RNA Velocity Wiki

Velocity skepticism

Properties1
tagscritique, physical-time-grounding, identifiability, normalization, discourse

5 min read

Velocity skepticism

中文導讀

這頁把對 RNA velocity 的「圈內質疑」整理成一條 thread,當作 FlowVelo intro 誠實面對批評的 素材。三股力量從不同方向打同一個洞:LiorPachter 罵 velocity-on-UMAP 是 “biological gobbledygook”;GennadyGorin 從 biophysics / normalization / chemical master equation 角度質疑;JianhuaXing 說 latent time 根本是 geometric reaction coordinate、跟 physical time 無關。三者其實都在講本 wiki 的 physical-time-grounding / identifiability 問題。重點: FlowVelo 要把這些當 honest backdrop,不是無視,也不是照單全收 polemic。

What it is

The cluster of internal critiques arguing that mainstream RNA velocity over-claims — that “arrows on a UMAP” and a relabeled clock outrun what snapshot data can identify. Three strands attack the same gap the wiki formalizes as physical-time-grounding:

  1. The visualization / over-claim critique — LiorPachter. “Putting arrows on UMAPs is biological gobbledygook” (2025); repeatedly amplifies papers showing velocity fields are unreliable or artifact-driven. The blunt, public-facing pole.
  2. The biophysics / normalization critique — GennadyGorin (Pachter lab). Reframes velocity through the chemical master equation and the (mis)handling of normalization, arguing much of the variation is technical and that biophysically careful models are needed. Constructive form: Monod (monod-gorin) — fit stochastic CME models to nascent+mature counts and minimize distortive normalization. The worry is upstream of velocity: distorted counts → distorted α/β/γ → distorted time.
  3. The dynamical-systems / “concept-swap” critique — JianhuaXing. Latent time is a geometric reaction coordinate, unrelated to physical time; scVelo silently swapped the physically defined 2018 velocity for a pseudo-velocity (xing-hu-regvelo-debate, dynamical-systems-formulation).
  4. The model-specification critique — Cleary lab (cell-growth-omission-2025, Shah, Ming & Cleary; full text). RNA velocity omits cell growth: in a growing population biomass accumulates, so a positive homeostatic velocity v = λx is required just to hold concentration (up/down-regulation = deviation from that line, not from zero). Quantitatively the fitted degradation absorbs the growth rate: γ* ≈ γ + λ, and — critically — the analysis targets metabolic-labeling (4sU pulse-label/chase) designs, i.e. the wiki’s supposed metric-time anchor (dynamo, metabolic-labeling). The residual artifacts make downstream trajectories point orthogonal to / reversed from the true direction: not just scale-off but direction-wrong. The sharpest strand — a missing term in the equation, and it implies a second anchor (a measured growth rate λ) beside the labeling Δt.

Empirical leg — two independent reliability benchmarks. Separate from the four theoretical strands, the field now has two benchmarks reaching the same verdict. (a) velocity-benchmark-17studies (Luo et al., Cell Rep Methods 2026; the preprint Pachter amplified) — 15 methods × 17 datasets, no method uniformly best, results lack uniformity. (b) velocity-benchmark-ancheta-2026 (Ancheta et al., PLoS Comput Biol 2026, CZ Biohub SF / UCSF — the “UCSF benchmark”) — 5 methods × 3 datasets in depth: cross-method agreement is low and complexity-dependent (consensus falls from 92% of cells on mouse pancreas to 39% on zebrafish NMP), driver-gene rankings are method-sensitive (only Pdx1 in a five-method intersection), “none of the methods perform highly,” so velocity is a hypothesis-generation tool requiring experimental validation. Breadth + depth, same conclusion — velocity output is method-dependent and fragile, without diagnosing why. Theory says why it’s under-identified; the benchmarks show that it’s unstable. Both score only direction (cosine similarity), never physical time — the evaluation culture itself omits the physical-time-grounding axis.

What they share (and don’t)

All four target identifiability / physical meaning, not whether direction is ever useful. They differ on the remedy: Pachter leans toward doubt the enterprise; Gorin toward fix the biophysics + normalization; Xing toward rebuild on dynamical-systems theory + metabolic-labeling physical time (metabolic-labeling); Cleary toward add the missing growth term. Note the authors’ own pragmatism (HuYizhou) — velocity as auxiliary evidence alongside lineage tracing.

Constructive responses (the critique is being answered)

Skepticism is not only destructive — several methods address it head-on, which FlowVelo should cite as the state of the art it builds on:

  • Uncertainty + applicability — veloVI (velovi-2023). Quantifies intrinsic/extrinsic velocity uncertainty and adds a permutation score testing whether RNA velocity is even appropriate for a dataset — a method-internal answer to the reliability worry.
  • Statistical control + validated metric time — VeloCycle (velocycle-2024). Makes velocity manifold-consistent by construction, adds credibility/significance testing on every velocity claim, and — most striking — reaches metric cell-cycle time validated against live imaging, directly rebutting the “pseudo-velocity, never physical” charge on the cell cycle. From velocyto-2018’s own author (GioeleLaManno).
  • Constructive biophysics — Cell2fate / monod-gorin: keep faithful biophysics, quantify where the model is trustworthy.

The honest reading: the skeptics identify a real gap; the best recent methods are closing parts of it (applicability tests, statistical significance, validated metric time on constrained geometry). The general developmental-trajectory metric-time problem remains — FlowVelo’s target.

Why it matters for FlowVelo

The honest intro move is to name this skepticism and locate FlowVelo relative to it: the temporal axis is genuinely under-identified from snapshots (the skeptics are right about the gap), and FlowVelo’s contribution is to be explicit about what it measures vs assumes (flow-matching / transport-operators) rather than to relabel a reaction coordinate as time. Avoid both extremes — neither dismiss the critique nor adopt its polemic tone (see regvelo-physical-time-critique caveats).

physical-time-grounding · latent-time · LiorPachter · GennadyGorin · monod-gorin · velocity-benchmark-17studies · velocity-benchmark-ancheta-2026 · cell-growth-omission-2025 · JianhuaXing · dynamical-systems-formulation · xing-hu-regvelo-debate · velocity-discourse-2025-2026 · metabolic-labeling · FlowVelo

Graph View

Backlinks