Why Isoform Structure Matters for Unspliced Counts
RNA velocity uses the relationship between unspliced and spliced RNA to estimate a gene’s local direction of change. At gene level, counts from transcripts assigned to the same gene are often pooled. That pooling is convenient, but it can combine transcripts with different exon structures and different processing behavior. Consequently, a gene’s unspliced count is not always a single, uniform precursor pool.
Alternative splicing can change which introns are removed, when they are removed, and which mature transcripts are produced. Intron retention can leave an intronic region in a transcript that would otherwise be spliced out. Transcript annotations determine which regions are recognized as exons or introns, and thus can affect whether reads are assigned to spliced or unspliced counts. These effects matter because a gene-level kinetic model commonly treats the pooled counts as if they followed one effective processing process.
A standard gene-level kinetic model writes \(U\) for unspliced abundance and \(S\) for spliced abundance. In a simple model, transcription produces unspliced RNA at rate \(\alpha\), unspliced RNA is converted to spliced RNA at rate \(\beta U\), and spliced RNA is degraded at rate \(\gamma S\). The resulting equations are:
This model assumes that the pooled unspliced quantity behaves as one kinetic pool with a common processing rate \(\beta\), and that the spliced pool has a common degradation rate \(\gamma\). The measured counts must also be sufficiently comparable across cells for their fitted relationship to represent changes in abundance rather than changes in measurement. The model is a useful approximation, not a guarantee that every transcript assigned to a gene has the same fate or processing rate.
Aggregation Can Hide Different Processing Rates
Suppose a gene has \(K\) transcript classes, indexed by \(k\). For a simple illustration, let \(U_k\) and \(S_k\) be their unspliced and spliced abundances, and let \(\beta_k\) be the rate at which unspliced RNA in class \(k\) is processed into spliced RNA. Assume each class follows the simple conversion relationship \(dS_k/dt=\beta_k U_k-\gamma_k S_k\). If gene-level abundances are sums across these classes, then:
For the conversion term to equal \(\beta U\) with one fixed rate \(\beta\), the weighted sum \(\sum_k \beta_k U_k\) must equal \(\beta\sum_k U_k\). When all \(\beta_k\) are the same, this holds. When rates differ, the corresponding effective rate is:
This is a weighted average of the class-specific rates, with weights determined by the current composition of the unspliced pool. It can therefore vary across cells or over time as isoform composition changes, even if each \(\beta_k\) remains constant. A model that imposes one fixed \(\beta\) can then misrepresent the pooled data. The same issue applies to degradation: differing \(\gamma_k\) values mean that \(\sum_k\gamma_kS_k\) need not equal one fixed \(\gamma S\).
Worked Example: A Changing Effective Splicing Rate
Consider two transcript classes with processing rates \(\beta_1=0.8\) and \(\beta_2=0.2\) per unit time. In one cell, let \(U_1=10\) and \(U_2=30\). Ignoring degradation for the moment, calculate the pooled conversion rate and compare it with a gene-level model that imposes \(\beta=0.5\).
The class-specific conversion contributions are \(0.8(10)=8\) and \(0.2(30)=6\), so total conversion is \(8+6=14\) abundance units per unit time. Total unspliced abundance is \(U=10+30=40\). Thus:
A model imposing \(\beta=0.5\) instead predicts \(0.5(40)=20\) conversion units per unit time, an overestimate of \(20-14=6\). Now suppose the isoform composition changes to \(U_1=30\) and \(U_2=10\), while the two rates remain unchanged. Total conversion becomes \(0.8(30)+0.2(10)=26\), and \(U\) remains 40, giving \(\beta_{\mathrm{eff}}=26/40=0.65\). The effective rate changed from 0.35 to 0.65 solely because the composition of the unspliced pool changed.
Alternative Splicing, Retention, and Annotation
Alternative splicing can make the meaning of a gene’s unspliced count depend on which transcripts are present. A read spanning an intronic region may support an unspliced assignment for one annotated transcript, while another transcript from the same gene may include that region as an exon. Intron-retaining transcripts add a related complication: an intronic sequence can persist in RNA rather than being a transient precursor that will promptly become a conventionally spliced transcript. Whether such reads enter a particular unspliced count depends on the quantification rules.
Annotation changes can alter the classification of reads without changing the underlying molecules. An annotation that adds an exon, removes an intron, or includes an alternative transcript can change which reads are considered evidence for spliced or unspliced RNA. Thus, a difference in unspliced counts between two analyses is not automatically evidence for a biological difference. It may reflect a difference in the reference annotation or in the counting procedure.
Worked Example: An Annotation-Dependent Velocity Change
Suppose the same hypothetical read data are processed using two annotations. Under annotation A, a gene has \(U=18\) and \(S=42\); under annotation B, 12 reads are assigned differently, giving \(U=30\) and \(S=42\). For illustration, suppose both analyses use the same fitted rates, \(\beta=0.5\) and \(\gamma=0.1\), per unit time. Calculate the velocity under each annotation.
Using \(v=\beta U-\gamma S\), annotation A gives:
Annotation B gives:
The estimated velocity differs by \(10.8-4.8=6.0\) abundance units per unit time. The calculation does not establish that one annotation is correct or that the cells changed biologically; it shows that the estimated velocity can be sensitive to how the same read data are classified. In an actual analysis, refitting the rates could also change the comparison.
It is important to distinguish an isoform effect from a simple counting error. A count may be technically accurate under a specified annotation while still being a poor summary for a single-rate kinetic model. The model’s biological interpretation depends on whether the counted RNA classes behave sufficiently like the assumed pooled precursor and product.
Worked Consequences for Gene-Level Velocity
Worked Example: Retained-Intron RNA in a Pooled Precursor Count
Consider a hypothetical gene with two unspliced classes. Class 1 has \(U_1=20\) and \(\beta_1=0.6\); class 2, representing an intron-retaining class with slower conversion in this illustration, has \(U_2=40\) and \(\beta_2=0.05\). Let total spliced abundance be \(S=30\), with common degradation rate \(\gamma=0.1\). Calculate the conversion-based velocity and compare it with the prediction obtained by applying \(\beta=0.6\) to all pooled unspliced RNA.
The class-specific conversion contributions sum to \(0.6(20)+0.05(40)=12+2=14\). Degradation is \(0.1(30)=3\), so the pooled class-specific velocity is:
The pooled unspliced count is \(U=20+40=60\). If a gene-level interpretation applies the first class’s rate to all of it, the predicted velocity is:
That prediction is 22 abundance units per unit time higher than the class-specific calculation. Its error arises from treating the slower-converting retained-intron class as if it had the faster rate. This is a deliberately simple illustration: real data may not identify these classes or their rates separately, and a fitted gene-level rate need not equal either class’s rate.
Diagnostics and Alternative Approaches
Isoform-related problems are plausible when gene-level kinetic behavior changes substantially under reasonable annotation or counting choices, or when genes with known complex transcript structures show systematic departures from the assumed relationship between unspliced and spliced counts. Diagnostics should be interpreted as evidence of sensitivity or model inadequacy, not as proof of a particular mechanism.
Review whether the gene has alternative exons, alternative introns, or annotated intron-retaining transcripts that could contribute differently to the counted pools.
Confirm which read features and annotation rules define unspliced and spliced counts. Do not assume that different pipelines use identical assignments.
Where feasible, compare estimates under defensible annotation or counting choices. Large changes indicate that the velocity interpretation depends on those choices.
Structured disagreement between the observed counts and a single-rate kinetic fit is consistent with model misspecification, but can also arise from measurement noise or other unmodeled biology.
The main alternative is an isoform-aware analysis, which attempts to distinguish transcript classes and estimate their behavior separately. It can represent different processing rates more directly, but requires enough informative reads to distinguish isoforms and depends on reliable transcript annotation. A gene-level model is simpler and can be more stable when data are sparse, but it conceals within-gene heterogeneity. A sensitivity analysis across annotations is useful for evaluating robustness, but does not by itself determine the correct transcript structure or recover unobserved rates.
These approaches have limits. Single-cell counts can be sparse, reads may not uniquely identify an isoform, and annotation may omit relevant transcripts. Even with an accurate annotation, measured counts may not cleanly separate transient precursor RNA from retained-intron RNA. When isoform-specific quantities cannot be estimated reliably, the appropriate conclusion is that gene-level velocity is conditional on the chosen counting and kinetic assumptions—not that the pooled estimate is necessarily an unbiased biological rate.
Common Pitfall
- Wrong reasoning: “Unspliced counts increased, so transcription must have increased.”
- Correction: An increase in the counted unspliced pool could reflect more transcription, slower processing, a change in isoform composition, intron retention, or a change in read assignment under the annotation. The count alone does not distinguish these explanations.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Two transcript classes have rates \(\beta_1=0.4\) and \(\beta_2=0.1\), with \(U_1=10\) and \(U_2=30\). What is the effective pooled rate?
- Why can a change in annotation alter a velocity estimate even when the underlying read data are unchanged?
- With \(\beta_1=0.5\), \(U_1=20\), \(\beta_2=0.1\), and \(U_2=20\), what is total conversion? What effective rate would a single-pool model need at this composition?
- Name one limitation of an isoform-aware approach and one limitation of a gene-level approach.
Solutions. (1) Total conversion is \(0.4(10)+0.1(30)=7\); total unspliced abundance is 40, so \(\beta_{\mathrm{eff}}=7/40=0.175\) per unit time. (2) Annotation rules can change which reads are assigned to spliced and unspliced counts, changing \(U\), \(S\), or both in the velocity formula. (3) Total conversion is \(0.5(20)+0.1(20)=12\), and the pooled unspliced abundance is 40, so the effective rate is \(12/40=0.3\) per unit time. (4) Isoform-aware estimation may be unstable when reads cannot distinguish transcripts; gene-level estimation can hide different isoform-specific processing rates.