RNA Velocity

RNA velocity captures the rate of change in gene expression by distinguishing newly transcribed, unspliced mRNA from mature, spliced mRNA. The ratio of unspliced to spliced reads provides a directional vector indicating whether a gene is being upregulated (induced) or downregulated (repressed) in each cell. This temporal derivative transforms a static snapshot into a predictive model of the immediate future state.

The high-dimensional velocity vectors are projected onto a low-dimensional embedding, such as a UMAP or t-SNE plot, creating a vector field. This field visually represents the flow of cells through transcriptional space. Streamlines traced through this field predict future cell states and reveal the underlying geometry of differentiation, enabling the identification of root cells, branch points, and terminal fates without requiring time-series data.

By integrating the velocity vectors, a latent time is assigned to each cell, ordering them along a continuous developmental trajectory. Unlike pseudotime methods that rely solely on transcriptomic similarity, latent time incorporates directional RNA kinetics. This resolves the directionality ambiguity inherent in static similarity measures, explicitly distinguishing a progenitor cell transitioning toward a differentiated state from a mature cell.

Two primary frameworks exist for estimating velocities:

• Steady-state model: Assumes a constant transcriptional state and fits a linear regression to the unspliced/spliced phase portrait. Computationally efficient but sensitive to violations of the steady-state assumption.
• Dynamical model: Solves the full differential equation of transcription, splicing, and degradation using an Expectation-Maximization (EM) algorithm. This recovers gene-specific kinetic rates and is more robust for transient populations.

Velocity analysis quantifies the probability that a cell is in a terminal differentiation state. Cells with high terminal state confidence exhibit velocity vectors that point inward toward an attractor point, indicating a cessation of transcriptional change. This metric is critical for identifying fully mature cell types and distinguishing them from intermediate, actively transitioning progenitors in complex tissues like the developing brain or tumor microenvironment.

A computational ordering of single cells along a continuous developmental path based on transcriptomic similarity. While RNA velocity provides a directional vector for each cell, pseudotime methods like Monocle and Slingshot assign a scalar position along a lineage. The most robust pipelines combine both: velocity vectors orient the pseudotime axis, resolving the directionality ambiguity inherent in static snapshot data.

The dynamical model of RNA velocity extends the steady-state assumption by solving the full transcriptional kinetics. scVelo recovers latent time, reaction rates for transcription (α), splicing (β), and degradation (γ), and identifies transient cell states. This approach captures non-stationary populations where unspliced/spliced phase portraits deviate from the linear steady-state line, common in rapid developmental transitions.

The computational inference of intercellular signaling networks by analyzing the co-expression of ligands and their cognate receptors across different cell types. Tools like CellChat and NicheNet model how paracrine signals from one population influence the velocity and differentiation trajectory of neighboring cells, linking extrinsic signaling to intrinsic transcriptional dynamics.

An experimental and computational approach that records the heritable history of cell divisions using genetic barcodes or CRISPR-induced scars. When combined with RNA velocity, lineage tracing provides ground-truth clonal relationships that validate or refine computationally inferred trajectories. This multimodal integration bridges the gap between predicted future states (velocity) and observed ancestral relationships (barcodes).

The computational alignment of multiple single-cell datasets from different conditions, technologies, or donors into a shared latent space. Methods like Harmony and scVI correct for batch effects while preserving biological variation. Integrating velocity vectors across batches requires specialized approaches that align not just static expression but also splicing dynamics, ensuring trajectory continuity across integrated datasets.