Dynamo turns single-cell RNA data into predictive vector fields

Single-cell RNA sequencing has moved far beyond static cell atlases, and Dynamo is built for that shift. The protocol presents a machine learning framework that uses noisy RNA velocity estimates to construct analytical and predictive vector fields of gene expression dynamics, converting snapshots of individual cells into a model of how those cells are likely to move through developmental or disease states. In practical terms, this means the data are no longer treated as a collection of isolated transcriptional profiles. They are interpreted as a dynamic system with direction, geometry, and state transition structure.

The protocol sits at the intersection of transcriptomics, systems biology, and computational modeling. Traditional RNA velocity tools provide qualitative directionality, but the framework described here pushes toward quantitative prediction: mapping the geometry of cell state transitions, identifying features of differentiation trajectories, inferring gene regulatory networks, predicting optimal reprogramming paths and transcription factor cocktails, and simulating in silico genetic perturbations. A single workflow therefore serves multiple biological questions, from lineage progression to fate diversion after perturbation. In a field where many analyses stop at descriptive arrows on a UMAP plot, that shift toward governed, model-based prediction is substantial.

Dynamo is not limited to one experimental format. It is designed to work across four data types and five downstream analyses, including conventional scRNA-seq, metabolic-labeling-enabled scRNA-seq, multiomic datasets, and datasets without velocity information. The protocol also spans hematopoiesis, pancreatic endocrinogenesis, and embryonic mouse brain data, showing that the same framework can be adapted across distinct developmental contexts. That flexibility matters because the biological process being modeled is not uniform across tissues; transcriptional kinetics in a developing brain, a blood lineage system, and endocrine differentiation will not unfold on the same timescale or through the same regulatory architecture.

At the core of the workflow is a mathematical reconstruction of the underlying vector field from noisy observed transcriptional states. Once those vector fields are learned, they can be interrogated for directions of movement, stable or unstable trajectories, and the likely consequences of altering key regulators. The protocol therefore treats cell fate as a dynamic system rather than a purely categorical endpoint. For translational researchers, the appeal is obvious: a model that can infer where a cell is headed, and which transcription factors might redirect that path, is more useful than a method that simply annotates where the cell currently sits.

Predictive modeling was not presented as an abstract aspiration. The protocol shows that Dynamo can predict cell fate diversions after genetic perturbations and accurately identify key transcription factors that enable lineage transitions in the hematopoietic system. In hematopoiesis, where lineage choice is tightly regulated and highly consequential, the ability to anticipate perturbation-induced diversion gives the framework real experimental value. A computational prediction that points to the wrong transcription factors wastes time; one that narrows the search space for reprogramming can accelerate both mechanism discovery and experimental design.

Hematopoietic lineage transition analysis anchors the protocol in biology

The hematopoietic case studies provide the most concrete demonstration of what the software is doing biologically. The reconstructed vector field enables differential geometry analyses of cell state progression, and those analyses are then used to examine how progenitor populations move along lineage branches. The protocol also includes in silico perturbation and lineage-associated prediction exercises, linking transcriptional motion to the regulatory levers that steer that motion. In this setting, gene regulatory networks are not merely inferred as static diagrams; they are embedded in a model of temporal change.

This approach is especially relevant in hematopoiesis because lineage commitment is one of the clearest examples of a high-dimensional biological decision process. Cells do not simply switch identity at a single point. They traverse intermediate states, become biased, and then consolidate into a stable fate. A framework that can detect those intermediate steps and estimate the consequences of perturbing the system can reveal how fate decisions are organized in time. The biological logic is straightforward: if transcriptional regulators shape the motion of a cell through state space, then modeling that motion should expose the regulators most likely to alter the destination.

The protocol also demonstrates compatibility with datasets that do not include velocity information, broadening its utility beyond a narrow experimental setup. That is important because many existing datasets were generated before velocity-aware or metabolic-labeling strategies became widespread. By accommodating legacy and modern datasets alike, Dynamo becomes more than a specialized method for a niche experimental design. It becomes a reusable analytical environment for retrospective and prospective single-cell studies.

Practical use, runtime, and access requirements for researchers

The workflow is intended for researchers with basic familiarity in scRNA-seq analysis and Python programming, without requiring advanced coding expertise. Typical RNA velocity vector field analyses can be completed in less than an hour, while the nine case studies in the protocol take about 7 hours and 35 minutes. That runtime is long enough to reflect the complexity of the analyses, but still short enough to remain usable for day-to-day computational biology work. For laboratories balancing exploratory analysis with experimental turnaround, the distinction between a rapid vector field estimate and a full multi-case protocol can shape whether the model is used only for demonstration or folded into a real research pipeline.

Public availability is central to the protocol. Dynamo is released as a Python package, with source code available through GitHub, and the protocol itself is accompanied by code and colab notebooks. Public datasets used in the study can also be downloaded directly through the framework. Those details matter because reproducibility in computational biology depends on more than a methods description; it depends on accessible code, clear workflows, and datasets that others can reanalyze under the same assumptions.

The protocol’s main limitation is also implicit in its design: the method is powerful, but it still depends on the quality of the underlying single-cell data and on the assumptions embedded in RNA velocity estimation and vector-field reconstruction. Predictive outputs can only be as reliable as the experimental input and the biological structure captured by the model. Future validation will need to determine how robust the framework remains across additional tissues, perturbation regimes, and experimental noise profiles beyond the demonstrating examples.


Reporting by Dr. A. Nikhil Krishnan, MBBS for Medicine World Foundation, with additional data provided by British Journal of Cancer.
PRIMARY CLINICAL EVIDENCE & REGISTRY DATA