RNA Velocity: Why a Confident Arrow of Cellular Direction Can Point the Wrong Way

Zetobit Bioinformatics Insight Series

Single-Cell Trajectory Inference

RNA Velocity: Why a Confident Arrow of Cellular Direction Can Point the Wrong Way

RNA velocity promises something a static snapshot cannot give: the direction each cell is heading. The arrows it produces are visually compelling and often biologically right — but they rest on kinetic assumptions that real data frequently violates, and the smooth stream you see can be an artifact of the projection rather than a readout of the biology.

By Kanna Nandakumar, PhD · Zetobit

Single-cell RNA sequencing captures a static snapshot: each cell is measured once, at the moment it was lysed, and then it is gone. From that frozen frame, trajectory-inference methods try to reconstruct the dynamic process the cells were caught in the middle of. Most such methods can order cells along a trajectory but cannot, on their own, tell which end is the beginning — they need external prior knowledge to orient the arrow of time.¹

RNA velocity was a genuinely clever answer to that limitation. Its insight: every cell’s transcriptome contains a short-term forecast of its own future. Newly transcribed unspliced pre-mRNA still carries introns; mature spliced mRNA does not. By measuring the ratio of unspliced to spliced molecules for each gene, you can infer whether that gene is being turned up or down — and aggregating across genes yields a per-cell velocity vector pointing toward the cell’s near-future state. It converts a snapshot into a direction without needing to be told which way time runs. When it works, it is a real advance.

The trouble is that “when it works” hides a set of assumptions strong enough to fail quietly, and an output attractive enough that the failure is easy to miss.

The model assumes kinetics that biology does not honor

The original steady-state formulation makes a strong simplifying assumption: for any given gene, all cells adhere to the same kinetic rates and therefore the same steady-state ratio of unspliced to spliced RNA.² One transcription rate, one splicing rate, one degradation rate per gene, shared across every cell in the dataset. The later dynamical model in scVelo relaxed the steady-state requirement by fitting the full induction-and-repression dynamics, but it kept a related assumption: a transcription rate that is constant within each kinetic state.

Real differentiation violates this. Genes can pass through multiple rate kinetics — different transcription or splicing rates in different lineages or stages — and these are precisely the genes that break the model. When such genes are present, the linear or time-invariant-rate assumptions are violated and can produce distorted or even reversed velocity estimates.³ Reversed is the dangerous word: the method does not fail loudly and return noise; it can return a confident arrow pointing the wrong way down the trajectory. In the mouse and human erythroid systems, genes with multiple rate kinetics led standard analysis to distorted lineage inference until the violating genes were handled specifically.³

The standard mitigation — manually identifying and removing multiple-rate-kinetics genes before running velocity — is itself a warning sign. It means the analyst must know which genes misbehave in their system, which is close to needing the answer in advance, and it makes the result contingent on a subjective curation step that is rarely reported in detail.

The signal itself is faint

Even where the kinetic assumptions approximately hold, the measurement is working from a weak signal. Unspliced reads — the entire basis for the forecast — are a small and noisy fraction of a droplet-based library. Velocity estimates are frequently inaccurate or inconsistent when recovering cellular transitions, partly because of the severely low signal-to-noise ratio in unspliced mRNAs.¹ The unspliced counts also depend on technical choices — the intron-annotation strategy, whether reads are counted from intronic regions or spanning junctions — so the same cells can yield different velocities under different preprocessing. A forecast built on a sparse, noisy input inherits that fragility, and nothing about the smoothness of the final plot reveals it.

This is also why the magnitude of a velocity vector deserves more suspicion than its direction. The direction of the velocity vector is generally the primary output, while its magnitude — the speed — is often less reliable and should be interpreted with caution.² A plot that appears to show some cells moving “fast” and others “slow” is reporting the least trustworthy component of the estimate.

The projection can manufacture a direction that isn’t there

The most seductive failure mode is visual. Velocity is computed in high-dimensional gene space, but it is almost always displayed as arrows on a two-dimensional UMAP or t-SNE embedding — the streamlines everyone recognizes. That projection step is not a neutral rendering. The projected velocity stream is highly dependent on the number of genes included and on the chosen plotting parameters, and projection quality degrades at the boundary of the low-dimensional embedding — so the same underlying estimates can be made to tell different stories by changing display settings.⁴

The strongest demonstration of the danger is that the projection can produce a coherent-looking flow from data that contains no directional signal at all. In a published critique, arbitrary directions were projected onto a UMAP using velocity estimates from only three genes that showed no transient states — where the correct result would have been a noisy vector field pointing nowhere — and simulated mature cell types produced the same false projections, coherent streams absent from the ground-truth vector field.⁵ The smoothing that makes velocity streams beautiful is the same smoothing that can impose order on noise. A clean arrow field is not evidence that a directional signal exists.

Smoothing can invent a current. The same cells whose per-gene velocity vectors point in scattered directions (left) can render as a single confident stream once projected onto a 2D embedding and smoothed (right). The coherence is a property of the visualization, not proof of an underlying directional signal — which is why a clean arrow field must be checked against the gene-level dynamics, not trusted on sight.

How to keep velocity honest

None of this means RNA velocity is unusable — it means the streamline plot is a hypothesis to be tested, not a result to be reported. Several checks separate a trustworthy velocity analysis from a decorative one:

• Inspect the phase portraits, not just the stream. For key genes, the unspliced-versus-spliced phase plot should show the expected almond/loop shape; if genes show multiple pronounced kinetics, velocity should be applied with caution and the data possibly subsetted to individual lineages.⁴ The gene-level evidence is the real signal; the embedding is downstream of it.
• Trust direction over speed. Treat the orientation of the flow as the interpretable output and the magnitude as unreliable.²
• Stress-test the projection. Vary the gene set and plotting parameters; a conclusion that survives only one specific configuration is a plotting artifact, not a finding.⁴
• Prefer methods that reason in high dimensions. Approaches such as CellRank use the velocity field to infer future states while operating on the higher-dimensional representation, avoiding the misleading streams that arise purely from the embedding.⁴
• Corroborate with orthogonal evidence. Known markers, lineage tracing, or metabolic-labeling experiments that directly enrich nascent RNA provide the ground truth a snapshot cannot; a velocity direction that contradicts established biology is the velocity that is wrong far more often than the biology.¹

RNA velocity belongs to the same family as the other confident-output methods in this series: a pipeline that always produces a clean, plausible answer, whether or not the assumptions behind it hold. The arrows are compelling precisely because they are smooth and directional — and smoothness is exactly the property a projection can supply for free. The discipline it demands is to treat the beautiful streamline plot as the beginning of an investigation into whether the gene-level dynamics actually support it, not as the end of one.

References

1. Gao M, Qiao C, Huang Y. UniTVelo: temporally unified RNA velocity reinforces single-cell trajectory inference. Nature Communications. 2022;13:6586. nature.com/articles/s41467-022-34188-7
2. RNA velocity and beyond: current advances in modeling single-cell transcriptional dynamics. ScienceDirect. 2025. sciencedirect.com/science/article/pii/S1323893025000899
3. Gao M, Qiao C, Huang Y. UniTVelo: on multiple-rate-kinetics (MURK) genes causing distorted or reversed velocity, and erythroid re-analysis. Nature Communications. 2022;13:6586. ncbi.nlm.nih.gov/pmc/articles/PMC9633790
4. RNA velocity. Single-cell best practices (sc-best-practices.org). sc-best-practices.org/trajectories/rna_velocity.html
5. Bergen V, Soldatov RA, Kharchenko PV, Theis FJ. RNA velocity — current challenges and future perspectives. Molecular Systems Biology. 2021;17(8):e10282. pubmed.ncbi.nlm.nih.gov/34435732

© Zetobit LLC · Bioinformatics Insight Series zetobit.com