RNA Velocity Wiki

Latent time

Properties1
tagsrna-velocity, pseudotime, identifiability, time

4 min read

Latent time

中文導讀

Latent time 是 velocity 模型從 splicing dynamics 反推出來的「內部時間座標」,跟純幾何的 pseudotime 不同——它是從 splicing-kinetics-ode 的動力學算出來的,理論上比 pseudotime 更接近真實時間。但實務上它幾乎都是 ordinal:驗證方式是對 FUCCI / known stage 做 rank correlation(例如 RegVelo 的 Spearman 0.68),這只證明 ordering 對,不證明 equal interval = equal real duration。要變成 metric latent time,得有 external anchor(見 physical-time-grounding)。

Definition

A per-cell (sometimes per-cell-per-gene) temporal coordinate inferred from a velocity model’s fitted dynamics, rather than from manifold geometry alone. Because it is derived from the splicing-kinetics-ode, latent time is meant to be closer to a kinetic clock than a purely geometric pseudotime (which orders cells by transcriptomic similarity to a root and carries no kinetic information).

Ordinal vs metric

This is the central distinction for the wiki (see physical-time-grounding):

  • Ordinal latent time — recovers the correct ordering of states; intervals are not in real units. Validation is a rank correlation (Spearman/Kendall) against a proxy (FUCCI phase, known sampling stage).
  • Metric latent time — equal intervals correspond to equal real durations (hours/days). Validation requires a real-time anchor.

Essentially all snapshot velocity methods produce ordinal latent time, because the splicing ODE’s absolute timescale is non-identifiable from snapshots.

Examples

  • scVelo “latent time” — the origin of the modern concept (scvelo-2020): gene-shared universal “internal clock”, dynamics-grounded (beats pseudotime), but the paper states it is metric only given the overall-timescale prior — otherwise a relative [0,1] ordering.
  • veloVI — per-gene latent time, ordinal; relative to “a given maximum time of the process” (velovi-2023), with the authors naming metabolic-labeling as the missing absolute anchor.
  • RegVelo — cell-gene-specific latent time, validated by Spearman ≈ 0.68 vs FUCCI → ordinal.
  • ddHodge — the potential (gradient component, potential-landscape) serves as the ordering, validated by correlation with developmental stage → ordinal. Notably conditional: faithful only where dynamics are gradient-dominated; cyclic regions (cell cycle) have no faithful potential ordering.
  • TopoVelo — inferred cell time is ordinal, but converted to physical units by assuming a 20-hour gene induction/repression cycle. This is metric-by-assumption: the ordering is real, the absolute spacing is imported as a convention (then cross-checked against live-imaging migration rates via real μm space). See spatial-velocity.
  • VeloCycle — the metric exception: on the cell cycle, the periodic phase φ is ordinal but the angular speed ω(φ), scaled by a measured average half-life, yields a period in hours validated against live imaging + EdU (velocycle-2024). Metric-by-validation, not by assumption — the inverse of TopoVelo’s move, and the first experimentally validated metric latent time.
  • A metric axis needs an external anchor: metabolic-labeling / time-series (dynamo with labels), or a known-period manifold + a measured rate (VeloCycle).

Author-level corroboration. JianhuaXing (dynamo/GraphVelo PI) makes the ordinal case strongly: latent time / pseudotime is a chemical-kinetics reaction coordinate with no intrinsic relation to physical time — a geometric quantity, hence unrelated to rate, so calling it “time” is a concept-swap from the physically defined 2018 velocity. See xing-hu-regvelo-debate and velocity-skepticism.

Beyond ordering: acceleration

ddHodge extracts acceleration (the change-of-change of cell state) from the potential’s curvature / divergence — i.e. whether a cell moves at constant speed or accelerates/decelerates. This is richer temporal structure than an ordering, but it remains in relative units; it does not by itself yield metric time.

NOTE: a high FUCCI rank correlation is necessary but not sufficient for metric time; it certifies ordering only. This is the crux of regvelo-physical-time-critique.

physical-time-grounding · splicing-kinetics-ode · RNA velocity · metabolic-labeling · potential-landscape · spatial-velocity · dynamical-systems-formulation · velocity-skepticism · scVelo · veloVI · velovi-2023 · RegVelo · TopoVelo · VeloCycle · velocycle-2024 · dynamo · ddHodge · JianhuaXing

Graph View

Backlinks