To understand why the new model matters, it helps to revisit what RNA velocity actually measures. Standard single-cell RNA sequencing counts mature, spliced transcripts, but many protocols also capture unspliced nascent RNA that has not yet been processed. Because unspliced RNA reflects genes that a cell has recently started transcribing, the ratio between unspliced and spliced counts for each gene provides a hint about whether that gene is being induced or repressed. By combining these signals across thousands of genes, RNA velocity methods can infer a vector field over the cell population, pointing each cell toward its likely future state. This allows researchers to reconstruct developmental trajectories, identify lineage relationships, and predict the endpoints of differentiation without needing time-series measurements.

The original formulation of RNA velocity relied on a dynamical model of splicing kinetics, treating the interplay between unspliced and spliced RNA as a set of ordinary differential equations whose parameters could be estimated from the data. While elegant, this approach struggles with the noise, dropout, and sparsity that plague single-cell datasets, and it often produces inconsistent velocity estimates across complex, multi-lineage systems. Subsequent methods have taken different routes: some simplify the problem by assuming steady-state kinetics, while others turn to neural networks to learn the velocity field directly from latent representations of gene expression. These neural approaches have improved robustness, but they come with their own limitations, particularly in how they model relationships between cells.

A key weakness of many current neural velocity models, the authors of the new study argue, lies in their reliance on local neighborhood structure. Most pipelines begin by constructing a k-nearest neighbor graph in a reduced-dimensional space, typically using principal component analysis to compress the high-dimensional expression profiles. The graph is meant to approximate the local geometry of the transcriptional manifold, and velocity information is propagated only between neighboring cells. Yet in real tissues, biologically related cells may not be nearest neighbors in the embedding space. Cells that lie far apart in the graph can still share developmental ancestry or be destined for similar fates, and these long-range dependencies carry information that a purely local model discards. Capturing them requires an architecture capable of attending to arbitrary pairs of cells rather than only to immediate neighbors.

That is precisely where the transformer comes in. Transformers, the architecture family that underpins modern large language models, are built around self-attention, a mechanism that lets each element in a set compute weighted relationships with every other element. Multi-head attention extends this by running several attention computations in parallel, each with its own learned projection, so that different heads can specialize in different kinds of relationships. Applied to single-cell data, a graph transformer can in principle learn which cells are relevant to one another regardless of their distance in the neighborhood graph, while the graph structure itself provides an inductive bias that keeps the model grounded in the observed topology of the data. GTVelo combines these ingredients within a neural ordinary differential equation framework, meaning that the learned attention-based representation feeds into a continuous-time dynamical model of how cell states evolve.

The neural ordinary differential equation component is significant in its own right. Rather than discretizing time into steps, a neural ODE defines the derivative of the cell’s latent state as the output of a neural network, and the trajectory is obtained by integrating that derivative forward. This continuous formulation is a natural fit for RNA velocity, which is fundamentally about the instantaneous rate of change of transcriptional state. It also allows the model to produce smooth trajectories through latent space, which is important when the underlying biology involves gradual differentiation rather than abrupt jumps between discrete states. By embedding the transformer-based encoder inside this dynamical system, GTVelo seeks to couple rich relational modeling of the cell population with principled temporal dynamics.

One of the more distinctive features of the new model is what the authors call a multi-origin state mechanism. Many existing approaches effectively assume that all trajectories in a dataset can be described starting from a single initial state, which becomes problematic in datasets containing multiple lineages, branching differentiation paths, or heterogeneous cell populations. A single origin forces the model to squeeze fundamentally different developmental stories into one shared starting point, degrading the accuracy of velocity estimates for every lineage involved. GTVelo instead allows inference to proceed from multiple origins, which the authors say enables flexible handling of multi-branch trajectories. This flexibility also extends to data integration: the model is designed to accommodate multi-omics inputs, reflecting the growing trend in genomics toward assays that measure RNA alongside chromatin accessibility, protein abundance, or other molecular layers in the same cells.

According to the study, GTVelo was evaluated across multiple benchmark datasets and compared against established RNA velocity methods using standard evaluation metrics. The reported results show competitive performance overall, with GTVelo outperforming the compared methods on most of the standard metrics in most cases. Two of the evaluation concepts highlighted in the paper are cross-boundary directedness, which measures whether inferred velocities correctly point across the boundaries between clusters in a biologically consistent direction, and in-cluster coherence, which assesses whether velocity vectors within a cluster are mutually consistent rather than pointing in conflicting directions. Together, these metrics probe whether a velocity model captures both the global organization of a trajectory and the local smoothness of the inferred dynamics, two properties that are often in tension.

The implications for downstream biology could be substantial. RNA velocity is widely used to annotate developmental hierarchies, trace the origins of disease-associated cell states, and generate hypotheses about lineage commitment in contexts ranging from embryogenesis to tumor evolution. Models that handle branching trajectories and long-range cellular associations more reliably could improve the fidelity of these analyses, particularly in complex tissues where multiple lineages coexist and where rare transitional populations carry outsized biological importance. The stated ability to integrate multi-omics data points toward applications in which velocity is estimated from joint measurements, potentially revealing regulatory dynamics that transcript-only models cannot see. As with any computational method, real-world performance will depend on how the benchmarks translate to noisy, dataset-specific conditions, and independent validation by the community will be an important next step.

The research was carried out by Shensi Huang, Hongyu Zhang, and Jianping Zhao of Xinjiang University, together with Zile Wang of Dalian University of Technology, Haiyun Wang of Wuhan University of Science and Technology, and Junfeng Xia of Anhui University, and was supported by funding including grants from the National Natural Science Foundation of China. The article was published open access in BMC Bioinformatics on 15 September 2026, with the accepted manuscript shared early under a permanent DOI. For a field that has seen a rapid succession of velocity methods, each claiming incremental gains, GTVelo’s contribution lies in bringing the attention machinery of modern deep learning to bear on a problem that has long been constrained by local graph assumptions and single-origin trajectory models. Whether graph transformers become a standard component of the single-cell dynamics toolbox, the study adds a technically grounded option for researchers wrestling with the messy, branching reality of cellular differentiation.

Subject of Research: RNA velocity inference in single-cell transcriptomics using a graph transformer-based neural ordinary differential equation model

Article Title: RNA velocity inference based on graph transformer

Article References: Huang, S., Wang, Z., Zhang, H., Wang, H., Zhao, J., & Xia, J. (2026). RNA velocity inference based on graph transformer. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06550-9

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06550-9

Keywords: RNA velocity, single-cell RNA sequencing, graph transformer, neural ordinary differential equations, trajectory inference, cell differentiation, deep learning, multi-head attention, multi-omics integration, transcriptomics, computational biology, BMC Bioinformatics

The problem veloAgent tackles is well documented. Existing tools such as scVelo, cellDancer, DeepVelo, veloVI, VeloVAE, UniTVelo, and SIRV each improve some aspect of velocity estimation, but most assume transcriptional kinetics that are uniform across cells, an oversimplification that erases the heterogeneity driving development and disease. More sophisticated models that estimate gene- and cell-specific rates, like cellDancer, become computationally prohibitive on atlas-scale datasets. And nearly all methods rely on embeddings of transcriptional similarity alone, discarding the tissue microenvironment. The result is velocity fields that contradict known tissue structure, misdirect flows from tumor cells toward healthy neighbors, or push neurons toward neuroblasts they can never become.

veloAgent is a hybrid of three components. First, a variational autoencoder compresses noisy spliced and unspliced count matrices from each cell into a shared latent representation, denoising the data while preserving the features that mark dynamic cellular states. Second, a deep neural network constrained by protein–protein interactions from the STRING database takes that latent vector and infers transcription, splicing, and degradation rates—α, β, and γ—for every gene in every cell. The biologically informed sparsity matters: each gene’s output depends only on its annotated interactors, grounding the learned kinetics in known regulatory relationships. These parameters feed ordinary differential equations that yield initial velocity estimates, which are then refined by aligning each cell’s vector with those of its transcriptionally similar neighbors.

The third component is the distinctive one. An agent-based model treats every cell as an autonomous agent sitting at its true position in the tissue, and updates its velocity using local interaction rules. Three factors shape each update: expression similarity between neighbors, local cell density, and inverse spatial distance, with closer and more transcriptionally alike cells exerting stronger influence. The result is a velocity field that respects tissue architecture rather than floating free of it. Crucially, the authors emphasize that the projected arrows represent inferred transcriptional state transitions, not physical cell migration—a distinction that keeps the spatial maps honest while still revealing how tissue organization constrains and guides cellular trajectories.

The benchmarking is extensive. The team applied veloAgent to four spatial transcriptomics datasets spanning different platforms and resolutions: a Stereo-seq mouse brain, a human breast cancer Visium sample, a developing chicken heart, and a HybISS mouse brain. In the Stereo-seq brain, competing methods produced unclear or contradictory velocities around fiber-tract cells, while veloAgent captured coherent transitions from hippocampal, thalamic sensory-motor, and polymodal association regions consistent with known neuronal projections. In breast cancer, SIRV incorrectly directed flow from tumor cells toward normal cells; veloAgent corrected the misprediction. In the chicken heart, it fixed reversed trajectories at the ventricle–atrium boundary, and in the HybISS brain it correctly modeled neurons as terminal states rather than steering them toward neuroblasts.

Quantitatively, veloAgent was evaluated against seven leading methods on three metrics: cross-boundary direction correctness, fate probability, and velocity confidence. It achieved the highest or near-highest scores across all datasets. For cross-boundary direction, it outperformed all competitors by an average of 92.03 percent on the Stereo-seq brain, 92.02 percent on breast cancer, 55.52 percent on the chicken heart, and 12.15 percent on the HybISS brain. Fate probability gains averaged between 33 and 66 percent, and velocity confidence remained near-perfect, at 0.998 or above, across every dataset. Marker-gene analyses reinforced the picture: Mbp velocity stayed confined to myelinating fiber tracts, ERBB2 velocity localized to tumor regions rather than macrophages, NPPA velocity remained restricted to the atria, and Nrg1 velocity marked neurons specifically.

The improved velocities also sharpen downstream biology. Feeding veloAgent’s output into CellRank, the researchers recovered lineage-specific driver genes ordered correctly along pseudotime. In the developing mouse brain, early En1 expression aligned with progenitor states, Dlk1 peaked mid-trajectory alongside pro-neural factors, and late genes such as Lhx5 and Klhl14 marked terminal neuronal populations, with Pax8 emerging as a key mid-hindbrain regulator. In the chicken heart, transient valve progenitors enriched for DICER1 were distinguished from terminal valve cells expressing extracellular matrix genes like COL1A1 and SPARC. Gene Ontology enrichment confirmed the programs: circulatory-system development and mesenchymal development in the heart, neuron differentiation and nervous-system development in the brain. SCENIC analysis further recovered known transcription factors, including Foxj1, Lmx1a, Msx1, and Msx2 in choroid-plexus patterning, Nfia and Tcf12 in fiber-tract formation, and Foxf1 and Gata4 in endoderm and cardiac lineages.

The most provocative feature, however, is the in silico perturbation module—something no prior velocity framework offers. Because veloAgent explicitly parameterizes α, β, and γ for each gene, researchers can set a gene’s transcription rate to zero after training, without retraining, and recompute the entire velocity field. The change in cross-boundary directionality then quantifies that gene’s causal influence on fate progression. Applied to the mouse brain, silencing top-ranked genes reversed differentiation toward the fiber-tract fate, and the perturbed genes were enriched for neurogenesis, axonogenesis, and synaptic organization. Applied to breast cancer, the top hits included CARD14 and ERBB2, both established oncogenes whose virtual silencing redirected velocity vectors away from malignant clusters. Kaplan–Meier analyses showed that patients with high expression of these genes had significantly poorer overall survival, and ERBB2 is already the target of FDA-approved drugs including trastuzumab and trastuzumab deruxtecan—a striking validation of the computational predictions.

Scalability is the other headline advance. veloAgent’s architecture scales with the number of genes, which is essentially fixed per organism, rather than with cell count, which grows exponentially with sequencing throughput. When the team duplicated the HybISS dataset from 50,000 up to one million cells, cellDancer’s memory usage climbed near-cubically to roughly 120 gigabytes, while veloAgent required only about 37 gigabytes with near-constant per-cell runtime. The agent-based module is also fully modular: applied post hoc to scVelo’s velocity estimates, it lifted velocity confidence dramatically across all four datasets—for example from 0.0677 to 0.9943 on the Stereo-seq brain—without retraining the underlying model. Ablation experiments confirmed that removing any component, especially the spatial ABM, significantly degraded performance.

The authors are candid about limitations. Visium spots can capture transcripts from multiple cells, so inferred dynamics represent averaged states, though consistent results on higher-resolution platforms suggest the signals are robust. The perturbation framework currently handles single genes rather than combinatorial interventions, interaction databases like STRING remain incomplete, and the model operates at gene rather than exon level. It has also not yet been run on a true million-cell atlas. Even so, veloAgent marks a conceptual shift: RNA velocity becomes a spatially coherent, causally probeable model of tissue dynamics rather than an abstract trajectory plot. For developmental biologists mapping organogenesis, oncologists tracing tumor evolution, and regenerative medicine researchers seeking to steer stem cells toward desired fates, the ability to simulate a gene knockout in seconds and watch the cellular movie change direction could reshape how hypotheses are generated—and which experiments get run first.

Subject of Research: A spatially informed RNA velocity framework combining deep generative modeling and agent-based simulation to infer and perturb cell state transitions in tissues

Article Title: Dissecting and steering cell dynamics using spatially-informed RNA velocity with veloAgent

Article References: Raghavan, V., Yoon, B., Fonseca, G. J., Li, Y., & Ding, J. (2026). Dissecting and steering cell dynamics using spatially-informed RNA velocity with veloAgent. Molecular Systems Biology, 22(7), 1180-1200. https://doi.org/10.1038/s44320-026-00213-w

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00213-w

Keywords: RNA velocity, spatial transcriptomics, single-cell RNA sequencing, agent-based modeling, variational autoencoder, computational biology, cell fate, in silico perturbation, breast cancer, ERBB2, developmental biology, therapeutic target discovery

That blind spot matters more than many researchers realized. A cell’s fate is not determined solely by the genes it expresses; it is shaped by its position within a tissue, by the density of surrounding cells, by physical constraints, and by local signaling interactions. Conventional RNA velocity methods treat cells as independent entities floating in an abstract expression space, and the result is often inferred dynamics that contradict the actual architecture of the organ being studied. Trajectories may point across anatomical boundaries where no plausible transition exists, or flows may run in directions that make no sense given the known organization of, say, the cerebral cortex. Now a team reporting in Molecular Systems Biology has built a computational framework designed to fix exactly this problem, and in doing so has added something genuinely new: the ability to simulate how perturbing specific genes would redirect cellular dynamics within intact tissue.

The new method, called veloAgent, was developed by Raghavan, Yoon, Fonseca, Li, and Ding and described in a companion News & Views commentary by Jiahao Wu and Suoqin Jin of Wuhan University. At its core, veloAgent takes spliced and unspliced mRNA counts as input and passes them through a variational autoencoder, a type of deep generative neural network that compresses high-dimensional expression data into a compact latent representation. This compression serves two purposes. It reduces the noise and dimensionality that plague raw count data, and it produces a robust summary of each cell’s transcriptional state that downstream models can work with more reliably than with raw counts alone.

From that latent representation, a neural network constructed in conjunction with gene-gene interaction networks predicts kinetic parameters for every cell and every gene: the transcription rate alpha, the splicing rate beta, and the degradation rate gamma. These are the same mechanistic quantities that underlie classical RNA velocity theory, but here they are estimated with the benefit of regulatory structure encoded in known interactions between genes. The initial velocity computed from these parameters is then refined in two stages. The first stage applies velocity smoothing, aligning the directions of transcriptionally similar cells to enforce local coherence and suppress erratic, single-cell noise. The second stage is where veloAgent departs most sharply from its predecessors.

That second stage is an agent-based model, a simulation technique borrowed from fields like epidemiology and economics in which individual entities, or agents, interact according to local rules and collective behavior emerges from those interactions. In veloAgent, each cell is an agent embedded in its spatial microenvironment. The velocity assigned to each cell is adjusted through neighborhood-based weighting that incorporates spatial distance, local cell density, and expression similarity. In practical terms, a cell’s inferred trajectory is constrained by what its actual physical neighbors are doing. Rather than treating spatial information as an auxiliary feature bolted onto the analysis, veloAgent uses it as a primary organizing principle, allowing cell-state transitions to emerge from local interactions rather than being imposed from above. The refined velocity field then supports downstream analyses including spatio-temporal dynamics inference, projection of velocity fields onto low-dimensional spaces and spatial tissue maps, and in silico gene perturbation.

The perturbation capability deserves particular attention because it moves the framework from description toward prediction. Instead of simply altering observed expression values, veloAgent modulates the underlying kinetic parameters directly. Increasing or decreasing alpha for a chosen gene simulates transcriptional activation or repression at the level of the dynamical system itself, after which the induced velocity field is recomputed. The functional impact of each simulated perturbation is quantified by measuring how the directionality of cellular velocity vectors changes, particularly with respect to predefined target states or trajectories. Genes are then ranked by their ability to reorient cellular dynamics toward those targets, yielding a prioritized list of candidate regulators of cell-state transitions. Because the perturbations operate within the learned parameter space, large-scale screening can be performed efficiently without retraining the model, while retaining mechanistic interpretability that purely correlative approaches lack.

The authors validated veloAgent on spatial transcriptomics datasets from multiple platforms and across diverse tissue types. In structured tissues such as the brain, the framework produced trajectories that closely aligned with known anatomical organization, avoiding the inconsistent or anatomically implausible directions that earlier methods generated. In silico perturbation analysis identified key regulatory genes whose simulated modulation reversed progression toward terminal fiber-tract cell states, demonstrating that the model can pinpoint levers that genuinely redirect cellular dynamics. In a breast cancer dataset, veloAgent’s inferred transitions were more consistent with expected tumor progression patterns, correcting misdirected flows produced by competing methods, and further perturbation analysis revealed potential therapeutic targets capable of suppressing malignant progression.

The quantitative benchmarks reinforce the qualitative picture. veloAgent outperformed alternative methods in cross-boundary directionality, fate probability, and velocity confidence, the metrics that capture whether inferred trajectories respect tissue boundaries and how reliably they predict cell fates. Notably, the framework also improved temporal velocity estimation for standard single-cell RNA sequencing data even when no spatial information was available, suggesting that the agent-based refinement and gene-gene regulatory constraints confer benefits that extend beyond the spatial setting. For a field in which velocity estimates have often been fragile and method-dependent, that robustness is a meaningful advance.

The broader significance, as Wu and Jin emphasize in their commentary, is that cellular behavior cannot be understood outside its native environment. Transcriptional state, spatial structure, physical constraint, and local interaction all conspire to determine what a cell becomes. By modeling cells as interacting agents within tissue rather than as independent observations, veloAgent offers a more realistic representation of how cellular processes actually unfold, and it provides a mechanistically grounded platform for simulating interventions before any experiment is run in the laboratory.

Challenges remain, and the commentary is candid about them. Spatial transcriptomics data are frequently collected across multiple tissue sections and experimental conditions, creating batch-correction problems and complicating robust cross-sample comparison of cellular dynamics. Integrating additional modalities such as epigenomics and proteomics would give a fuller picture of the regulatory processes driving cell dynamics. Incorporating causal machine learning frameworks could sharpen both interpretability and predictive power, and directly modeling intercellular signaling, rather than approximating neighborhood effects under spatial constraints, would deepen the mechanistic account of how cells influence one another. Still, the trajectory of the field is clear. The long-term ambition, articulated in recent proposals for building virtual cells with artificial intelligence, is to move beyond descriptive analysis toward predictive models capable of simulating cellular behavior in silico, accelerating therapeutic target discovery and advancing precision medicine. veloAgent represents a concrete step along that path: a tool that not only decodes where cells in a tissue are going, but suggests how researchers might steer them somewhere else.

Subject of Research: Spatially resolved RNA velocity modeling and in silico perturbation of cell-state transitions in tissue

Article Title: Decoding and steering spatially resolved cellular dynamics

Article References: Decoding and steering spatially resolved cellular dynamics. (n.d.). https://doi.org/10.1038/s44320-026-00216-7

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00216-7

Keywords: RNA velocity, spatial transcriptomics, single-cell RNA sequencing, variational autoencoder, agent-based model, cell-state transitions, gene perturbation, computational biology, tumor progression, deep learning, gene regulation, Molecular Systems Biology