RNA Velocity Wiki

VeloCycle

Properties1
tagsmethod, rna-velocity, manifold-constrained, bayesian, cell-cycle, metric-time, statistical-inference

4 min read

VeloCycle

中文導讀

VeloCycle(Lederer et al., Nat Methods 2024,EPFL,通訊作者 Felix Naef & Gioele La Manno—— 就是 2018 RNA velocity origin paper 的第一作者)是一個 manifold-constrained、Bayesian 的 velocity 方法:它把 manifold learning 跟 velocity field 估計綁進同一個 generative 模型,強制 velocity tangent to manifold,解決 gene-wise 不一致的老問題。聚焦 cell cycle(1D periodic manifold,phase φ),velocity = angular speed ω(φ)。對本 wiki 最關鍵:它把 velocity 用 mean-half-life 為單位,再乘上實測 average half-life,得到以小時計的 cell-cycle period, 而且用 live-imaging + EdU 直接驗證(dHF ≈15 h、RPE1 ≈17.7 h)——是 RNA velocity 第一次被 real-time 實驗直接驗證。所以它是繼 dynamo 之後第二個 reach metric time 的方法,而且走的是 不同的路(periodic manifold + measured half-life,不需要 metabolic labeling)。等於 origin 作者回頭把 physical-time grounding 用 statistical rigor 重新立起來。

What it is

A manifold-constrained, fully Bayesian RNA-velocity framework. It jointly estimates (1) the low-dimensional gene-expression manifold and (2) a velocity field constrained to lie tangent to it, as one generative dynamical system (Pyro / SVI), rather than fitting gene-wise kinetics independently and reconciling afterward. The reference application, VeloCycle, models the cell cycle as a 1D periodic manifold (phase φ ∈ [0,2π], expression as a Fourier series of φ) and infers a phase-dependent angular speed ω(φ). It is the first method to provide statistical velocity inference — credibility tests for nonzero velocity and significance tests for velocity differences between conditions — and the first whose velocity is validated against live-imaging real-time cell-cycle periods.

Physical-time scorecard

AxisVeloCycle
1. Latent time — ordinal / metric?METRIC (on the cell cycle). Phase is ordinal-periodic, but angular speed × measured half-life → period in hours validated by live imaging + EdU. One of only two metric-time methods (with dynamo); the first experimentally validated in real time.
2. Scale degeneracyBroken — by a measured rate, not labeling. Velocity in mean-half-life units; an externally measured average half-life scales it to hours. Degradation/splicing ratio recovered at r≈0.99.
3. External time anchorYes, three: known cell-cycle periodicity (closed 1D manifold) + measured half-lives (rpmh→h) + live-imaging / EdU validation. Notably not metabolic labeling for the fit.
4. Constant-rate assumptionsβ, γ gene-specific constants, but angular speed ω(φ) is phase-dependent — relaxes the constant-speed-along-the-cycle assumption (the “speed modulations”).

Placement: with dynamo, the metric-time pole of the wiki — but reached by a different route (constrained periodic geometry + a measured rate scale + real-time validation, no labeling). It is the constructive resolution of the velocity-skepticism worries: manifold consistency by construction + statistical control on every velocity claim. From velocyto-2018’s author, it closes the origin-reframe loop (physical → ordinal → back to validated physical).

Relation to other methods

  • vs dynamo — both reach metric time; dynamo via metabolic-labeling, VeloCycle via periodicity + measured half-lives + live-imaging (matches dynamo’s LAP period without labels).
  • vs GraphVelo — both make velocity manifold-consistent; GraphVelo by post-hoc tangent projection of any method’s output, VeloCycle by construction inside one generative model.
  • vs veloVI / Cell2fate — shares the Bayesian-uncertainty lineage but adds statistical significance testing (nonzero / differential velocity), not just uncertainty bands.
  • vs TopoVelo — TopoVelo is metric-by-assumption (assumes a 20 h cycle); VeloCycle measures the period and validates it — metric-by-validation, the stronger claim.
  • vs UniTVelo — both exploit the cell cycle’s periodic structure, but VeloCycle yields validated metric time where UniTVelo stays ordinal.

Connections

Graph View

Backlinks