TFvelo

中文導讀

TFvelo(Li, Pan, Yuan & Shen, Nat Commun 2024,Shanghai Jiao Tong University)是一個完全 不用 splicing 的 RNA velocity 方法:它把 target gene 的 velocity 直接寫成TF expression 的 線性組合(dy/dt = W_g·X_g − γ_g·y_g,y 是 total mRNA),用 GRN(ENCODE + ChEA TF-target)取代 unspliced/spliced 的 phase delay。跟 UniTVelo 一樣是 top-down(直接設計 y_g(t) 的 sine profile,generalized EM 交替更新 latent time / TF weights / shape)。最大賣點是 robustness——能跑 在 splicing 不可得或太 sparse 的資料(FISH、privacy-restricted human embryo、只有 total mRNA), 甚至在不用 ATAC 的情況下重現 MultiVelo 的方向。對本 wiki 的定位:它是 regulatory axis 的 極端——保留 RegVelo 的 GRN-driven α,卻把 splicing kinetic basis 整個丟掉,因此離 physical-time grounding 更遠(velocity 變成 total abundance 的 regression-derived rate,連 velocyto 當年那個 ds/dt-in-hours 的物理意義都沒了)。

What it is

A splicing-free, GRN-driven RNA-velocity method. The motivating finding (via LASSO) is that a target gene’s velocity is well-approximated by a linear combination of its transcription factors’ expression, with the selected TFs significantly enriched in TF–target databases. TFvelo therefore models the time-derivative of a target’s total mRNA abundance as

dy_g(t)/dt = W_g·X_g(t) − γ_g·y_g(t)

where X_g are the gene’s TF expressions, W_g learned regulatory weights, γ_g degradation. The phase portrait is recast as WX (weighted TF combination) vs y (target) — a clockwise co-expression loop replacing the unspliced/spliced loop. It is top-down (designs the abundance profile y_g(t) = α_g·sin(ω_g t + θ_g) + β_g directly, ω=2π) and fits with a generalized EM over three parameter groups: cell-specific latent time t, TF weights W, and shape params [α,β,θ,γ].

Its practical edge is applicability: needing only total mRNA + a static GRN, it runs where splicing-based methods cannot — FISH data, ATAC-paired multiome (matching MultiVelo without using ATAC), and privacy-restricted human preimplantation-embryo data whose raw reads are unavailable.

Physical-time scorecard

Axis	TFvelo
1. Latent time — ordinal / metric?	Ordinal. Cell-specific latent time on [0,1] (sine window) → velocity-pseudotime + transition matrix (scVelo-style); validated by CBDir / expression-trend. No physical units.
2. Scale degeneracy	Inherited, arguably deepened. Sine profile in normalized [0,1] time; W, γ relative. With no splicing kinetics, even the ds/dt-in-hours reading of the velocyto origin is gone.
3. External time anchor	None — the furthest from one. Drops the splicing signal too; works from total mRNA + a static GRN prior. No labeling, no unspliced.
4. Constant-rate assumptions	Production replaced by W_g·X_g(t) — GRN-driven, time-varying “transcription” (TF expression used directly, cf. RegVelo’s learned GRN function); γ_g gene-specific constant.

Placement: the splicing-free extreme of the regulatory axis. TFvelo confirms the wiki’s orthogonality thesis from the limit — you can advance regulation so far that splicing itself is discarded, and the temporal axis is still untouched (ordinal, no anchor). It buys robustness at the cost of moving “velocity” from a physical rate to a regulation-implied direction on abundance.

Relation to other methods

• vs RegVelo — both make transcription regulation-driven, but RegVelo couples a GRN into the splicing ODE (keeps unspliced/spliced + latent time); TFvelo replaces splicing with TF regression on total mRNA. RegVelo = GRN + kinetics; TFvelo = GRN − kinetics.
• vs UniTVelo — shares the top-down profile-function design (TFvelo sine vs UniTVelo RBF) and explicitly cites it; differs in the observable (TF→target abundance vs spliced reads).
• vs MultiVelo — both gate transcription on context; MultiVelo uses chromatin (ATAC), TFvelo uses TF expression, and TFvelo reproduces MultiVelo’s direction without ATAC.
• vs monod-gorin — opposite responses to “splicing counts are noisy”: Monod models the count stochasticity rigorously; TFvelo sidesteps splicing counts via regulation.

Connections

• tfvelo-2024 — the source page (full-text); model, generalized EM, validation datasets.
• grn-informed-velocity — TFvelo is the splicing-free extreme of this concept.
• RegVelo / GRN — the GRN-coupled relatives; TFvelo drops the kinetic basis they keep.
• UniTVelo — shared top-down profile design (sine vs RBF).
• scVelo / MultiVelo / velocyto — the splicing baselines it replaces / matches without ATAC.
• RNA velocity / splicing-kinetics-ode — the splicing-kinetics definition TFvelo diverges from.
• velocity-skepticism — constructive answer to “splicing signal is weak” (drop it, use regulation).
• physical-time-grounding / physical-time-grounding-across-methods — regulatory axis advances, temporal untouched.
• FlowVelo — contrast on what “velocity” denotes once splicing is gone.