VeloVAE

中文導讀

VeloVAE(Gu, Blaauw & Welch, ICML 2022)是 Welch lab 的 variational-autoencoder velocity: 用 variational mixtures of ODEs 推 cell-specific 的 transcription rate ρ(t) 跟一個 shared latent time,不像 scVelo 假設單一 lineage、也不像 scVelo/veloVI 把每個 gene 放在獨立 timescale。 它是 TopoVelo 的直接前身(TopoVelo 在它上面加 spatial coupling)。physical time 上一樣 ungrounded:latent time ordinal、β γ relative、沒 external anchor。

What it is

A variational-autoencoder RNA-velocity model (Gu, Blaauw & Welch, ICML 2022) using variational mixtures of ODEs to infer cell-specific transcription rates ρ(t) and a single shared latent time across genes, supporting branching trajectories. The direct predecessor of TopoVelo (which adds spatial coupling). Distinct from scVelo/veloVI, which fit each gene on a separate timescale.

Physical-time scorecard

Axis	VeloVAE
Latent time	ordinal (shared across genes)
Rate scale	relative; ρ cell-specific, β,γ relative constants
External anchor	none
Verdict	ungrounded

It estimates cell-specific transcription rates and time well in benchmarks, but does not address absolute time.

Relation to other methods

Predecessor extended by TopoVelo (space) much as veloVI is extended by RegVelo (GRN). Benchmarked against scVelo, veloVI, DeepVelo, cellDancer, STT.

Related

TopoVelo · RNA velocity · splicing-kinetics-ode · latent-time · physical-time-grounding · scVelo · veloVI · RegVelo