Overview
A single-topic knowledge base on RNA velocity and single-cell trajectory inference, maintained to support critical method evaluation and the drafting of the FlowVelo / JuloVelo line of work. The standing analytical lens is physical-time-grounding: for every method, is inferred time ordinal (a mere ordering) or metric (calibrated to real units), and does the method break, ignore, or inherit the splicing-kinetics-ode’s rate–time scale degeneracy?
Current synthesis
The field is advancing along four largely orthogonal axes, none of which is physical-time grounding per se:
-
The mechanistic / regulatory axis — coupling splicing kinetics to regulation. RegVelo is the current frontier: it makes transcription α regulation-dependent via a prior GRN (grn-informed-velocity), enabling in silico perturbation and strong lineage-driver recovery. This axis is maturing fast.
-
The geometric / manifold axis — making the velocity field consistent with the data manifold and dissecting its structure. GraphVelo projects any method’s velocity onto the tangent space (manifold-consistent-velocity), uniquely preserving magnitude (cell speed) and relaxing the constant-rate assumption (cell-context-specific α, γ for MURK genes). ddHodge goes further with a full hodge-decomposition (gradient/curl/harmonic), yielding a potential-landscape (empirically validating Waddington: embryogenesis ~88% gradient), divergence as differentiation potency, and acceleration. Both are geometry-preserving and reject UMAP velocity-embedding — but both inherit scale rather than anchoring it.
-
The spatial axis — coupling dynamics to spatial context. TopoVelo is the spatial sibling of RegVelo: transcription is predicted from a cell’s spatial neighbors (GNN) instead of a GRN (spatial-velocity), and it adds a second velocity — physical cell migration in μm/hour.
-
The temporal / physical-time axis — calibrating the inferred clock to real time. This is open for general developmental trajectories, but now solved on the cell cycle. Snapshot methods (scVelo, veloVI, RegVelo, GraphVelo, ddHodge) all produce ordinal latent-time / pseudotime, validated by rank correlation (RegVelo’s FUCCI Spearman ≈ 0.68; GraphVelo’s ρ ≈ 0.83 vs embryo time; ddHodge’s potential vs stage), with rates known only up to scale. TopoVelo reaches physical units only by assuming a 20 h gene cycle (metric-by-assumption). But VeloCycle (velocycle-2024, from velocyto-2018’s author GioeleLaManno) now reaches validated metric time on the cell cycle — a manifold-constrained Bayesian model that gives a cell-cycle period in hours matching live-imaging + EdU, without metabolic labeling, by exploiting the cycle’s known periodicity and a measured half-life. So the lesson sharpens: space can be anchored (μm), regulation coupled (GRN/neighbors), geometry made consistent (manifold/Hodge) — and metric time is now reachable too, given constrained geometry + a measured rate scale. Two methods reach metric time by different routes: dynamo via metabolic-labeling, VeloCycle via periodicity + measured half-lives + live-imaging validation. The residual open problem is the general, non-periodic, unknown-geometry trajectory.
Critique backdrop
The temporal axis is not just our framing — there is an active velocity-skepticism current saying mainstream velocity over-claims. JianhuaXing (dynamo/GraphVelo PI) argues, author-to-author, that latent time is a geometric reaction coordinate unrelated to physical time and that scVelo silently swapped the 2018 physical velocity for a “pseudo-velocity” (xing-hu-regvelo-debate); LiorPachter (“gobbledygook”) and GennadyGorin (biophysics / normalization) attack the same identifiability gap. Two independent reliability benchmarks now back the critique empirically: velocity-benchmark-17studies (Luo et al.; breadth, 14×17) and velocity-benchmark-ancheta-2026 (Ancheta et al., PLoS Comput Biol 2026; depth, 5×3) both conclude no method is uniformly best and velocity is hypothesis-generating — and both score only direction, never physical time. This is author-level + empirical corroboration of the wiki’s lens — honest backdrop for FlowVelo, not polemic (see velocity-discourse-2025-2026).
The seam FlowVelo targets is axis 2: replacing an inferred consensus clock with flow-matching / transport-operators dynamics that can carry physically interpretable time — while being honest about whether metric time is reachable without an absolute-time anchor at all (see regvelo-physical-time-critique).
Method scorecard (physical time)
| Method | Latent time | Rate scale | External anchor | Verdict |
|---|---|---|---|---|
| velocyto (2018 origin) | physical rate, hours | ratio γ; labeling pins step | metabolic labeling | physically grounded origin |
| scVelo | ordinal (admits scale needs prior) | ratios only | none | ungrounded |
| veloVI | ordinal | relative | none | ungrounded |
| RegVelo | ordinal (FUCCI ρ≈0.68) | α GRN-driven; β,γ relative | none | mechanistically grounded, temporally ungrounded |
| UniTVelo | ordinal unified ([0,1]) | RBF fixes shape, scale inherited | none | time-centric yet ordinal |
| TFvelo | ordinal ([0,1] sine) | inherited, deepened (no splicing) | none (drops splicing too) | splicing-free GRN extreme; furthest from physical time |
| GraphVelo | ordinal (ρ≈0.83 vs embryo time) | magnitude preserved but inherited; α,γ relaxed | none of its own | manifold-consistent, scale-preserving but scale-inheriting |
| ddHodge | ordinal potential (+ acceleration) | inherited (relative) | none of its own | geometry-preserving reconstruction; potency + acceleration axes |
| TopoVelo | metric-by-assumption (20 h gene cycle) | expression time assumed; space real (μm) | partial: μm + live-imaging migration match | closest to physical units; timescale imported, not identified |
| veloVI | ordinal (+ uncertainty + applicability test) | relative | none | ungrounded but self-aware |
| VeloCycle | metric on cell cycle (period in hours) | broken by measured half-life | periodicity + half-lives + live-imaging/EdU | metric & experimentally validated (periodic processes) |
| dynamo | metric (w/ labels) | absolute (w/ labels) | metabolic labeling | grounded with labeling |
| FlowVelo | metric (target) | absolute (target) | TBD | the gap this work targets |
Sources ingested
Syntheses: physical-time-grounding-across-methods contrasts all methods on the four axes (draft-ready Related Work); regvelo-physical-time-critique drills into RegVelo.
- regvelo — RegVelo (Wang et al., Cell 2026). GRN-coupled deep-generative velocity.
- graphvelo — GraphVelo (Chen et al., Nat Commun 2025). Tangent-space velocity refinement, multimodal.
- ddhodge — ddHodge (Maehara & Ohkawa, Nat Commun 2025). Hodge-decomposition vector-field reconstruction; potential landscape, divergence-potency, acceleration.
- topovelo — TopoVelo (Gu et al., Nat Biotechnol 2025). Spatially coupled velocity from spatial transcriptomics; migration velocity in μm/hour; metric-by-assumption.
- xing-hu-regvelo-debate — Xing↔Hu correspondence (2026-06-05). Primary-source critique; author-level physical-time-grounding thesis. Private — don’t quote verbatim.
- velocity-discourse-2025-2026 — tweet digest. Cell2fate, spVelo, Pachter/Gorin skepticism, RegVelo announcement, UCSF benchmark.
- velocity-benchmark-17studies — Luo et al. (Cell Rep Methods 2026). 14 methods × 17 datasets; no method uniformly best (the benchmark Pachter amplified).
- velocity-benchmark-ancheta-2026 — Ancheta et al. (PLoS Comput Biol 2026, CZ Biohub SF/UCSF). 5 methods × 3 datasets in depth; no winner, hypothesis-generating; the “UCSF benchmark.”
- monod-gorin — Gorin et al. (Nat Methods 2025). CME stochastic modeling minimizing distortive normalization. Full text.
Foundational + method papers (2026-06-12 batch): velocyto-2018 (origin), scvelo-2020 (dynamical model + latent time), multivelo-2022 (chromatin), cell2fate-2025 (Bayesian modules), spvelo-2025 (multi-batch spatial), deepvelo-2022 (neural-ODE field), celldancer-2023 (cell-specific rates), unitvelo-2022 (unified time + top-down RBF), tfvelo-2024 (splicing-free, GRN/TF-driven velocity on total mRNA), velovi-2023 (variational; uncertainty + applicability test), velocycle-2024 (manifold-constrained Bayesian; validated metric cell-cycle time — the temporal-axis landmark; velocycle-nv-2024 commentary), cell-growth-omission-2025 (missing growth term; full text — γ*≈γ+λ, growth-biased labeling rates, orthogonal/reversed trajectories).
All raw/papers/ PDFs ingested (origin → latest), plus the 2026-06 notes, tweet bookmarks, and
fetched skeptic sources. The corpus now spans the full lineage; the origin reframe (physical
time present in 2018, lost in snapshot successors) and the four-strand velocity-skepticism
are the load-bearing additions for FlowVelo.