replacing it with zero may be convenient for a display, but changes the estimator and can bias summaries upward; the truncation should not be mistaken for evidence that the true signal is exactly zero.

A process for assessing background should keep estimation, comparison, and interpretation distinct:

Specify the measurement and controls.
Identify the tag counts, the relevant cell and batch structure, and which features are intended as isotype or negative controls.

Check comparability.
Ask whether controls experienced the same processing and whether their binding and count behavior plausibly track the target feature's background.

Describe the control distribution.
Inspect its counts across cells and batches. Do not reduce a variable control pattern to one rate without checking the consequences.

Compare target counts to that reference.
Use contrasts or a stated background model, and report the assumptions and uncertainty involved.

Interpret as evidence, not a molecule-by-molecule label.
State whether counts are unusual relative to the controls, without claiming that every low count is true expression or every excess count is certainly specific.

Diagnostics, Alternatives, and Limitations

Several patterns warn that a simple common Poisson background is not adequate. Control counts may differ substantially among cells, vary by batch, or show many more high counts than a Poisson distribution with the same mean would predict. Target-to-control differences may also change with cell type, total tag capture, or staining conditions. These patterns suggest that the background rate is not constant or that the Poisson variance assumption is too restrictive. Plots of control counts by cell and batch, alongside target-versus-control comparisons, help reveal such failures. A control feature that is nearly always zero may also be uninformative about background in the range relevant to a target.

The main alternatives answer different questions. Simple subtraction gives an easily summarized excess but can produce negative values and does not model count uncertainty. Thresholding labels counts above a chosen cutoff; it is easy to apply but ignores cell-specific background and can misclassify observations when control distributions overlap target counts. A Poisson background tail calculation makes its reference model explicit, but is sensitive to rate estimation and extra-Poisson variation. More flexible count models can accommodate varying rates or overdispersion, but require enough appropriate controls and additional assumptions. None of these approaches repairs a poorly matched control.

The model is a starting point, not a universal account of CITE-seq background. Background can be feature-specific, cell-specific, and batch-dependent; signal and background need not be independent in all experimental settings. A negative control that is not truly negative, or an isotype that does not mimic the target antibody's nonspecific behavior, can mislead even a carefully calculated analysis. The defensible conclusion is therefore calibrated: controls can show whether counts are compatible with a measured background pattern and can flag unusual counts, but they do not make low counts automatically biological or fully identify the latent components.

Common Mistakes

Calling every nonzero count expression. A positive count can arise from background. Compare it with controls and describe the evidence rather than equating detection with specific binding.
Treating the isotype rate as identical to every target's background. Similarity is an assumption to assess, not a guaranteed property of the control.
Calling a background tail probability the probability of background. It is a probability of data at least this extreme under a specified background model, not a posterior probability that a count is nonspecific.
Interpreting subtraction as exact signal recovery. Observed counts do not label their source. A target-minus-control difference is an estimate or contrast, not a direct count of signal molecules.
Ignoring extra variation. If controls are more variable than the Poisson model predicts, Poisson tail probabilities can be too small and apparent evidence of signal overstated.

Key takeaway: CITE-seq protein-tag counts can be modeled as an additive combination of target-related signal and nonspecific background. Isotype controls and negative-control features help characterize background, but their conclusions depend on comparability, model fit, and uncertainty; a low count is not automatically true expression.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

Under the independent Poisson signal-plus-background model, what are the mean and variance of the observed count?
A negative-control feature has counts \(0,1,1,2\) in four hypothetical cells. What is its mean count, and what does that number estimate under the stated control interpretation?
Why is a target count of 2 not automatically evidence of target expression, even if it is nonzero?
A researcher subtracts an estimated background of 2.4 from an observed target count of 1 and reports signal \(-1.4\). What is the problem with this interpretation?
What does a background-only tail probability of 0.02 mean, and what does it not mean?

Solutions. (1) The observed count is Poisson with rate \(\lambda_S+\lambda_B\), so its mean and variance are both \(\lambda_S+\lambda_B\). (2) The mean is \((0+1+1+2)/4=1.0\) count per cell. If the feature is a valid negative control, this is a sample summary informing the background level for that control, not necessarily the exact background rate of every target. (3) Background can generate nonzero counts, so the count must be assessed relative to a suitable control distribution and its assumptions. (4) A physical signal count cannot be negative. The negative difference is a noisy excess estimate, not a recovered latent count; truncating it to zero would also change the estimator. (5) If the specified background model is correct, 2% of background-only observations would be at least as large as the observed value. It is not a 2% probability that the observation is background, nor a 98% probability of true expression.