GRN (gene regulatory network)

中文導讀

GRN 是 TF→target 的調控圖。傳統 GRN inference(SCENIC+、CellOracle、GRNBoost、Pando)只給 network,不模 dynamics;傳統 velocity 又把 gene 當 independent。RegVelo 的切點就是把 GRN 當成 weight matrix W 接進 transcription rate(α=h(|Ws+b|)),讓 network 跟 dynamics 一起 fit。本 wiki 關心的是:GRN 補的是 mechanistic 軸,跟 physical-time-grounding 的 temporal 軸正交。

Definition

A directed, signed network in which transcription factors regulate target genes (activation / repression). In velocity modeling, a GRN is encoded as a weight matrix W whose entries scale each regulator’s effect on a target’s transcription rate.

Two cultures it bridges

• GRN inference without dynamics — SCENIC+, CellOracle, GRNBoost, Pando, Dictys, UniBind/ChIP-Atlas priors. Produce networks but model cell dynamics only post hoc.
• Velocity without regulation — scVelo, veloVI. Model dynamics but treat genes as independent with constant transcription.

grn-informed-velocity (i.e. RegVelo) couples the two: a prior GRN seeds W, which is then refined data-drivenly under sparsity/Jacobian penalties while splicing kinetics are fit jointly.

Caveat for this wiki

A GRN grounds the mechanistic axis (which regulator drives which gene; in silico perturbation), which is orthogonal to physical-time grounding. Coupling α to a GRN does not break the rate–time scale degeneracy of the splicing-kinetics-ode.

Related

grn-informed-velocity · RegVelo · RNA velocity · physical-time-grounding