GRN-informed velocity

中文導讀

把 GRN(gene regulatory network)接進 velocity 模型,讓 transcription rate α 不再是 gene-independent constant,而是由 regulators 決定。代表作是 RegVelo(α=h(|Ws+b|))。 好處是 mechanistic:能做 in silico perturbation、找 lineage driver。光譜的另一極端是 TFvelo——它把 splicing 整個丟掉,直接用 TF expression 的線性組合定義 velocity (dy/dt = W·X − γ·y,y 是 total mRNA),能跑在沒有 splicing 的資料上。但要注意——這條軸 ground 的是 regulation,不是 time;GRN coupling 不會 break rate–time scale degeneracy (見 physical-time-grounding),latent time 依舊 ordinal,TFvelo 甚至離 physical time 更遠。

Definition

A class of RNA velocity methods that couple splicing kinetics to a gene regulatory network, making the transcription rate α a function of regulator abundances rather than a gene-independent constant. This relaxes two classic assumptions of scVelo/veloVI: gene independence and constant transcription.

Why it matters

• Mechanistic interpretability — links which regulator drives which gene to the dynamics, not post hoc.
• In silico perturbation — masking/deleting a regulon and re-solving predicts the causal effect of a TF knockout on velocity and on CellRank fate probabilities.
• Driver discovery — ranks lineage drivers (e.g. RegVelo’s tfec, elf1 in zebrafish neural crest).

The exemplar: RegVelo

RegVelo predicts α_g = h(|W·s + b|_g) with a learnable, prior-constrained GRN weight matrix W, fit jointly with splicing kinetics in one coupled ODE. Predecessor GRN-ignoring methods (scVelo, veloVI) and dynamics-ignoring GRN methods (CellOracle, GRNBoost, Pando, SCENIC+) sit on opposite sides of the gap RegVelo bridges.

Spatial analog: context = neighbors

The same “make α a function of context” idea has a spatial version: TopoVelo predicts transcription ρ from a cell’s spatial neighbors (via a GNN) instead of from a GRN (see spatial-velocity). RegVelo and TopoVelo are siblings — regulatory context vs spatial context — both leaving β, γ as constants and both ordinal on time.

The splicing-free extreme: TFvelo

TFvelo (tfvelo-2024, Li et al. 2024) pushes “make α a function of regulators” to its limit: it discards the splicing kinetic basis entirely and models the target gene’s velocity as W_g·X_g − γ_g·y_g on total mRNA, where X_g are TF expressions. The phase portrait becomes TF-combination (WX) vs target (y) instead of unspliced vs spliced. This buys real robustness — it runs on FISH, no-splicing, and privacy-restricted data, and reproduces MultiVelo’s direction without ATAC — but it moves further from physical time: its “velocity” is a regulation-implied derivative of abundance, not even the ds/dt-in-hours the velocyto origin had. RegVelo = GRN + kinetics; TFvelo = GRN − kinetics. The contrast is a clean illustration that the regulatory and temporal axes are independent (see physical-time-grounding-across-methods).

Standing caveat: regulation ≠ time

GRN coupling grounds the mechanistic axis but not the temporal one. The rate–time scale degeneracy of the splicing-kinetics-ode carries through with the GRN-predicted α (β, γ still relative); latent time stays ordinal. The two axes are orthogonal — see physical-time-grounding and regvelo-physical-time-critique.

Related

RegVelo · TFvelo · tfvelo-2024 · GRN · splicing-kinetics-ode · RNA velocity · veloVI · MultiVelo · CellRank · TopoVelo · spatial-velocity · physical-time-grounding · FlowVelo