Why RNA-seq Counts Are Overdispersed

Why the Poisson Model Is a Useful Baseline

An RNA-seq experiment records how many reads are assigned to each gene in each sample. The gene-by-sample count matrix makes it possible to compare a gene’s counts across biological replicates, but the counts naturally fluctuate even when the underlying expression level is unchanged. A basic probability model helps separate this expected read-sampling variation from additional variability across samples.

The Poisson distribution is a useful baseline because it describes counts generated at a fixed rate over a fixed exposure. In this setting, the exposure can be thought of as the opportunity to observe reads, such as a specified sequencing depth. The simple model is a starting point, not a claim that all RNA-seq counts follow it.

For a particular gene, suppose comparable samples have the same underlying rate and sequencing exposure, and that read assignments behave as independent events at that rate. Under this simplified model, counts from different samples are independent Poisson observations with a common mean \(\lambda\). These assumptions are deliberately restrictive: biological samples can differ in expression, and technical factors can change the effective opportunity to observe reads. The model is most useful here as a reference against which extra variation can be understood.

If there are \(n\) replicate counts \(y_1,\ldots,y_n\), their sample mean is \(\bar{x}\) and their sample variance is \(s^2\). Under the common-rate Poisson model, both estimate the same underlying quantity. Thus \(s^2\) substantially greater than \(\bar{x}\) is a descriptive sign that the observed counts vary more than the baseline predicts. It is not, by itself, a formal test or proof: with few replicates, both estimates can be noisy.

Comparing the Mean and Variance

Consider a hypothetical gene with counts from four biological replicates: 4, 8, 12, and 16 reads, all measured under comparable sequencing exposure. The average count summarizes the scale of expression, while the sample variance summarizes how spread out the replicate counts are.

The units differ formally: a count mean is in reads, while a variance is in squared reads. Their numerical comparison is nevertheless meaningful for this model because the Poisson distribution equates the numerical value of the mean rate and variance. In practice, the ratio \(s^2/\bar{x}\) can summarize the comparison when \(\bar{x}>0\). A ratio above one suggests excess sample variability; it should be interpreted alongside the number of replicates, the design, and the range of counts.

A single gene is not the only possible diagnostic. One can examine mean–variance patterns across genes, or inspect replicate counts for genes of similar mean. A broad tendency for observed variances to exceed their corresponding means is more informative than an isolated large ratio. Even then, differences in exposure or sample composition can create apparent excess variability if they have not been accounted for.

How Biological Heterogeneity Adds Variance

The fixed-rate assumption is often unrealistic for biological replicates. Individuals, tissues, or cell populations can differ in their underlying expression for the same gene. A useful way to express this is to let the Poisson rate itself vary among samples. Write \(\Lambda\) for the sample-specific rate and suppose that, conditional on \(\Lambda\), the observed count \(Y\) is Poisson with mean \(\Lambda\).

Conditional on a particular rate, read-sampling variation has variance \(\Lambda\). Across samples, however, the rates may also vary. The law of total variance separates these two contributions: the average conditional variance and the variance of the conditional means.

The first term is the expected Poisson sampling variance; the second is the extra variance caused by differences in rates across samples. Also, \(E(Y)=E(\Lambda)\). Consequently, whenever \(\operatorname{Var}(\Lambda)>0\), the marginal count variance exceeds the marginal count mean. This result explains how biological heterogeneity can produce overdispersed counts without changing the Poisson assumption for read sampling within each sample.

What Can Create Excess Variation?

Biological heterogeneity can arise because samples differ in the proportions of cell types, the physiological state of the tissue, or other biological characteristics that affect expression. Differences among organisms or patients may also matter. These are examples of variation in the underlying rate, not merely random read assignment. If the study’s biological replicates represent a diverse population, their counts may vary more than counts from repeated sequencing of the same material.

Not every apparent excess is biological. Unequal sequencing exposure can change expected counts between samples, and unrecorded batch or processing differences can produce additional variation. A gene’s counts can also vary because of sampling error alone. The comparison with a Poisson baseline is informative only when the expected counts being compared are reasonably comparable, or when relevant differences in exposure and sample structure have been accounted for.

Diagnostics, Alternatives, and Limitations

A useful first diagnostic is to plot or tabulate replicate counts for selected genes, while noting their mean levels and whether samples are comparable. Across many genes, one can compare observed variability with the Poisson baseline at similar mean counts. Look for systematic patterns rather than relying on one gene’s ratio. Large counts, unusual zero patterns, and samples with broadly elevated or depressed counts can all be clues that the simple assumptions need review; none alone establishes a biological explanation.

Technical replicates and biological replicates answer different questions. Repeated sequencing of the same material can help characterize read-sampling and technical variation, but it does not represent variation among biological samples. Conversely, biological replicates capture sample-to-sample differences relevant to inference about a biological condition. Treating technical repetitions as if they were independent biological units can make the apparent evidence more precise than the design warrants.

The fixed-rate Poisson model is attractive because its mean–variance relationship is simple. If observed variability is close to that baseline and the assumptions are plausible, it may provide a useful description. When variation is consistently larger, one can first inspect exposure, sample labels, batch structure, and biological composition. Stratifying samples by known design features or studying technical and biological replicates separately can help diagnose the source. A count model that allows extra variability is another general alternative; selecting and fitting such a model is a separate task.

A key limitation is that \(s^2/\bar{x}\) is noisy when the number of replicates is small, especially for low counts. A ratio above one does not mean that every sample is biologically different, and a ratio below one does not prove the Poisson assumptions. Across genes, the counts are not independent summaries of a single shared process, so a pattern should be interpreted in context rather than treated as a universal test. Overdispersion describes a variance relationship, not its cause.

• Mistake: “The sample variance is greater than the sample mean, so biological heterogeneity has been proved.”
  Correction: The comparison is descriptive. Chance, unequal exposure, technical effects, or other departures from the model can also produce the pattern.
• Mistake: “A Poisson count has no variability beyond its mean.”
  Correction: A Poisson count has sampling variance equal to its mean. Excess variance appears when the fixed-rate assumption fails or additional sources of variation are present.
• Mistake: “Many technical replicates can replace biological replicates.”
  Correction: Technical replication helps characterize measurement variation; it does not substitute for sampling biological units when the scientific target is biological variability.