Multi-omic single-cell velocity models epigenome–transcriptome interactions (MultiVelo)

MultiVelo (Li et al., Nat Biotechnol 2022)

Li, Virgilio, Collins & Welch. Nature Biotechnology (2022). Welch lab (U Michigan) — the same lineage as VeloVAE and TopoVelo.

Summary

MultiVelo extends the dynamical RNA-velocity model with a third ODE for chromatin accessibility, jointly modeling epigenome→transcriptome timing from single-cell multi-omic (10x Multiome / SHARE-seq) data. Transcription rate is made proportional to chromatin accessibility c(t) (so α is chromatin-driven, not constant), and a probabilistic latent-variable EM infers switch times, rates, a discrete state, and a gene-shared latent time. It resolves priming (chromatin opens before transcription) and decoupling (chromatin closing vs transcriptional repression at different times), improving fate prediction over RNA-only scVelo. For this wiki it is both (a) a context-coupled-α method (epigenome analog of RegVelo’s GRN) and (b) JianhuaXing’s cautionary example of velocity↔expression inconsistency.

Key Claims

• Three-ODE model. dc/dt = α_co(1−c) − α_cc·c; du/dt = α·c − βu; ds/dt = βu − γs. α (max transcription) is gated by chromatin accessibility c — time-varying, not constant.
• Two orderings, four states. Model M1 (chromatin closes before transcription ends) vs M2 (after); each cell is primed / coupled-on / coupled-off / decoupled. ~41% genes M1, ~27% M2.
• Gene-shared latent time + time lags. Infers per-cell latent time; quantifies priming (~21% of total time) and decoupling (~19%) intervals, and TF-expression↔binding-site and SNP-accessibility↔expression lags — all as fractions of normalized time, not physical units.
• Improves fate prediction. Chromatin disambiguates genes densely clustered at the (u,s) origin; corrects scVelo’s “implausible backflows” in cortex upper-layer neurons; recovers inside-out cortical development and TAC→hair-shaft direction.

Physical-time grounding (standing lens)

1. Latent time — ordinal or metric? Ordinal (gene-shared, [0,1]); validated by recovering known direction / cell cycle, like scVelo. Time lags are quantified as fractions of latent time — relative, not hours.
2. Scale degeneracy. Inherited; switch times and latent time on [0,1], no absolute anchor.
3. External anchor. None — snapshot multi-omic (10x Multiome / SHARE-seq); no labeling/time-series.
4. Constant-rate assumptions. α relaxed — chromatin-gated (epigenome context), the analog of RegVelo’s GRN-gated α; β, γ and the chromatin rates α_co, α_cc are gene-specific constants.

Two-faced for the wiki: MultiVelo does relax constant-α via the epigenome (a real advance on the mechanistic axis), yet stays ordinal on time and is the method JianhuaXing and TopoVelo flag for velocity↔expression inconsistency. The “decoupling” it models is intended (chromatin and RNA legitimately move oppositely during priming/repression) — so the critique is better stated as “downstream coarse-grained transitions can be inconsistent,” not “the model is wrong.” Keep that distinction (see MultiVelo).

Connections

• MultiVelo — the method entity.
• VeloVAE / TopoVelo — same Welch-lab lineage; TopoVelo reports MultiVelo inconsistency.
• scVelo — the RNA-only dynamical model it extends with chromatin.
• grn-informed-velocity — chromatin-gated α is the epigenome analog of GRN-gated α.
• latent-time / physical-time-grounding — ordinal time; relative lags.
• GraphVelo — offered (by Xing) as the self-consistent multi-omics alternative.
• velocity-skepticism / xing-hu-regvelo-debate — the inconsistency critique.

Contradictions

• Nuances the second-hand MultiVelo entity: the velocity↔expression “inconsistency” reflects intended decoupling modeling; the fair critique is about downstream transition coarse-graining (TopoVelo), not the ODE model itself.