Single-Cell RNA Velocity: Inferring Cellular Dynamics
RNA velocity is a computational approach that estimates the future transcriptional state of individual cells from single-cell RNA sequencing data by measuring the ratio of unspliced precursor messenger RNA to spliced mature messenger RNA. This method allows researchers to infer cellular dynamics, predict cell state transitions, and reconstruct developmental trajectories from static snapshot data. For students, researchers, analysts, and life-science professionals working with single-cell transcriptomics, understanding RNA velocity is essential for moving beyond static cell typing toward dynamic interpretations of differentiation, response to perturbation, and disease progression.
The core insight behind RNA velocity is that standard single-cell RNA sequencing captures only a static snapshot of gene expression at a point in time. RNA abundance is a powerful indicator of the state of individual cells, and single-cell RNA sequencing can reveal RNA abundance with high quantitative accuracy, sensitivity, and throughput. However, this approach captures only a static snapshot at a point in time, posing a challenge for the analysis of time-resolved phenomena such as embryogenesis or tissue regeneration. RNA velocity addresses this limitation by directly estimating the time derivative of the gene expression state, distinguishing between unspliced and spliced mRNAs in common single-cell RNA sequencing protocols. The result is a high-dimensional vector that predicts the future state of individual cells on a timescale of hours.
This article provides a practical framework for running RNA velocity analysis, interpreting results, and avoiding common pitfalls. It covers the biological basis of the method, the primary computational tools available, workflow decisions, quality controls, and interpretation limits.
At a Glance
RNA velocity analysis requires specific data inputs, computational choices, and interpretation strategies. The table below summarizes the key decisions researchers face when planning a velocity analysis.
| Workflow Component | Primary Options | Key Considerations |
|---|---|---|
| Data input | Standard scRNA-seq with spliced and unspliced counts from velocyto or STARsolo | Requires alignment output that distinguishes exonic and intronic reads, not all scRNA-seq protocols preserve this information |
| Core tool | velocyto, scVelo, or newer methods such as TIVelo, spVelo, NeuroVelo, NARVI | Tool choice affects model assumptions, scalability, and ability to handle transient states or spatial data |
| Model assumption | Steady-state ratio, dynamical modeling, or trajectory-based orientation | Steady-state assumes full splicing dynamics are observed, dynamical modeling relaxes this assumption |
| Dimensionality reduction | UMAP or t-SNE embedding for visualization | Velocity vectors are projected onto low-dimensional representations, choice affects visual interpretation |
| Validation | Comparison with known biology, pseudotime methods, or lineage tracing | Independent validation is essential because velocity predictions can be wrong |
The Biological Basis of RNA Velocity
RNA velocity relies on the biology of gene expression and messenger RNA maturation. When a gene is transcribed, the initial product is unspliced precursor messenger RNA containing introns. This precursor is subsequently processed through splicing to produce mature spliced messenger RNA that can be translated into protein. The ratio of unspliced to spliced mRNA for a given gene provides information about whether transcription is being activated or repressed at the time the cell was captured.
RNA velocity describes the rate of gene expression change for an individual gene at a given time point based on the ratio of its spliced and unspliced messenger RNA. When a gene is being upregulated, the cell will contain a higher proportion of newly transcribed unspliced mRNA relative to the steady-state level. When a gene is being downregulated, the proportion of unspliced mRNA will be lower than expected. By measuring these ratios across thousands of genes simultaneously, the method constructs a high-dimensional vector that predicts the future state of each cell.
The foundational work on RNA velocity validated the approach in the neural crest lineage, demonstrated its use on multiple published datasets and technical platforms, revealed the branching lineage tree of the developing mouse hippocampus, and examined the kinetics of transcription in human embryonic brain. The method was expected to greatly aid the analysis of developmental lineages and cellular dynamics, particularly in humans.
Core Principles of Velocity Estimation
The Splicing Kinetics Model
The mathematical foundation of RNA velocity is a model of transcriptional dynamics that relates the abundance of unspliced and spliced mRNA for each gene. The model assumes that transcription produces unspliced mRNA at a certain rate, splicing converts unspliced to spliced mRNA at a rate governed by a splicing constant, and degradation removes spliced mRNA at a degradation rate. The ratio of unspliced to spliced mRNA at steady state provides a baseline against which observed ratios can be compared.
The original implementation of RNA velocity relied on the assumption of a common splicing rate and the observation of the full splicing dynamics with steady-state mRNA levels. When these assumptions are violated, errors in velocity estimates arise. The scVelo method overcomes these limitations by solving the full transcriptional dynamics of splicing kinetics using a likelihood-based dynamical model. This generalization extends RNA velocity to systems with transient cell states, which are common in development and in response to perturbations.
Steady-State Versus Dynamical Modeling
The choice between steady-state and dynamical modeling is one of the most consequential decisions in RNA velocity analysis. The steady-state model, implemented in the original velocyto approach, assumes that the system has reached equilibrium and that the observed unspliced to spliced ratios can be compared against a fitted steady-state line. This assumption is often violated in real datasets because cells may be captured during transient phases of differentiation or response.
The dynamical model implemented in scVelo estimates gene-specific rates of transcription, splicing, and degradation, and recovers each cell's position in the underlying differentiation processes. This approach can detect putative driver genes and disentangle subpopulation kinetics in systems such as neurogenesis and pancreatic endocrinogenesis. The dynamical model is generally preferred for datasets with complex trajectories or transient cell states.
Trajectory-Based Orientation
Newer methods have moved away from explicit ordinary differential equation assumptions. TIVelo first determines the velocity direction at the cell cluster level based on trajectory inference before estimating velocity for individual cells. This approach calculates an orientation score to infer the direction at the cluster level without an explicit ODE assumption, which effectively captures complex transcriptional patterns and avoids potential inconsistencies in velocity estimation for genes that do not follow the simple ODE assumption. TIVelo was validated on 16 real datasets and compared with six benchmarking methods.
Computational Tools for RNA Velocity
Velocyto
Velocyto was the pioneering implementation of RNA velocity. It introduced the concept of distinguishing unspliced and spliced mRNA from standard single-cell RNA sequencing data and provided a computational pipeline for estimating velocity vectors. The tool requires alignment files that retain intronic reads, typically generated with STARsolo or similar aligners that can output spliced and unspliced count matrices.
scVelo
scVelo represents the second generation of RNA velocity tools. It generalizes the framework by solving the full transcriptional dynamics of splicing kinetics using a likelihood-based dynamical model. This approach overcomes the limitations of the steady-state assumption and extends RNA velocity to systems with transient cell states. scVelo infers gene-specific rates of transcription, splicing, and degradation, recovers each cell's position in the underlying differentiation processes, and detects putative driver genes.
Advanced Methods
Several newer methods address specific limitations of earlier tools. NARVI is a deep learning framework that learns the relationship between the expression patterns and velocities of computable genes to accurately estimate velocities for otherwise incalculable genes. This approach recovers velocities for thousands of genes that were previously impossible to estimate, broadening the scope of downstream analyses.
NeuroVelo couples learning of an optimal linear projection with non-linear neural ordinary differential equations. Using dynamical systems theory in the optimized latent space, NeuroVelo can simultaneously determine cellular transitions and identify gene interactions that drive the observed temporal dynamics of gene expression. It reconstructs correct cell-type transitions and identifies gene-regulatory networks that drive cell fate directly from the data.
spVelo is designed for multi-batch spatial transcriptomics data. It utilizes the spatial information inherent in spatial transcriptomics and supports uncertainty quantification, complex trajectory pattern discovery, driver marker identification, gene regulatory network inference, and temporal cell-cell communication inference.
Practical Workflow for RNA Velocity Analysis
Step 1: Data Preparation and Quality Control
The first step in any RNA velocity analysis is ensuring that the input data contains the necessary information. Standard single-cell RNA sequencing protocols can provide unspliced and spliced count matrices if the alignment pipeline is configured to distinguish intronic and exonic reads. The velocyto pipeline or STARsolo can generate these matrices from raw sequencing data.
Quality control should follow established single-cell analysis practices. Filter cells with low total counts, high mitochondrial content, or other indicators of poor quality. Filter genes that are expressed in very few cells. Normalize the data to account for differences in sequencing depth across cells.
Step 2: Preprocessing and Dimensionality Reduction
Before running velocity estimation, the data should be preprocessed using standard single-cell analysis workflows. This includes normalization, identification of highly variable genes, and dimensionality reduction using principal component analysis. The resulting low-dimensional representation is used for clustering and for visualizing velocity vectors.
Uniform manifold approximation and projection (UMAP) is commonly used to explore the cellular composition of tissues and to visualize velocity results. The choice of low-dimensional representation affects the visual interpretation of velocity fields, but the Hodge decomposition of RNA velocity fields has been shown to reveal key cell dynamic features such as cell cycle, bifurcation, and cell lineage differentiation regardless of the choice of low-dimensional representation.
Step 3: Velocity Estimation
Run the velocity estimation using your chosen tool. For scVelo, this involves computing velocity on the normalized data, then projecting the velocity vectors onto the low-dimensional embedding. The dynamical model should be used when the data contains transient cell states or when the steady-state assumption is questionable.
For spatial transcriptomics data, spVelo provides a scalable framework for multi-batch datasets and supports downstream applications including uncertainty quantification and trajectory pattern discovery.
Step 4: Visualization and Interpretation
Velocity results are typically visualized as arrows or streamlines on a UMAP or t-SNE embedding. The direction and magnitude of the arrows indicate the predicted future state of each cell. Streamline plots show the overall flow of cells through the transcriptional landscape.
The Hodge decomposition approach provides a more comprehensive analysis of cell state transitions by decomposing the velocity field into curl-free, divergence-free, and harmonic components. This decomposition distinctly reveals key cell dynamic features such as cell cycle, bifurcation, and cell lineage differentiation, and the consistency across different representations demonstrates that the Hodge decomposition is a reliable and robust way to extract these cell dynamic features.
Step 5: Validation
Velocity predictions should be validated against independent evidence. This can include comparison with known developmental trajectories, pseudotime analyses, lineage tracing experiments, or longitudinal samples. In studies of tumor cell evolution, RNA velocity and reverse graph embedding were used to determine tumor cell relationships, and longitudinal samples were studied to determine tumor cellular state and its evolution.
Options and Tradeoffs in Tool Selection
The choice of RNA velocity tool involves tradeoffs between model flexibility, computational cost, and interpretability. The original steady-state model is computationally efficient but makes strong assumptions that are often violated. The dynamical model in scVelo is more flexible but requires more computational resources and can be sensitive to model misspecification.
Trajectory-based methods such as TIVelo avoid explicit ODE assumptions and can capture complex transcriptional patterns, but they depend on the quality of the initial trajectory inference. Deep learning methods such as NARVI and NeuroVelo can recover velocities for genes that are otherwise incalculable and identify gene-regulatory networks, but they require more computational resources and their results can be harder to interpret.
The performance of RNA velocity methods varies depending on the method and dataset. A comparative benchmark analyzed the performance of five RNA velocity methods on three datasets based on local consistency, method agreement, identification of driver genes, and robustness to sequencing depth. This benchmark provides a resource for scientists to understand the strengths and limitations of different RNA velocity methods.
Observations and Measurements in Applied Studies
RNA velocity has been applied across diverse biological contexts, providing insights into cell fate decisions and disease mechanisms. In a study of the osteoimmunology microenvironment in periodontitis, researchers performed single-cell RNA sequencing to identify 51,248 single cells from healthy controls, patients with severe chronic periodontitis, and patients after initial periodontal therapy. Pseudotime cell trajectory and RNA velocity analysis, combined with gene enrichment analysis, were used to reveal the molecular pathways underlying cell fate decisions. The analysis identified ten major cell types and revealed that CXCL12+ MSC-like pericytes could convert their identity into other cell types.
In skeletal muscle research, single-cell RNA sequencing of porcine longissimus dorsi and soleus muscles generated a transcriptomic dataset of myogenic cells and other cell types. RNA velocity analysis was performed alongside unsupervised clustering, marker gene analysis, gene set variation analysis, and pseudotime analysis to explore the heterogeneity of myogenic cells. The analysis identified six subpopulations of myo-lineage cells and revealed a higher proportion of cycling and MYF6+ cells in the soleus population.
In cancer research, RNA velocity has been used to understand tumor cell evolution and heterogeneity. A study of hepatocellular carcinoma and intrahepatic cholangiocarcinoma analyzed 46 biopsies from 37 patients and used RNA velocity and reverse graph embedding to determine tumor cell relationships. The analysis revealed that increased functional clonality was tightly linked to patient prognosis and was accompanied by a polarized immune cell landscape.
In glioblastoma research, single-cell RNA sequencing of 53,586 adult glioblastoma cells and 22,637 normal human fetal brain cells revealed a conserved neural tri-lineage cancer hierarchy centered around glial progenitor-like cells. RNA velocity analysis showed that this progenitor population contained the majority of the cancer's cycling cells and was often the originator of the other cell types.
In wound healing research, single-cell RNA sequencing of murine skin wounds revealed that wound fibroblasts group into twelve clusters. Pseudotime and RNA velocity analyses revealed that some clusters likely represent consecutive differentiation states toward a contractile phenotype, while others appear to represent distinct fibroblast lineages.
Records and Documentation for Reproducibility
Reproducibility in RNA velocity analysis requires careful documentation of every computational decision. The following records should be maintained for each analysis:
| Record Type | Specific Information | Purpose |
|---|---|---|
| Data provenance | Source of raw data, sequencing platform, alignment software and version | Ensures data can be traced to origin and reanalyzed if needed |
| Parameter settings | All parameters used for filtering, normalization, dimensionality reduction, and velocity estimation | Allows exact reproduction of the analysis |
| Tool versions | Versions of velocyto, scVelo, or other tools and their dependencies | Computational environments change, version control is essential |
| Model choices | Whether steady-state or dynamical modeling was used, and the rationale | Model assumptions affect results and interpretation |
| Validation results | Comparison of velocity predictions with independent evidence | Documents the reliability of the analysis |
The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable. Applying these principles to RNA velocity analyses ensures that results can be shared and reproduced by other researchers. The National Institutes of Health Genomic Data Sharing Policy provides additional guidance for sharing genomic data generated with federal funding.
Common Failure Patterns and How to Avoid Them
Failure Pattern 1: Ignoring the Steady-State Assumption
The most common failure in RNA velocity analysis is applying the steady-state model to data that contains transient cell states. When cells are captured during active differentiation or response to perturbation, the steady-state assumption is violated and velocity estimates contain errors. The dynamical model in scVelo should be used for such datasets.
Failure Pattern 2: Poor Quality Input Data
RNA velocity requires accurate quantification of unspliced and spliced mRNA. If the alignment pipeline does not properly distinguish intronic and exonic reads, the velocity estimates will be unreliable. Quality control should include verification that the unspliced to spliced ratios are biologically plausible.
Failure Pattern 3: Overinterpreting Velocity Vectors
Velocity vectors predict the future state of cells on a timescale of hours, but they are estimates based on a model. The predictions can be wrong, especially for genes that do not follow the simple splicing kinetics model. Velocity results should be validated against independent evidence before drawing biological conclusions.
Failure Pattern 4: Ignoring Technical Artifacts
Sequencing depth and other technical factors can affect velocity estimates. A comparative benchmark found that robustness to sequencing depth varies across methods. Researchers should assess whether their results are robust to technical variation.
Failure Pattern 5: Using Inappropriate Dimensionality Reduction
The choice of low-dimensional representation affects the visual interpretation of velocity fields. Different embeddings can produce different apparent trajectories. The Hodge decomposition approach provides a way to extract cell dynamic features that is consistent across different representations.
Limitations of RNA Velocity
RNA velocity has several inherent limitations that researchers must understand. The method estimates the time derivative of the gene expression state, but this estimate is based on a model of splicing kinetics that may not hold for all genes or all biological contexts. Errors in velocity estimates arise if the central assumptions of a common splicing rate and the observation of the full splicing dynamics with steady-state mRNA levels are violated.
Existing RNA velocity estimation tools often fail to calculate velocities for a substantial portion of genes due to technical limitations or model assumptions. This restricts downstream analyses that rely on velocity estimations. Deep learning approaches such as NARVI can recover velocities for thousands of genes that were previously impossible to estimate, but these methods have their own limitations.
The predictive power of RNA velocity methods is limited. Reconstructing temporal cellular dynamics from static single-cell transcriptomics remains a major challenge. Methods based on RNA velocity are useful, but interpreting their results to learn new biology remains difficult.
RNA velocity captures cell dynamic information in biological processes, yet a comprehensive analysis of cell state transitions and their associated chemical and biological processes remains a gap. The Hodge decomposition approach addresses this gap by examining decomposed components of the velocity field, but it is a visualization and analysis tool instead of a fundamental improvement in velocity estimation.
Quality Controls and Validation Strategies
Quality control in RNA velocity analysis should include multiple layers of validation. First, verify that the unspliced and spliced count matrices are biologically plausible. Genes that are known to be rapidly induced or repressed should show appropriate velocity signals.
Second, compare velocity results with pseudotime analyses. If velocity predicts a trajectory that contradicts pseudotime ordering, one of the two methods is likely wrong. In the wound healing study, pseudotime and RNA velocity analyses were used together to reveal that some clusters represent consecutive differentiation states while others represent distinct lineages.
Third, use lineage tracing or longitudinal data when available. In the liver cancer study, longitudinal samples from 4 patients were used to determine tumor cellular state and its evolution. This provided independent validation of the velocity-based predictions.
Fourth, assess the robustness of results to parameter choices and technical variation. The comparative benchmark of five RNA velocity methods provides a framework for assessing local consistency, method agreement, identification of driver genes, and robustness to sequencing depth.
Safety and Regulatory Context
RNA velocity analysis is a computational method that does not involve direct manipulation of biological samples or human subjects. However, the data used for such analyses may be subject to regulatory requirements. The National Institutes of Health Genomic Data Sharing Policy governs the sharing of genomic data generated with NIH funding. Researchers should ensure that their data sharing practices comply with applicable policies.
The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable. Applying these principles to RNA velocity analyses ensures that results can be shared and reproduced by other researchers.
For researchers working with human data, privacy and consent considerations apply. Single-cell RNA sequencing data can reveal sensitive information about individuals, and researchers must ensure that data are de-identified and used in accordance with ethical guidelines.
Professional Escalation Criteria
Researchers should seek additional expertise or escalate concerns in the following situations:
| Situation | Recommended Action |
|---|---|
| Velocity results contradict established biology | Consult with a computational biologist or bioinformatician to assess whether the model assumptions are appropriate |
| Results are highly sensitive to parameter choices | Seek guidance on parameter selection and consider using multiple methods for cross-validation |
| Unspliced and spliced count matrices appear unreliable | Revisit the alignment pipeline and verify that intronic reads are properly quantified |
| Velocity predictions cannot be validated with any independent evidence | Consider whether the biological system is appropriate for RNA velocity analysis |
| Computational resources are insufficient for the chosen method | Consider using a simpler method or seeking access to high-performance computing resources |
Frequently Asked Questions
What is the difference between RNA velocity and pseudotime analysis?
RNA velocity estimates the time derivative of the gene expression state from the ratio of unspliced to spliced mRNA, predicting the future state of individual cells on a timescale of hours. Pseudotime analysis orders cells along a trajectory based on transcriptional similarity, but it does not provide directional information. RNA velocity provides directionality that pseudotime cannot, but pseudotime can be used to validate velocity results.
What data do I need for RNA velocity analysis?
You need single-cell RNA sequencing data that has been aligned in a way that distinguishes unspliced and spliced mRNA. The velocyto pipeline or STARsolo can generate spliced and unspliced count matrices from raw sequencing data. Not all scRNA-seq protocols preserve the intronic information needed for velocity analysis.
Which tool should I use for RNA velocity analysis?
The choice of tool depends on your data and biological question. scVelo with dynamical modeling is a good default for most datasets because it relaxes the steady-state assumption. For spatial transcriptomics data, spVelo is designed for multi-batch datasets. For recovering velocities of genes that are otherwise incalculable, NARVI may be appropriate. For identifying gene-regulatory networks that drive cell fate, NeuroVelo is a newer option.
How do I interpret RNA velocity plots?
Velocity plots show arrows or streamlines on a low-dimensional embedding such as UMAP or t-SNE. The direction of the arrows indicates the predicted future state of each cell. Streamline plots show the overall flow of cells through the transcriptional landscape. The Hodge decomposition approach can reveal cell cycle, bifurcation, and cell lineage differentiation features in the velocity field.
What are the main limitations of RNA velocity?
RNA velocity relies on assumptions about splicing kinetics that may not hold for all genes or all biological contexts. Errors arise when the steady-state assumption is violated. Many tools fail to calculate velocities for a substantial portion of genes. The predictive power of the method is limited, and results should be validated with independent evidence.
Can RNA velocity be used for spatial transcriptomics data?
Yes, spVelo is a scalable framework for RNA velocity inference of multi-batch spatial transcriptomics data. It utilizes the spatial information inherent in spatial transcriptomics and supports uncertainty quantification, complex trajectory pattern discovery, driver marker identification, gene regulatory network inference, and temporal cell-cell communication inference.
How do I validate RNA velocity results?
Validation strategies include comparing velocity predictions with pseudotime analyses, using lineage tracing or longitudinal data when available, assessing robustness to parameter choices and technical variation, and comparing results across multiple methods. The comparative benchmark of five RNA velocity methods provides a framework for assessing local consistency, method agreement, identification of driver genes, and robustness to sequencing depth.
What are the emerging directions in RNA velocity analysis?
Emerging directions include deep learning-based methods such as NARVI and NeuroVelo, integration with spatial and multimodal data, and the use of Hodge decomposition for comprehensive analysis of cell state transitions. These methods aim to address the limitations of existing tools and provide deeper insights into gene transcriptional dynamics.
Related Bioinformatics Guides
- Single-Cell RNA-Seq Analysis Pipelines for Veterinary Immunology
- Single-Cell RNA Sequencing: From Bulk to Resolution
- Master Guide: Single-Cell RNA Sequencing Bioinformatics Workflows
- Single-cell RNA-seq Trajectory Inference and Cell Lineage Tracing
- Single-Cell RNA-seq Clustering and Cell-Type Annotation Pipelines
References and Further Reading
- EMBL-EBI Training. European Bioinformatics Institute.
- NCBI Data Resources. National Center for Biotechnology Information.
- Genomic Data Sharing Policy. National Institutes of Health.
- The FAIR Guiding Principles. Scientific Data.
- Single-cell RNA landscape of the osteoimmunology microenvironment in periodontitis.. Theranostics, 2022.
- Single-cell RNA-seq reveals novel interaction between muscle satellite cells and fibro-adipogenic progenitors mediated with FGF7 signalling.. Journal of cachexia, sarcopenia and muscle, 2024.
- RNA velocity of single cells.. Nature, 2018.
- Single-cell atlas of tumor cell evolution in response to therapy in hepatocellular carcinoma and intrahepatic cholangiocarcinoma.. Journal of hepatology, 2021.
- Single-cell analysis reveals fibroblast heterogeneity and myeloid-derived adipocyte progenitors in murine skin wounds.. Nature communications, 2019.
- Single-cell RNA-seq reveals that glioblastoma recapitulates a normal neurodevelopmental hierarchy.. Nature communications, 2020.
- Generalizing RNA velocity to transient cell states through dynamical modeling.. Nature biotechnology, 2020.
- Hodge Decomposition of Single-Cell RNA Velocity.. Journal of chemical information and modeling, 2024.
- Neural network-assisted RNA velocity imputation for empowering transcript dynamics-based analyses.. 2026.
- Interpretable learning of temporal cellular dynamics from single-cell data.. 2026.
- TIVelo: RNA velocity estimation leveraging cluster-level trajectory inference.. 2025.
- Challenges and progress in RNA velocity: Comparative analysis across multiple biological contexts.. 2026.
- spVelo: RNA velocity inference for multi-batch spatial transcriptomics data.. 2025.
- RNA velocity and beyond: Current advances in modeling single-cell transcriptional dynamics.. Allergology International, 2025.
- Dynamical Systems Model of RNA Velocity Improves Inference of Single-cell Trajectory, Pseudo-time and Gene Regulation. Journal of Molecular Biology, 2022.
- Deep generative modeling of transcriptional dynamics for RNA velocity analysis in single cells. Nature Methods, 2024.
- Inference of high-resolution trajectories in single-cell RNA-seq data by using RNA velocity. Cell Reports Methods, 2021.
This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.