UniTVelo

中文導讀

UniTVelo(Gao, Qiao & Huang, Nat Commun 2022,Yuanhua Huang lab @ HKU)是把 unified latent time 當作主賣點的 RNA velocity 方法:不像 scVelo 每個 gene 各自一條 time,它用 一條全 transcriptome 共用的 latent time 做 temporal regularization,並用 **top-down / RBF(radial basis function)**直接擬合 spliced reads 隨 time 的 profile(spliced-oriented), 讓 transcription activity 變 time-varying。對本 wiki 最關鍵的一點:即使是「以 time 為核心命題」 的方法,它的 unified time 仍然是 ordinal([0,1]、靠 trajectory/rank 驗證),不是 metric—— unified ≠ physical。在 velocity-benchmark-ancheta-2026 裡它 Local Consistency 最高(field 最平滑),但在較難的 zebrafish 資料上跟其他方法的 agreement 偏低(smoothness ≠ correctness)。

What it is

A statistical RNA-velocity framework whose central idea is temporal unification: instead of fitting an independent latent time per gene (the scVelo dynamical model) and stitching a consensus afterward, UniTVelo optimizes one latent time shared across all genes as explicit temporal regularization. Its second distinctive choice is a top-down, spliced-oriented design: the spliced-expression profile s(t) is modeled directly with a radial basis function (RBF) (a smooth bell-shaped induction/repression curve over time) and the unspliced/kinetics are derived from it — the inverse of scVelo’s bottom-up “integrate the ODE from transcription α” route. This gives flexible (time-varying) transcription activity and is designed to fix erroneous velocity directions caused by model violation or lack of temporal regularization (its headline case: recovering the expected hematopoietic differentiation direction where earlier methods reversed it).

UniTVelo offers two modes: a unified-time mode (default — one global clock, best when genes are temporally synchronized along a trajectory) and an independent mode (per-gene time, closer to scVelo, for less-synchronized systems).

Primary paper now full-text ingested as unitvelo-2022 (Gao, Qiao & Huang, Nat Commun 13:6586, 2022 — s41467-022-34188-7); the four-axis scorecard below is paper-grounded. Also benchmarked in velocity-benchmark-ancheta-2026.

Physical-time scorecard

Axis	UniTVelo
1. Latent time — ordinal / metric?	Ordinal. A unified latent time on [0,1], validated by trajectory/direction recovery and rank correlation — its whole thesis is a consistent global ordering, not physical units. The most time-centric method that still stops at ordinal.
2. Scale degeneracy	Inherited. The RBF fixes the shape of s(t) but the time axis is normalized; no absolute timescale, no anchor to break the rate↔time rescaling.
3. External time anchor	None. Snapshot scRNA-seq only; no metabolic-labeling or real-time series.
4. Constant-rate assumptions	α relaxed — transcription activity is time-varying via the top-down RBF (a flexible α(t)); β, γ gene-specific constants. The flexibility is in the spliced profile, not in splicing/degradation rates.

The instructive point for physical-time-grounding: UniTVelo makes the time variable the explicit object of inference (a unified clock) and still lands ordinal. “Unified” buys cross-gene consistency, not physical calibration — so it strengthens, rather than weakens, the claim that metric time needs its own external signal (the seam FlowVelo targets).

Benchmark behavior

In velocity-benchmark-ancheta-2026 (Ancheta et al., PLoS Comput Biol 2026) UniTVelo posts the highest Local Consistency (≈0.8–1.0 across datasets) — unsurprising given its unified-time temporal regularization smooths the field — yet shows low cross-method agreement on the harder zebrafish data (A2 ≈ 0.305 on ZF NMP, 0.392 on ZF embryo vs 0.848 on pancreas) and falls into the {scv-Dyn, UniTVelo} driver-gene camp (only ~55% overlap with the other cluster). Read together: high self-consistency is not the same as cross-method correctness — a smooth, internally coherent field can still disagree with every other method on a complex system.

Connections

• unitvelo-2022 — the source page (full-text); RBF formula, τ_g shape classification, two modes, CBDir table.
• scVelo — the bottom-up dynamical baseline UniTVelo inverts (top-down, unified time vs per-gene time).
• TFvelo — borrows the same top-down profile-function idea but with a sine on TF-driven abundance (no splicing).
• latent-time — UniTVelo is the strongest expression of the unified latent-time idea; still ordinal.
• physical-time-grounding / physical-time-grounding-across-methods — the four-axis lens; UniTVelo as “time-centric yet ordinal.”
• velocity-benchmark-ancheta-2026 — benchmarked here (highest LC; low agreement on hard data).
• Cell2fate — a different take on the time variable (cell-specific T_c + uncertainty) vs UniTVelo’s single unified clock.
• metabolic-labeling / dynamo — the anchor route UniTVelo, like all snapshot methods, forgoes.
• FlowVelo — contrast: a unified ordinal clock vs targeting metric time.