RegVelo: Gene-regulatory-informed dynamics of single cells

Wang, Hu, Weiler, Mayes, Lange, Fountain, Haug, Wang, Xue, Sauka-Spengler & Theis. Cell 189, 3773–3800 (2026). https://doi.org/10.1016/j.cell.2026.04.022

Summary

RegVelo is a Bayesian deep-generative model that jointly infers transcriptome-wide splicing kinetics and a gene regulatory network (grn-informed-velocity), extending the veloVI variational line. Its central move is to replace the constant-transcription assumption of scVelo/veloVI with a regulation-dependent transcription rate α predicted from each gene’s regulators through a prior GRN, so that splicing kinetics and regulatory wiring are fit in one coupled, high-dimensional ODE system. The payoff is mechanistic: integrated with CellRank, RegVelo supports in silico regulon perturbation and lineage-driver prediction, validated by CRISPR-Cas9 knockout and Perturb-seq in zebrafish neural crest. Crucially for this wiki, RegVelo grounds the regulatory axis while leaving the temporal axis ungrounded — latent time stays ordinal and rates stay relative (see regvelo-physical-time-critique).

Key Claims

• Regulation-dependent transcription. Each gene’s transcription rate is α_g = h(|W·s + b|_g), with W the learnable GRN weight matrix, h a neural net, and s spliced abundance — so α is cell- and gene-specific and regulation-driven, relaxing the constant-α assumption of prior velocity models.
• Single coupled high-dimensional ODE. Unlike scVelo/veloVI’s decoupled one-gene closed-form ODEs, RegVelo solves one genome-wide coupled ODE numerically (no analytic solution; parallelizable solver). β (splicing) and γ (degradation) remain gene-specific constants; only α is made dynamic via the GRN.
• Bayesian uncertainty. Stochastic variational inference yields posterior velocity distributions and intrinsic (state-change) vs extrinsic (future-state) uncertainty, extended to per-edge GRN confidence.
• Actionable perturbation. Masking a regulon and re-solving gives a perturbed velocity field; cosine dissimilarity (“perturbation effect score”) and a CellRank “depletion likelihood” (0–1) rank lineage drivers causally.
• Side-information-guided model selection. A ModelComparison class picks among trained models by agreement with known developmental stages / transitions; the Jacobian-L1-regularized “Soft-r” variant wins (sparsity improves NeuralODE dynamics).
• Benchmarks. Across simulated GRNs (dyngen), cell cycle (FUCCI), pancreatic endocrinogenesis, human hematopoiesis, embryonic myogenesis, hindbrain, and zebrafish neural crest, RegVelo improves CBC, velocity consistency, latent-time correlation, terminal-state identification (TSI), and driver-TF AUROC over scVelo, veloVI, and dynamo — notably out-predicting dynamo’s driver ranking without using dynamo’s metabolic-labeling data.
• Biology. Identifies tfec as an early driver and elf1 as a regulator of pigment-cell fate in zebrafish neural crest; recovers SPI1–GATA1 toggle in hematopoiesis; Neurod2 in pancreatic epsilon-cell genesis.

Physical-time grounding (standing lens)

1. Latent time — ordinal or metric? Ordinal. RegVelo infers a cell-gene-specific latent time (as in veloVI), validated by Spearman ≈ 0.68 against a FUCCI cell-cycle score — a rank statistic. No calibration to hours/days. → ordinal.
2. Rate–time scale degeneracy. Inherited, not broken. GRN coupling constrains the functional form of α(regulators) but the time↔rate rescaling carries through with the GRN-predicted α; β, γ stay known only up to a common scale (argument on physical-time-grounding).
3. External time anchor. None. Snapshot-only; deliberately forgoes metabolic-labeling — and beats label-using dynamo on driver prediction, i.e. it discards the very signal that would fix absolute scale.
4. Constant-rate assumptions. α is freed (regulation-dependent); β and γ remain gene-specific constants. Wherever splicing/degradation rates drift along the trajectory the implied time mapping is biased, and with no clock the bias is unobservable.

Detailed critique and draft-ready manuscript paragraph: regvelo-physical-time-critique.

Key Quotes

“RNA velocity relies on restrictive assumptions, like gene independence and constant transcription rates, thereby excluding transcriptional regulation underlying cell differentiation.” — Introduction, motivating the GRN coupling.

“existing methods for inferring RNA velocity consider a set of decoupled one-dimensional ODEs for which analytic solutions exist, but RegVelo relies on a single, high-dimensional ODE … now coupled through gene regulation-informed transcription α_g = h(|Ws + b|_g).”

“its estimated latent time correlated positively with the FUCCI score used as ground truth (Spearman correlation = 0.683).” — the headline temporal validation.

Connections

• RegVelo — the method this source defines.
• veloVI — direct predecessor; RegVelo generalizes its variational ODE inference.
• scVelo — the constant-rate baseline RegVelo argues against.
• dynamo — label-using competitor RegVelo out-predicts on drivers without labels.
• CellRank — fate-probability / terminal-state layer RegVelo plugs into.
• grn-informed-velocity — the concept RegVelo most directly advances.
• splicing-kinetics-ode — the kinetic backbone, here coupled via the GRN.
• latent-time — RegVelo’s temporal coordinate (ordinal).
• physical-time-grounding — the axis RegVelo does not address.
• metabolic-labeling — the absolute-time signal RegVelo forgoes.
• FlowVelo — our work; RegVelo grounds mechanism, FlowVelo targets the temporal seam.

Contradictions

• No factual contradiction with existing wiki pages. The pre-seeded regvelo-physical-time-critique and physical-time-grounding pages are confirmed by the source: β, γ constant, α GRN-driven, latent time ordinal (FUCCI Spearman 0.68), snapshot-only.