# RNA Velocity vs. Pseudotime: When to Use Each Method for Inferring Cell State Transitions in Single-Cell Data


## Key Takeaways

- RNA velocity infers cell state transitions by modeling the ratio of unspliced to spliced mRNA, estimating the direction and speed of transcriptional change, whereas pseudotime orders cells along a learned trajectory based on transcriptional similarity.
- RNA velocity requires intron-aware alignment and quantification to distinguish unspliced (intronic) from spliced (exonic) reads, while pseudotime analysis can be performed on standard single-cell RNA sequencing count matrices.
- The steady-state assumption in simpler RNA velocity models posits equilibrium for most genes, while dynamical models relax this to capture active induction and repression phases, offering greater accuracy for rapidly changing biological processes.
- Pseudotime algorithms vary in their structural assumptions, with cluster-based methods like Slingshot requiring user-defined clusters and graph-based methods like PAGA preserving global topology, necessitating careful algorithm selection and validation.
- Both methods are sensitive to technical confounders such as cell cycle progression and batch effects, requiring rigorous quality control, normalization, and potentially data integration or regression strategies before trajectory inference.
- Combining RNA velocity and pseudotime analyses on the same dataset can reveal agreement or divergence in inferred transitions, with velocity providing local directionality and pseudotime offering global ordering, aiding in the comprehensive characterization of complex biological processes.

---

Researchers analyzing single-cell RNA sequencing data face a fundamental choice when studying developmental processes, disease progression, or cellular responses: should they use RNA velocity or pseudotime analysis to infer cell state transitions? The direct answer is that pseudotime analysis orders cells along a learned trajectory based on transcriptional similarity, while RNA velocity uses the ratio of unspliced to spliced messenger RNA to estimate the direction and speed of transcriptional change for each cell. Pseudotime answers the question of where a cell sits along a continuum of states, whereas RNA velocity answers the question of where a cell is heading next based on its current transcriptional activity. The choice between them depends on whether your data includes spliced and unspliced counts, whether the biological process follows a linear or branching structure, and whether you need local directionality or global ordering. This article provides a decision framework grounded in the computational biology literature and practical workflow guidance for laboratory professionals and bioinformatics trainees.

## The Conceptual Divide Between Ordering and Direction

Pseudotime analysis and RNA velocity address related but distinct biological questions. Pseudotime methods such as Monocle, Slingshot, and PAGA construct a manifold from single-cell transcriptomes and then place each cell along a path that represents a continuous biological process. The resulting pseudotime value is a scalar that orders cells from an inferred start state to an inferred end state. This ordering is useful for identifying genes whose expression changes progressively during differentiation, for locating branch points where cell fates diverge, and for comparing the relative maturity of cells within a population.

RNA velocity, by contrast, estimates the transcriptional dynamics of each gene by modeling the relationship between unspliced precursor messenger RNA and spliced mature messenger RNA. The core assumption is that an increase in unspliced counts relative to spliced counts indicates recent transcriptional activation, while a decrease indicates repression. By projecting these per-gene dynamics onto a low-dimensional embedding, RNA velocity produces a vector field that points from each cell toward its predicted future transcriptional state. This vector field provides local directionality that pseudotime cannot offer, because pseudotime assumes a global ordering instead of estimating cell-specific transitions.

The practical consequence of this distinction is that pseudotime is appropriate when you have prior knowledge that cells follow a defined progression, such as hematopoiesis or embryonic development, and when you want to characterize the continuum of states. RNA velocity is appropriate when you want to discover transitions without assuming a predefined path, when you suspect that multiple cell types may be transitioning simultaneously, and when your data includes the spliced and unspliced count information required for velocity estimation.

## Data Requirements and Library Preparation Constraints

The most immediate practical constraint is data availability. Pseudotime analysis requires only a standard single-cell RNA sequencing count matrix, which nearly all commercial platforms produce. RNA velocity requires additional information: the alignment and quantification pipeline must distinguish reads originating from unspliced pre-messenger RNA from reads originating from spliced mature messenger RNA. This distinction is achieved by detecting reads that span exon-exon junctions, which indicate spliced transcripts, versus reads that map to intronic regions, which indicate unspliced transcripts.

Standard 10x Genomics Chromium data can be used for RNA velocity if the alignment step retains intronic reads. The velocity estimation tools velocyto and scVelo both provide workflows for generating spliced and unspliced count matrices from standard single-cell FASTQ files. However, the quality of velocity estimates depends on sequencing depth and read length. Short reads that do not span exon-exon junctions may be misclassified, and low sequencing depth can produce sparse unspliced counts that lead to noisy velocity estimates.

Single-nucleus RNA sequencing presents a different consideration. Nuclei contain a higher proportion of unspliced pre-messenger RNA than whole cells, because nuclear RNA has not yet undergone export and cytoplasmic processing. This enrichment can improve the signal for RNA velocity, but it also changes the interpretation of the data. The transcriptional state captured by single-nucleus sequencing reflects nuclear RNA content, which may differ from the cytoplasmic mRNA pool that drives protein production. Researchers using single-nucleus data for velocity analysis should verify that their quantification pipeline handles intronic reads appropriately and should compare velocity results with known biological expectations.

For pseudotime analysis, the data requirement is simpler but still consequential. Pseudotime methods assume that the cells being analyzed represent a continuous sampling of a biological process. If the process is asynchronous, such as cells at different stages of the cell cycle, pseudotime can still recover the cycle. If the process is synchronous, such as a time-course experiment with discrete sampling points, pseudotime may struggle to order cells that are transcriptionally similar within each time point. The quality of pseudotime inference depends on the density of sampling along the true biological trajectory, and sparse sampling can produce artifacts.

## The Splicing Kinetics Model and Its Assumptions

RNA velocity methods model gene expression using a system of ordinary differential equations that describe the rates of transcription, splicing, and degradation. The simplest model, used by the original velocyto implementation, assumes that each gene has a constant transcription rate, a constant splicing rate, and a constant degradation rate. Under this model, the ratio of unspliced to spliced counts for each gene follows a predictable pattern: when transcription is induced, unspliced counts rise first, followed by spliced counts, when transcription is repressed, unspliced counts fall first.

The steady-state assumption, which underlies the original velocity model, states that the system has reached equilibrium for most genes in most cells. Under this assumption, the relationship between unspliced and spliced counts can be fit with a linear regression, and deviations from the regression line indicate active induction or repression. The scVelo package introduced a dynamical model that relaxes the steady-state assumption by estimating the full transcriptional dynamics for each gene, including the induction and repression phases. This dynamical model is more accurate for genes that are actively changing during the sampled process, but it requires more data and is computationally more intensive.

The limitations of the splicing kinetics model are well documented. The model assumes that splicing and degradation rates are constant across cells, which may not hold for genes subject to post-transcriptional regulation. The model also assumes that the observed unspliced and spliced counts accurately reflect the true transcriptional state, which can be violated by technical dropout, sequencing noise, and alignment errors. Recent work has extended velocity modeling to integrate additional data modalities, such as chromatin accessibility, to capture regulatory dynamics that RNA abundance alone cannot reveal. A 2025 study introduced a multi-omic relay velocity framework that infers velocity parameters at single-cell resolution without pre-assigned latent time, enabling locally adaptive estimation of gene expression kinetics across brain, skin, and blood cells. This approach distinguished chromatin-dependent and independent transcriptional regulation and identified asynchronous gene repression in neural progenitors, demonstrating that velocity analysis can move beyond the limits of RNA-only models.

For researchers considering whether to use RNA velocity, the key question is whether the splicing kinetics assumptions are reasonable for their biological system. If the process under study involves rapid transcriptional changes, such as an immune response or cell cycle progression, the steady-state assumption may fail and the dynamical model should be used. If the process involves slow developmental transitions, the steady-state assumption may be adequate for a subset of genes, but the velocity estimates should be validated against known biology.

## Pseudotime Algorithms and Their Structural Assumptions

Pseudotime methods vary in how they construct the trajectory and assign cell orderings. Monocle 2 uses reversed graph embedding to learn a principal graph that captures the branching structure of the data. Monocle 3 extends this approach to handle larger datasets and integrates with UMAP embeddings. Slingshot uses cluster-based lineage inference, identifying clusters of cells and then fitting smooth curves through the clusters to define lineages. PAGA builds a partition-based graph abstraction that preserves the global topology of the data while reducing noise, and it can be combined with pseudotime estimation along graph paths.

The choice of pseudotime algorithm affects the interpretation of results. Cluster-based methods such as Slingshot require the user to specify the number of clusters and the starting cluster, which introduces subjectivity. Graph-based methods such as PAGA are more automated but may produce trajectories that are difficult to interpret when the data contains many interconnected states. The LAIOR framework, introduced in a 2026 study, addresses the challenge of learning embeddings that simultaneously preserve local cell-state structure, global hierarchy, and smooth developmental trajectories. This hyperbolic neural ODE variational framework was benchmarked across 118 single-cell datasets and demonstrated improvements in manifold continuity, trajectory coherence, and embedding fidelity compared to 23 baseline methods. The study highlights a persistent problem in trajectory inference: classical methods emphasize either local neighborhoods or global variance, deep generative models cluster cell types well but often fracture trajectory continuity, and hyperbolic embeddings capture hierarchy but are numerically fragile.

For practical purposes, researchers should test multiple pseudotime algorithms and compare the resulting trajectories. If different algorithms produce substantially different orderings, the biological interpretation should be treated with caution. The robustness of pseudotime results can be assessed by subsampling the data, varying the number of nearest neighbors, and checking whether the inferred trajectory is stable across these perturbations.

## Comparing Velocity and Pseudotime on the Same Dataset

A practical approach for many research questions is to run both RNA velocity and pseudotime analysis on the same dataset and compare the results. This comparison can reveal whether the two methods agree on the direction of cell state transitions and can highlight regions of the trajectory where the methods diverge. A 2022 study introduced Cytopath, a trajectory inference method that uses RNA velocity information to define a Markov chain model and simulate an ensemble of possible differentiation trajectories. The study demonstrated that Cytopath can reconstruct differentiation trajectories with varying topologies, including bifurcated, circular, convergent, and mixed structures, and can assess the association of RNA velocity-based pseudotime with actually elapsed process time. The authors identified drawbacks in current state-of-the-art trajectory inference approaches, emphasizing that velocity-based pseudotime does not always correspond to real time.

This distinction between pseudotime and real time is critical. Pseudotime is an abstract ordering that reflects transcriptional similarity, not elapsed time. Two cells that are transcriptionally similar may be separated by different amounts of real time depending on the speed of the underlying biological process. RNA velocity provides an estimate of the direction and magnitude of transcriptional change, but the magnitude does not directly translate to a time scale. The Cytopath study showed that the association between velocity-based pseudotime and actually elapsed process time varies across datasets and topologies, meaning that researchers should not interpret pseudotime or velocity magnitudes as direct measures of time.

A 2026 study introduced PseudoVelo, a kinetics-free alternative that infers gene expression derivatives along pseudotime as pseudo-velocity. This method fits the expression of each gene as a function of pseudotime using generalized additive models, then calculates the derivative of gene expression with respect to pseudotime using a central difference approximation. PseudoVelo demonstrated strong performance compared to CellRank 2 across multiple developmental processes and showed high resilience against data perturbations. This approach is useful when spliced and unspliced counts are unavailable or when the splicing kinetics assumptions are violated, because it derives directionality from the pseudotime ordering itself instead of from splicing dynamics.

## Quality Control Before Trajectory Inference

The quality of trajectory inference depends on the quality of the input data. Standard single-cell quality control steps should be completed before running either pseudotime or RNA velocity. These steps include filtering cells by the number of detected genes, the total number of counts, and the percentage of mitochondrial reads. Cells with very low counts may be empty droplets or dying cells, while cells with very high mitochondrial reads may be stressed or apoptotic. The thresholds for these filters depend on the tissue type and the sequencing platform, and they should be determined by examining the distribution of these metrics across the dataset.

For RNA velocity, additional quality control is required. The spliced and unspliced count matrices should be examined for consistency, and genes with very low unspliced counts should be excluded from velocity estimation because their velocity estimates will be noisy. The proportion of unspliced reads varies by tissue and cell type, and datasets with very low overall unspliced content may not support reliable velocity estimation. The scVelo package provides diagnostic plots that show the phase portrait for each gene, which can be used to assess whether the splicing kinetics model fits the data.

Data integration is another consideration. If the dataset combines multiple samples, batches, or conditions, the integration method can affect trajectory inference. Batch effects can create artificial separation between cells from different samples, which can distort pseudotime orderings and velocity estimates. Integration methods such as Harmony, scVI, and Seurat's integration workflow can correct for batch effects, but the choice of integration method and the strength of correction should be documented and justified. The nf-core community provides standardized pipeline documentation that emphasizes reproducibility in single-cell analysis, and following these standards can help ensure that trajectory inference results are reproducible across runs and researchers.

## Imputation and Its Effects on Trajectory Inference

Technical dropout, where genes are not detected in some cells due to limited sequencing depth, is a pervasive challenge in single-cell RNA sequencing. Dropout can obscure true biological expression and create spurious zeros that distort trajectory inference. A 2026 study introduced scZN, an imputation framework that models observed single-cell data as a combination of RNA's two-state transcription process and dropout, then formulates imputation as nonnegative matrix factorization. The method decomposes the raw count matrix into interpretable nonnegative factors and applies constraints from prior knowledge and multiple regularizations to reconstruct the cellular expression landscape. The study showed that scZN captures the true distributional characteristics at both the gene and cell levels and suppresses spurious activation of genes that should not be expressed. In complex experimental design scenarios, scZN markedly improved trajectory inference for embryonic stem cells and mouse dentate gyrus data.

The decision to impute before trajectory inference is consequential. Imputation can fill in dropout zeros and improve the continuity of trajectories, but it can also introduce false signals if the imputation model makes incorrect assumptions. The scZN study emphasizes the importance of distinguishing biological zeros, where a gene is genuinely not expressed, from technical zeros, where a gene is expressed but not detected. Imputation methods that cannot make this distinction may activate genes that should remain silent, creating artifacts in downstream trajectory analysis.

For researchers using pseudotime or RNA velocity, the recommendation is to run the analysis on both the raw and imputed data and compare the results. If the trajectory structure is stable across both versions, the conclusions are more robust. If the trajectory changes substantially after imputation, the results should be interpreted with caution, and the imputation parameters should be examined.

## Cell Cycle Effects and Other Confounders

Cell cycle progression is a common confounder in trajectory inference. Cells in different phases of the cell cycle have distinct transcriptional programs, and these programs can create apparent trajectories that reflect cell cycle position instead of the biological process of interest. Many single-cell analysis workflows include a cell cycle scoring step that assigns each cell to a phase based on the expression of known cell cycle genes. Researchers can then either regress out the cell cycle effect or exclude cycling cells from the trajectory analysis.

The choice between regression and exclusion depends on the biological question. If the process of interest is cell cycle progression itself, such as studying the transcriptional dynamics of the cell cycle, then the cell cycle signal should be preserved. If the process of interest is a differentiation trajectory that is confounded by cell cycle, regression may be appropriate. However, regression can remove real biological signal if the cell cycle genes overlap with the genes driving the differentiation process.

RNA velocity is particularly sensitive to cell cycle effects because the splicing kinetics of cell cycle genes are tightly regulated. The velocity estimates for cell cycle genes may dominate the vector field, pulling cells toward the next cell cycle phase instead of toward the differentiation trajectory. Some velocity implementations allow the user to exclude cell cycle genes from the velocity estimation, and this option should be considered when the biological process of interest is not cell cycle related.

## Choosing Between Velocity and Pseudotime for Specific Biological Questions

The decision framework for choosing between RNA velocity and pseudotime can be organized around the biological question, the data availability, and the expected trajectory structure.

For identifying the direction of cell state transitions without prior knowledge of the trajectory, RNA velocity is the preferred method. The vector field provides local directionality that can reveal which cell types are transitioning into which, even when the trajectory has complex topology. This is particularly useful for discovering unexpected transitions, such as dedifferentiation or transdifferentiation events.

For characterizing the continuum of states along a known developmental path, pseudotime is the preferred method. The scalar ordering allows for straightforward visualization of gene expression changes along the trajectory and for identifying genes that are upregulated or downregulated at specific stages. Pseudotime also provides a natural framework for differential expression analysis between early and late states.

For analyzing branching trajectories where cells diverge into multiple fates, both methods can be used in combination. Pseudotime can identify the branch point and the genes that are differentially expressed between branches, while RNA velocity can confirm the direction of the divergence and identify the transcriptional programs that drive each fate.

For analyzing circular trajectories, such as the cell cycle or circadian rhythms, pseudotime methods that support circular topologies should be used. Standard pseudotime methods that assume a linear or branching structure will produce artifacts when applied to circular processes. RNA velocity can handle circular trajectories because the vector field naturally follows the cycle, but the interpretation of velocity magnitude along a cycle requires care.

For analyzing datasets without spliced and unspliced counts, pseudotime is the only option unless a kinetics-free velocity method such as PseudoVelo is used. PseudoVelo derives directionality from the pseudotime ordering itself, making it applicable to standard count matrices, but it inherits the limitations of the underlying pseudotime ordering.

## The At a Glance Decision Table

| Consideration | RNA Velocity | Pseudotime |
| --- | --- | --- |
| Primary question answered | Where is each cell heading next based on current transcriptional activity? | Where does each cell sit along a continuum of transcriptional states? |
| Data requirement | Spliced and unspliced count matrices from intron-aware alignment | Standard single-cell count matrix |
| Trajectory structure | Handles complex topologies including bifurcations, cycles, and convergent transitions | Best suited for linear or branching trajectories with defined start and end states |
| Key assumption | Splicing and degradation rates are constant and can be modeled from unspliced to spliced ratios | Cells are sampled continuously along the biological process |
| Main limitation | Sensitive to sequencing depth, dropout, and violations of splicing kinetics assumptions | Does not provide local directionality and may not reflect real elapsed time |
| Recommended use case | Discovering transitions, validating trajectory direction, analyzing multi-omic data | Characterizing gene expression changes along a known developmental path, identifying branch points |

## Practical Workflow for Trajectory Inference

The following workflow provides a structured approach for researchers deciding between RNA velocity and pseudotime analysis.

First, assess the data availability. Check whether the alignment and quantification pipeline produced spliced and unspliced count matrices. If not, determine whether the raw FASTQ files can be re-aligned with an intron-aware pipeline. The Galaxy Training Network provides accessible tutorials for single-cell RNA sequencing analysis that include intron-aware alignment and velocity estimation workflows. If re-alignment is not feasible, pseudotime analysis is the only option.

Second, perform standard quality control. Filter cells by gene count, total count, and mitochondrial read percentage. Examine the distribution of these metrics and set thresholds based on the data. Document the filtering decisions and the number of cells retained at each step.

Third, normalize the data and identify highly variable genes. The choice of normalization method affects downstream analysis, and the method should be appropriate for the data type. For single-cell data, log-normalization with a scale factor is standard, but alternative methods such as SCTransform may perform better for datasets with high technical variation.

Fourth, run pseudotime analysis using at least two different algorithms. Compare the resulting trajectories and assess their agreement. If the algorithms produce substantially different orderings, investigate the source of the discrepancy. The Bioconductor project provides official documentation for many trajectory inference packages, including Slingshot and Monocle, and the documentation includes guidance on parameter selection and interpretation.

Fifth, if spliced and unspliced counts are available, run RNA velocity analysis. Use the dynamical model in scVelo instead of the steady-state model, because the dynamical model is more accurate for genes that are actively changing. Examine the phase portraits for key genes to assess whether the splicing kinetics model fits the data.

Sixth, compare the velocity vector field with the pseudotime trajectory. Project the velocity vectors onto the same embedding used for pseudotime and check whether the vectors point along the pseudotime trajectory. Regions where the velocity vectors point against the pseudotime ordering may indicate incorrect pseudotime assignment or complex dynamics such as dedifferentiation.

Seventh, validate the results against known biology. If the dataset includes cells from a well-characterized developmental system, check whether the inferred trajectory matches the known developmental path. If the trajectory contradicts established biology, investigate the cause before drawing conclusions.

## Records and Measurements for Reproducible Analysis

Reproducibility in trajectory inference requires careful record keeping. The following records should be maintained for each analysis:

The version numbers of all software packages used, including the alignment tool, the quantification tool, the normalization method, and the trajectory inference algorithm. The nf-core documentation emphasizes the importance of version pinning and containerization for reproducible workflows, and these practices should be applied to single-cell analysis.

The exact parameters used for each step, including quality control thresholds, normalization parameters, the number of principal components used for dimensionality reduction, the number of nearest neighbors, and the resolution parameter for clustering. These parameters should be recorded in a configuration file or a notebook.

The random seed used for any stochastic steps, such as clustering or dimensionality reduction. Different random seeds can produce different results, and the seed should be recorded to allow exact reproduction.

The input data files and their checksums, to ensure that the same data is used for all analyses.

The output files from each step, including the filtered count matrix, the normalized data, the embedding coordinates, and the trajectory assignments.

The Carpentries lessons provide foundational training in data organization and reproducible analysis practices, and these skills are directly applicable to single-cell trajectory inference. Researchers who follow these practices can share their analysis with confidence that others can reproduce the results.

## Common Failure Patterns in Trajectory Inference

Several failure patterns recur in trajectory inference analyses, and recognizing them can prevent incorrect biological conclusions.

The first failure pattern is the interpretation of pseudotime as real time. Pseudotime is an abstract ordering based on transcriptional similarity, and it does not correspond to elapsed time. Two cells that are far apart in pseudotime may be separated by a short real time interval if the transcriptional changes are rapid, and two cells that are close in pseudotime may be separated by a long real time interval if the transcriptional changes are slow. The Cytopath study demonstrated that the association between velocity-based pseudotime and actually elapsed process time varies across datasets, and researchers should not assume a linear relationship.

The second failure pattern is the overinterpretation of velocity magnitudes. The length of a velocity vector reflects the estimated rate of transcriptional change, but this rate is not calibrated to real time. A long velocity vector does not mean that the cell will transition quickly, and a short velocity vector does not mean that the cell is static. The velocity magnitude is also sensitive to the scaling of the embedding, and comparing velocity magnitudes across different embeddings is not meaningful.

The third failure pattern is the use of velocity on data with insufficient unspliced counts. If the proportion of unspliced reads is very low, the velocity estimates will be dominated by noise, and the resulting vector field will be unreliable. Diagnostic plots that show the distribution of unspliced counts and the phase portraits for key genes should be examined before interpreting velocity results.

The fourth failure pattern is the application of pseudotime to data with disconnected clusters. If the cells form discrete clusters with no continuous bridge between them, pseudotime algorithms will force a trajectory through the gaps, creating an artificial ordering. The connectivity of the data should be assessed before running pseudotime, and disconnected clusters should be analyzed separately or the trajectory should be interpreted with caution.

The fifth failure pattern is the neglect of batch effects. If the dataset combines multiple samples or batches, the batch effects can create artificial trajectories that reflect technical variation instead of biology. Integration should be performed before trajectory inference, and the results should be examined for batch-specific patterns.

The sixth failure pattern is the failure to validate against known biology. Trajectory inference methods can produce plausible-looking trajectories that contradict established biology. Validation against known markers, known developmental paths, or independent experimental data is essential for building confidence in the results.

## Multi-Omic Integration and Emerging Methods

The field of trajectory inference is moving toward multi-omic integration, where multiple data modalities are measured in the same cells or matched across cells. The 2025 MoFlow study demonstrated that integrating chromatin accessibility data with RNA expression can reveal regulatory dynamics that RNA-only velocity models miss. The study applied the relay velocity framework to single-cell multi-omic datasets from brain, skin, and blood cells and identified chromatin-dependent and independent transcriptional regulation, validated transcription repression models, and identified asynchronous gene repression in neural progenitors.

A 2026 study on early zebrafish development used a multimodal measurement of full-length transcriptome and histone modifications in individual cells to show that chromatin and transcription states are uncoupled before germ layer formation and become progressively connected during gastrulation and somitogenesis. The study found that silencing of developmental genes is achieved by local spreading of repressive chromatin together with cell type-specific demethylation. This work demonstrates that trajectory inference based on RNA alone may miss important regulatory events that occur at the chromatin level.

For researchers considering multi-omic trajectory inference, the practical considerations include the availability of multi-omic data, the computational resources required for integration, and the interpretability of the resulting models. The MoFlow study emphasizes that existing RNA velocity models primarily rely on RNA abundance and globally inferred latent time, which limits their ability to capture local regulatory dynamics. Multi-omic approaches that infer velocity parameters at single-cell resolution without pre-assigned latent time can provide a more comprehensive and locally adaptive estimation of gene expression kinetics.

The LAIOR framework, introduced in a 2026 study, addresses the challenge of learning embeddings that preserve local cell-state structure, global hierarchy, and smooth developmental trajectories. The framework combines Lorentz geometric regularization for tree-like latent hierarchy, a dual-path information bottleneck for coordinated biological programs, and neural ODE regularization for stable latent trajectories. The study benchmarked LAIOR across 118 single-cell datasets and demonstrated improvements in manifold continuity, trajectory coherence, and embedding fidelity. This work highlights the ongoing effort to develop trajectory inference methods that are both accurate and interpretable.

## Application Examples from the Literature

Several recent studies illustrate the practical application of pseudotime and RNA velocity to biological questions.

A 2026 study of CD27-positive cytotoxic T cells in colorectal cancer used pseudotime trajectory analysis to reveal progressive exhaustion pathways with three distinct differentiation fates. The study found that exhausted T cells exhibited elevated expression of checkpoint molecules and reduced cytotoxic capacity compared to effector populations. RNA velocity and PAGA connectivity analysis demonstrated that exhaustion represents a terminal differentiation state with limited plasticity. This study exemplifies the combined use of pseudotime and velocity to characterize a disease-relevant trajectory.

A 2026 study of tumor-initiating cells in histologically normal colonic mucosa used single-cell RNA sequencing with multiple trajectory inference methods, including Monocle2, CytoTRACE, and RNA velocity. The study identified tumor-specific stem-like populations enriched in adenomas and localized tumor-initiating cell subsets to the root of lineage trajectories leading to polyp-enriched states. The study identified ETS2, SLC12A2, and LEFTY1 as tumor-initiating cell-specific markers. This study demonstrates the power of combining multiple trajectory inference methods to identify early events in neoplastic transformation.

A 2026 study of oxidative stress in benign prostatic hyperplasia integrated single-cell RNA sequencing, spatial transcriptomics, and machine learning to characterize oxidative stress-linked stromal states. The study found elevated oxidative stress scores across cell types, with the highest scores in fibroblasts, and identified ACOX2-high fibroblasts that were expanded in benign prostatic hyperplasia and exhibited trajectory-associated increases in ACOX2 expression. This study exemplifies the integration of trajectory inference with spatial and multi-omic data.

A 2026 study of ferroptosis-pyroptosis crosstalk in sepsis-induced acute respiratory distress syndrome used single-cell RNA sequencing to characterize cell type-specific expression patterns of cell death pathway genes. The study identified 10 ferroptosis-pyroptosis crosstalk genes with differential expression between ARDS and sepsis-only groups, with enrichment of selected signals in myeloid cells. This study demonstrates the application of single-cell analysis to identify prognostic biomarkers and therapeutic targets.

These examples illustrate the range of biological questions that can be addressed with trajectory inference methods, from developmental biology to cancer immunology to inflammatory disease.

## Limitations and Interpretation Boundaries

Both RNA velocity and pseudotime analysis have inherent limitations that should be acknowledged in any research report.

Pseudotime analysis assumes that the cells in the dataset represent a continuous sampling of the biological process. If the process involves discrete states with rapid transitions, the intermediate states may be underrepresented, and the pseudotime ordering will be unreliable. Pseudotime also assumes that the transcriptional similarity between cells reflects their proximity along the biological trajectory, which may not hold if the trajectory involves large transcriptional jumps or if unrelated cell types are transcriptionally similar.

RNA velocity assumes that the splicing kinetics model accurately describes the transcriptional dynamics of each gene. This assumption is violated for genes with complex regulation, such as genes with multiple isoforms, genes subject to nonsense-mediated decay, or genes with cell type-specific splicing rates. The velocity estimates for these genes will be inaccurate, and the overall vector field may be distorted.

Both methods are sensitive to the choice of embedding. The low-dimensional representation of the data, typically UMAP or t-SNE, affects the local neighborhood structure and therefore affects both pseudotime ordering and velocity projection. Different embeddings can produce different trajectories, and the choice of embedding should be justified and documented.

The interpretation of trajectory inference results should always be framed as hypothesis generation instead of confirmation. Trajectory inference methods produce computational predictions about cell state transitions, and these predictions should be validated with independent experimental approaches, such as lineage tracing, time-course experiments, or perturbation studies.

## Professional Escalation Criteria

Researchers should consider escalating their analysis to a more specialized bioinformatics team or seeking expert consultation in the following situations:

If the trajectory inference results are inconsistent across multiple algorithms or parameter settings, and the source of the inconsistency cannot be identified, the analysis may require expert review.

If the velocity estimates are extremely noisy or the phase portraits for key genes do not show the expected patterns, the data quality may be insufficient for velocity analysis, and re-sequencing or re-alignment may be necessary.

If the dataset includes multiple batches or samples with strong batch effects that cannot be corrected with standard integration methods, specialized integration approaches may be required.

If the biological process under study involves complex regulatory mechanisms, such as chromatin remodeling or post-transcriptional regulation, multi-omic integration may be necessary, and this analysis typically requires specialized expertise.

If the trajectory inference results contradict established biology, the analysis should be reviewed before any conclusions are drawn.

The EMBL-EBI Training portal provides learning pathways for bioinformatics analysis, including single-cell RNA sequencing and trajectory inference, and researchers who need to build their skills can use these resources. The NCBI provides access to the Gene Expression Omnibus and other databases where single-cell datasets are deposited, allowing researchers to compare their results with published data.

## Frequently Asked Questions

### What is the main difference between RNA velocity and pseudotime?

RNA velocity estimates the direction and speed of transcriptional change for each cell based on the ratio of unspliced to spliced messenger RNA. Pseudotime orders cells along a learned trajectory based on transcriptional similarity. RNA velocity provides local directionality, while pseudotime provides a global ordering.

### Can I run RNA velocity on standard single-cell RNA sequencing data?

Yes, if the raw FASTQ files are re-aligned with an intron-aware pipeline that distinguishes spliced from unspliced reads. Standard 10x Genomics data can be used for velocity estimation, but the quality of the estimates depends on sequencing depth and read length.

### What should I do if my data does not have spliced and unspliced counts?

You can use pseudotime analysis, which requires only a standard count matrix. Alternatively, you can use a kinetics-free velocity method such as PseudoVelo, which derives directionality from the pseudotime ordering itself.

### How do I choose between the steady-state and dynamical models for RNA velocity?

The steady-state model assumes that the splicing system has reached equilibrium for most genes, while the dynamical model estimates the full transcriptional dynamics for each gene. The dynamical model is more accurate for genes that are actively changing during the sampled process, but it requires more data and is computationally more intensive.

### Can RNA velocity and pseudotime be used together?

Yes, running both methods on the same dataset and comparing the results can reveal whether the methods agree on the direction of cell state transitions. The velocity vector field can be projected onto the pseudotime trajectory to check for consistency.

### How should I validate my trajectory inference results?

Validate against known biology, such as known developmental markers or established differentiation paths. Compare results across multiple algorithms and parameter settings. If possible, validate with independent experimental approaches such as lineage tracing or time-course experiments.

### What are the most common mistakes in trajectory inference?

Interpreting pseudotime as real time, overinterpreting velocity magnitudes, using velocity on data with insufficient unspliced counts, applying pseudotime to disconnected clusters, neglecting batch effects, and failing to validate against known biology.

### When should I seek expert help for trajectory inference?

Seek expert help when results are inconsistent across algorithms, when velocity estimates are extremely noisy, when batch effects cannot be corrected, when multi-omic integration is needed, or when results contradict established biology.

## Related Bioinformatics Guides

- [Single-Cell RNA Velocity: Inferring Cellular Dynamics](/knowledge/bioinformatics/single-cell-rna-velocity-inferring-cellular-dynamics)
- [Single-Cell Sequencing Workflow: From Sample Preparation to Data Analysis](/knowledge/bioinformatics/single-cell-sequencing-workflow-from-sample-preparation-to-data-analysis)
- [Single-Cell Sequencing Analysis Pipeline: From Raw Data to Biological Insights](/knowledge/bioinformatics/single-cell-sequencing-analysis-pipeline-from-raw-data-to-biological-insights)
- [Single-Cell RNA Sequencing Quality Control: A Practical Guide to Filtering and Metrics](/knowledge/bioinformatics/single-cell-rna-sequencing-quality-control-a-practical-guide-to-filtering-and-metrics)
- [Single-Cell RNA Sequencing Depth: A Cost-Benefit Analysis for Experimental Design](/knowledge/bioinformatics/single-cell-rna-sequencing-depth-a-cost-benefit-analysis-for-experimental-design)

## Related Clinical & Scientific Guides

* [A Practical Guide to Detecting Antimicrobial Resistance Genes in Shotgun Metagenomic Data](/knowledge/bioinformatics/a-practical-guide-to-detecting-antimicrobial-resistance-genes-in-shotgun-metagenomic-data)
* [Computational Immunology: Modeling the Immune System](/knowledge/bioinformatics/computational-immunology-modeling-the-immune-system)
* [How to Set Hard Filters for Germline Variant Calling: A Practical Guide to GATK Best Practices](/knowledge/bioinformatics/how-to-set-hard-filters-for-germline-variant-calling-a-practical-guide-to-gatk-best-practices)


## References and Further Reading

- [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information.
- [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute.
- [Bioconductor](https://bioconductor.org/). Bioconductor Project.
- [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project.
- [nf-core Documentation](https://nf-co.re/docs). nf-core.
- [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries.
- [Multi-omic relay velocity modeling uncovers dynamic chromatin-transcription regulation across cell states.](https://doi.org/10.1038/s41467-025-67259-6). 2025.
- [LAIOR: a hyperbolic neural ODE variational framework for interpretable single-cell manifold learning and trajectory inference.](https://doi.org/10.3389/fgene.2026.1838613). 2026.
- [Prior-guided factorization for reliable imputation of scRNA-seq data.](https://doi.org/10.1371/journal.pcbi.1014051). 2026.
- [Ferroptosis-pyroptosis crosstalk signature as prognostic biomarkers and therapeutic targets in sepsis-induced ARDS.](https://doi.org/10.3389/fcell.2026.1810674). 2026.
- [Single-cell co-mapping reveals relationship between chromatin state and gene expression in early zebrafish development.](https://doi.org/10.7554/elife.110400). 2026.
- [Single-cell transcriptomic profiling reveals CD27&lt,sup&gt,+&lt,/sup&gt, cytotoxic T Cell heterogeneity and exhaustion dynamics in colorectal cancer tumor microenvironment.](https://doi.org/10.3389/fgene.2026.1808171). 2026.
- [Identification of tumor initiating cells and early marker genes in histologically normal colonic mucosa that lead to neoplastic transformation.](https://doi.org/10.1016/j.neo.2026.101300). 2026.
- [Oxidative stress reprograms benign prostatic hyperplasia microenvironments: insights from integrative multi-omics and machine learning.](https://doi.org/10.1186/s12967-026-08055-8). 2026.
- [Simulation-based inference of differentiation trajectories from RNA velocity fields](https://doi.org/10.1016/j.crmeth.2022.100359). Cell Reports Methods, 2022.
- [PseudoVelo: Inferring Gene Expression Derivatives Along Pseudotime as Pseudo-Velocity](https://doi.org/10.3390/ijms27146420). International Journal of Molecular Sciences, 2026.

> This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.