RNA Velocity Wiki

RNA velocity

Properties1
tagsrna-velocity, single-cell, trajectory, kinetics

3 min read

RNA velocity

中文導讀

RNA velocity 是用 unspliced/spliced mRNA 比例,從單一 snapshot 推斷每個細胞的「即時變化方向」 的框架,底層是 splicing-kinetics-ode。它最大的貢獻是不需要 predefined root 就能給出 local direction;最大的限制是 snapshot 只定得出 direction、定不了 absolute timescale (見 physical-time-grounding)。本 wiki 把各方法沿四個軸評:latent time 是 ordinal 還是 metric、scale degeneracy 有沒有被 break、有沒有 external time anchor、哪些 rate 被當 constant。

Definition

A framework for inferring the instantaneous direction of cell-state change from a single scRNA-seq snapshot, by exploiting the ratio of unspliced (nascent) to spliced (mature) mRNA. The velocity vector v = ds/dt = β·u − γ·s (from the splicing-kinetics-ode) points toward each cell’s near-future state, enabling trajectory direction without a predefined root — the limitation that pseudotime (see latent-time) requires.

Origin (La Manno et al., 2018)

The founding paper (velocyto-2018; La Manno, Linnarsson, Kharchenko — preprocessing reworked with HuYizhou; tool velocyto) is, by both critics’ and authors’ account, the physically sound version — it even calibrated the extrapolation step to hours via metabolic labeling. From intron-bearing (u) and intron-free (s) reads and ds/dt = βu − γs, it estimates γ/β from cells at steady state (ds/dt = 0), then infers ds/dt for non-steady cells. Two points JianhuaXing stresses (xing-hu-regvelo-debate): (1) du/dt = α − βu was not used in the ds/dt estimate; (2) α is cell-state-dependent, not a constant — it encodes quantitative gene regulation. The later silent move to scVelo’s constant-α “pseudo-velocity” and a relabeled latent-time is the contested step (see velocity-skepticism).

The method family (this wiki’s scope)

  • Steady-state / dynamical — scVelo (the field’s workhorse baseline).
  • Deep-generative — veloVI (variational), extended to GRNs by RegVelo.
  • Metabolic-labeling-aware — dynamo (can recover absolute rates with labels).
  • Manifold/graph-consistent — refinements that constrain velocity to the data manifold or a cell graph.
  • GRN-informed — grn-informed-velocity / RegVelo; at the extreme, TFvelo drops splicing entirely (velocity = TF regression on total mRNA).
  • Unified-time / top-down — UniTVelo (RBF-fit spliced profile + one transcriptome-wide clock).
  • Our line — FlowVelo, JuloVelo.

Downstream, CellRank turns a velocity field into transition probabilities, terminal states, and fate probabilities.

”Velocity” is becoming an umbrella observable

The newest methods stretch what “RNA velocity” denotes, which matters when comparing them. At one pole the velocyto origin’s velocity is a physical ds/dt (unspliced→spliced, calibrated to hours); at the other, TFvelo (tfvelo-2024) computes a GRN-driven derivative of total mRNA with no splicing at all, and MultiVelo gates transcription on chromatin. They are not estimating the same quantity — so directional agreement and physical-time grounding are separate questions for each. Flag this when reading cross-method benchmarks (velocity-benchmark-ancheta-2026) and when positioning FlowVelo.

The standing critique: physical time

Snapshot RNA velocity recovers the shape and direction of a trajectory but not its duration — the splicing-kinetics-ode’s absolute timescale is non-identifiable. Every method in this wiki is evaluated on how it handles physical time; see physical-time-grounding.

splicing-kinetics-ode · latent-time · physical-time-grounding · grn-informed-velocity · metabolic-labeling · scVelo · veloVI · RegVelo · TFvelo · UniTVelo · MultiVelo · dynamo · CellRank · FlowVelo · JuloVelo

Graph View

Backlinks