UniTVelo: temporally unified RNA velocity reinforces single-cell trajectory inference

UniTVelo (Gao, Qiao & Huang, Nat Commun 2022)

Mingze Gao, Chen Qiao & Yuanhua Huang. Nature Communications 13:6586 (2022-11-03). HKU (School of Biomedical Sciences; Dept of Statistics & Actuarial Science). Full-text ingest grounding the UniTVelo entity.

Summary

UniTVelo reformulates RNA velocity with two innovations: (1) a top-down, spliced-RNA-oriented design — instead of scVelo’s bottom-up route (define transcription α(t) as a step function, then integrate the splicing-kinetics-ode to get s(t)), UniTVelo directly models the spliced profile s(t) = f(t;θ) with a radial basis function (RBF) and derives u(t) and α(t) from it; and (2) a unified latent time shared across the whole transcriptome, aggregating dynamic information across genes as temporal regularization. Together these fix the erroneous/reversed directions earlier methods produce on systems with weak kinetics, multiple-rate-kinetics (MURK) genes, cell cycle, or complex branches. The headline cases: it recovers the expected erythroid maturation and hematopoietic directions where scVelo reverses them. For the wiki it is the strongest expression of the unified latent time idea — and the proof that “unified” buys cross-gene consistency, not physical units (see physical-time-grounding).

Key Claims

• Top-down RBF model. du/dt = α(t) − βu(t); ds/dt = βu(t) − γs(t), but s(t) is fit directly as s_g(t) = h_g·e^{−a_g(t−τ_g)²} + o_g (RBF, params h,a,τ,o); velocity = ds_g/dt = s_g(t)·(−2a_g(t−τ_g)) — taken from the derivative of the fitted spliced curve, not the deviation from a steady-state line. More robust because it leans on the more reliable spliced counts and admits cells either above or below steady state.
• Peak time τ_g classifies gene shape. τ_g (the gene’s time of highest expression in the normalized [0,1] window) labels each gene: τ<0 repression, 0<τ<1 transient, τ>1 induction — recovered shapes match expectation (most mouse-dataset genes τ<0 = repression).
• Unified latent time. A single gene-shared time t_n = (1/G)Σ_g Q[t_ng] (quantile-aggregated across genes) regularizes the ordering — letting stably/monotonically changing genes contribute rather than over-fitting each gene’s almond-shaped portrait independently (the scVelo failure).
• Two modes. Unified-time mode (default; for cell-cycle / stably-changing genes) and an independent mode (per-gene time, scVelo-like; for high-SNR / complex / sparse datasets).
• Resolves MURK genes. On Abcg2, Smim1 (transcriptional boosting) UniTVelo rescues directions scVelo’s gene-independent mode reverses, by jointly fitting all genes under one time.
• Broadly validated. CBDir (cross-boundary direction correctness) best on most of 10 datasets (Table 1): mouse erythroid 0.793, human bone marrow 0.804, dentate gyrus 0.746, scNT-seq 0.483, intestinal organoid 0.594, hindbrain 0.609 — vs scVelo stochastic/dynamical. R² per-gene goodness-of-fit flags well- vs poorly-explained genes.

Physical-time grounding (standing lens)

1. Latent time — ordinal or metric? Ordinal, but the most time-centric of the ordinal set: the unified latent time is its headline. Rescaled/truncated to [0,1], validated by CBDir and concordance with diffusion-pseudotime — a consistent ordering, never physical units.
2. Scale degeneracy. Inherited. The RBF fixes the shape of s(t); τ_g is a peak time within the normalized [0,1] window; no absolute anchor, so the rate↔time rescaling stands.
3. External time anchor. None. Snapshot scRNA-seq — and notably, even when applied to scNT-seq / scEU-seq (metabolic-labeling data), UniTVelo uses only the unspliced/spliced counts and does not ingest the labeling to anchor time (it deliberately mimics standard scRNA-seq there).
4. Constant-rate assumptions. α(t) relaxed — “flexible transcription activities,” time-varying and derived from the top-down RBF (the inverse of assuming a constant/step α). β, γ remain gene-specific constants.

The instructive point: UniTVelo makes the time variable the explicit object of inference (a unified transcriptome-wide clock) and still lands ordinal. Unified ≠ metric — it buys cross-gene consistency, not physical calibration. That the strongest latent-time method stops at ordinal is itself the argument that metric time needs a separate external signal — the seam FlowVelo targets.

Key Quotes

“we present UniTVelo, a statistical framework of RNA velocity that models the dynamics of spliced and unspliced RNAs via flexible transcription activities. Uniquely, it also supports the inference of a unified latent time across the transcriptome.” — Abstract.

Connections

• UniTVelo — the method entity (now paper-grounded).
• scVelo — the bottom-up dynamical baseline UniTVelo inverts (top-down RBF; unified vs per-gene time).
• splicing-kinetics-ode — the kinetics; UniTVelo solves it spliced-first via RBF.
• latent-time — UniTVelo is the strongest unified latent-time expression; still ordinal.
• manifold-consistent-velocity / GraphVelo / cellDancer — the MURK-gene problem it also targets.
• MultiVelo — UniTVelo notes its top-down design aligns with MultiVelo’s chromatin-predicted α.
• physical-time-grounding / physical-time-grounding-across-methods — “time-centric yet ordinal.”
• FlowVelo — contrast: a unified ordinal clock vs targeting metric time.

Contradictions

• None. Confirms the UniTVelo entity built earlier from web sources; adds the RBF formula, τ_g shape classification, the two modes, and the Table 1 CBDir numbers.