veloVI

中文導讀

veloVI(velocity variational inference)把 RNA velocity 從 closed-form ODE 估計搬進 deep-generative / variational 框架:用 encoder–decoder 推 per-gene latent time 跟 posterior velocity distribution,能量化 uncertainty。它仍假設 gene-independent、β γ constant、α 非 regulation-driven。RegVelo 就是直接 generalize veloVI——把 GRN 接進 transcription。對 physical time 一樣 ungrounded:latent time 還是 ordinal、沒 external anchor。

What it is

A deep-generative RNA-velocity model that recasts velocity inference as variational inference: an encoder–decoder produces per-gene latent-time and a posterior velocity distribution, giving principled uncertainty (intrinsic / extrinsic). It is the direct predecessor that RegVelo generalizes by adding a GRN coupling. Now paper-grounded as velovi-2023 (Gayoso, Weiler et al., Nat Methods 2023). Two features beyond uncertainty: a permutation score that tests whether RNA velocity is appropriate for a dataset at all (a method-internal applicability check), and an optional time-dependent α(t) extension relaxing constant transcription.

Model

Per-gene splicing-kinetics-ode with gene-specific constant β, γ and a neural-parameterized α per latent state — but genes are still modeled independently (no regulatory coupling), and transcription is not regulation-driven.

Physical-time scorecard

Axis	veloVI
Latent time	ordinal (per-gene)
Rate scale	relative constants
External anchor	none
Verdict	ungrounded

Relation to other methods

Builds on scVelo’s kinetic model with a variational/generative formulation; generalized by RegVelo (GRN); benchmarked against dynamo. Feeds CellRank for fate analysis.

Related

velovi-2023 · RNA velocity · splicing-kinetics-ode · latent-time · physical-time-grounding · scVelo · RegVelo · spVelo · Cell2fate · VeloCycle · FabianTheis · velocity-skepticism · dynamo · CellRank