Statistical inference with a manifold-constrained RNA velocity model uncovers cell cycle speed modulations

VeloCycle (Lederer et al., Nat Methods 2024)

Lederer, Leonardi, Talamanca, Bobrovskiy, Herrera, Droin, Khven, Carvalho, Valente, Dominguez Mantes, Mulet Arabí, Pinello, Felix Naef & Gioele La Manno. Nature Methods 21:2271–2286 (2024-10-31). EPFL (La Manno = first author of the 2018 origin paper velocyto-2018; Naef = circadian/systems-biology) + Broad/MGH (Pinello). Companion News & Views: velocycle-nv-2024.

Summary

VeloCycle reframes RNA velocity as statistical inference on a manifold-constrained generative dynamical system: instead of fitting gene-wise kinetics independently and reconciling post hoc, it jointly learns the low-dimensional expression manifold and a velocity field constrained to lie tangent to that manifold, in one Bayesian model (Pyro / SVI). Applied to the cell cycle — a 1D periodic manifold parameterized by phase φ ∈ [0,2π] with gene expression as a Fourier series — it infers an angular speed ω(φ) for cell-cycle progression. The landmark result for this wiki: by expressing velocity in mean-half-life units (rpmh) and scaling by measured average mRNA half-lives, VeloCycle reports a cell-cycle period in hours that matches live-imaging time-lapse microscopy and EdU labeling (dHF ≈ 15.3–15.8 h; RPE1 ≈ 17.7 h) — “the first example of a direct validation of RNA velocity estimation with experimental methodologies … justifies the use of VeloCycle output in units of real (not pseudo) time.” It also delivers statistical velocity inference for the first time: credibility testing for nonzero velocity, and differential-velocity significance between conditions (erlotinib drug response, genome-wide Perturb-seq). From velocyto-2018’s author, this is the origin paper’s physical-time grounding, re-established with statistical rigor — a second route to metric time distinct from dynamo’s metabolic labeling.

Figures

Fig 1 — manifold-constrained velocity framework (Lederer et al., Nat Methods 2024; velocycle-2024). (a) Joint estimation of the gene-expression manifold x and a velocity field V(x) constrained tangent to the manifold — vs (b) standard gene-wise velocity, whose components live on different timescales and predict off the manifold. (c) Plate diagram: spliced S and unspliced U are negative-binomial around manifold + kinetics. (e) The VeloCycle cell-cycle model: expression as a Fourier series of phase φ; velocity = angular speed ω(φ) around the ring. (f) The four new capabilities: (i) statistical velocity significance testing, (ii) structured-uncertainty posterior, (iii) direct validation against live imaging, (iv) transfer learning to small datasets. The governing identity: ∇ₓs·V(x) = β·u(x) − γ·s(x).

Fig 5 — validation in real (not pseudo) time (Lederer et al., Nat Methods 2024). VeloCycle’s velocity, scaled by measured average half-lives, yields a cell-cycle period in hours that overlaps the period measured by time-lapse microscopy (dHF 15.0 h; RPE1 16 h 40 min) and by continuous EdU labeling (RPE1 16.8 h). Panel (q) compares cell-cycle periods from VeloCycle (standard scRNA-seq), Dynamo (LAP), and Dynamo-Metabolic — VeloCycle’s posterior overlaps the gold standard without metabolic labels. This is the figure that earns “units of real time.”

Key Claims

• Manifold-constrained, dynamically consistent velocity. Velocity is parameterized as an autonomous field V(x) on manifold coordinates, forced tangent to the manifold — eliminating the gene-wise inconsistency where velocity components sit on different timescales and point off the traversed manifold. Resolves a critique GraphVelo also targets, but here by construction in a single generative model rather than as post-hoc projection.
• Cell cycle as a 1D periodic manifold. Phase φ ∈ [0,2π]; gene expression = Fourier series of φ; velocity = angular speed ω(φ) (radians per mean half-life, rpmh), allowed to vary with phase (speed modulation). Manifold learning recovers phases at circular correlation r≈0.95 vs ground truth; outperforms DeepCycle (autoencoder; 60% lower MSE).
• Metric time, validated. Scaling ω by measured average half-lives → cell-cycle period in hours, matching live-imaging and EdU ground truth across cell types with distinct periods (dHF ≈15.3–15.8 h, RPE1 ≈17.7 h). G2/M progresses faster (0.47 rpmh) than G1 (0.37) / S (0.36).
• Statistical velocity inference (first of its kind). Full Bayesian posterior (SVI; MCMC/NUTS for the correlated structure) supports (1) credibility testing for nonzero velocity, (2) differential-velocity significance between samples — demonstrated on erlotinib (slower mitotic cell cycle at D3, localized to G2/M) and genome-wide Perturb-seq knockdowns in RPE1.
• Structured uncertainty + transfer learning. A low-rank multivariate-normal posterior captures the genuine correlation between degradation γ and angular speed ω (mean-field SVI underestimates velocity uncertainty); manifold parameters transfer from large references to small/sparse target datasets.
• Spatiotemporal biology. Radial-glia cell-cycle speed varies along the developing anterior–posterior axis (FB < MB < HB); regionally defined progenitor speed differences.

Physical-time grounding (standing lens) — a rare metric-time case

1. Latent time — ordinal or metric? METRIC (on the cell cycle). The phase φ is ordinal-periodic, but the inferred angular speed, scaled by measured half-lives, gives a period in hours validated against live imaging + EdU — genuine metric time, not pseudo-time. With dynamo, one of only two methods in the wiki reaching metric time, and the first with direct real-time experimental validation.
2. Rate–time scale degeneracy. Broken — by a measured-rate anchor, not labeling. Velocity is reported in mean-half-life units; converting to hours requires an externally measured average mRNA half-life (≈1 h assumed/measured). So the absolute timescale is fixed by a measured degradation rate, a different anchor than metabolic-labeling’s known Δt. The degradation–splicing rate ratio is recovered near-perfectly (r≈0.99 in simulation).
3. External time anchor. Yes — three, and none is metabolic labeling for the estimate itself: (i) the cell cycle’s known periodicity (a closed 1D manifold), (ii) measured average half-lives to scale rpmh→hours, (iii) live-imaging period + EdU as validation ground truth. (scEU-seq labeling is used only for cross-comparison, not to fit the velocity.)
4. Constant-rate assumptions. β, γ gene-specific constants (jointly inferred), but the angular speed ω(φ) is phase-dependent — it relaxes the constant-speed assumption along the cycle (the “speed modulations” of the title), the axis most relevant to physical-time mapping.

The decisive entry for the wiki’s thesis. The “temporal axis is open / only dynamo reaches metric” framing must now be qualified: on a process with known periodicity and measured rate scale (the cell cycle), metric time is reached and experimentally validated — by velocyto-2018’s own author, closing the loop on the origin reframe (physical→ordinal→back to physical). The general (non-periodic, developmental) trajectory remains open. For FlowVelo: metric time is attainable when (a) the manifold geometry is constrained/known and (b) the rate scale is anchored by a measured rate (half-life) or labeling — VeloCycle is the proof of principle to extend beyond the cell cycle.

Key Quotes

“this is the first example of a direct validation of RNA velocity estimation with experimental methodologies and justifies the use of VeloCycle output in units of real (not pseudo) time.” — Results.

“we introduce a Bayesian model of RNA velocity that couples velocity field and manifold estimation in a reformulated, unified framework, identifying the parameters of an explicit dynamical system.” — Abstract.

Connections

• VeloCycle — the method entity.
• velocyto-2018 / GioeleLaManno — the 2018 origin and its author; VeloCycle re-grounds physical time.
• physical-time-grounding — a second route to metric time (periodic manifold + measured half-lives + live-imaging), beside dynamo’s labeling.
• metabolic-labeling — the alternative anchor it does not need (validates against EdU/scEU-seq instead).
• manifold-consistent-velocity / GraphVelo — manifold-tangent velocity, here by construction in one generative model.
• latent-time — periodic phase that becomes metric via angular speed + half-life scaling.
• veloVI / Cell2fate — the Bayesian-uncertainty lineage; VeloCycle adds significance testing.
• dynamo — the other metric-time method; VeloCycle matches its LAP period estimate without labels.
• velocity-skepticism — a constructive answer (statistical control + manifold consistency) to the “heuristics / not manifold-consistent” critiques.
• FelixNaef — co-corresponding author (cell-cycle / circadian systems biology).
• velocycle-nv-2024 — the News & Views commentary.
• FlowVelo — proof that metric time is reachable with constrained geometry + a rate anchor; extend beyond the cell cycle.

Contradictions

• Updates the wiki’s central “temporal axis is open” claim. Prior pages framed metric time as reachable only via metabolic-labeling (dynamo), with everything else ordinal. VeloCycle reaches validated metric time on the cell cycle without labeling. The corrected framing: the temporal axis is solved on periodic/known-geometry processes with a measured rate scale, still open for general developmental trajectories. Propagated to physical-time-grounding, overview, and physical-time-grounding-across-methods.