Geometry-preserving vector field reconstruction of high-dimensional cell-state dynamics using ddHodge

ddHodge: geometry-preserving vector-field reconstruction of cell-state dynamics

Maehara & Ohkawa. Nature Communications 16, 11342 (2025). https://doi.org/10.1038/s41467-025-67782-6 (Kyushu University). Julia package.

Summary

ddHodge is a vector-field reconstruction framework that takes scRNA-seq data plus velocities (e.g. RNA velocity) as input and decomposes the field via Hodge decomposition into gradient (potential), divergence-free rotational (curl), and harmonic components on the data manifold. Its distinctive yields are (1) a potential landscape that empirically validates the Waddington gradient picture with real embryogenesis data, (2) divergence as a measure of differentiation potency / cell-state stability, and (3) acceleration — the second-order change of cell state, which prior velocity methods could not extract. Like GraphVelo it is a geometry-preserving post-processing layer (manifold-consistent-velocity); its temporal output (the potential) is an ordinal pseudotime and, importantly, only faithful where the dynamics are gradient-dominated.

Key Claims

• Hodge decomposition on the manifold. ω = grad α + curl·β + γ (gradient + rotational + harmonic residual). ddHodge reconstructs all three plus the Jacobian and second-order (acceleration) information, even from biased/sparse samples.
• Geometry-preserving manifold approximation. Local PCA tangent spaces patched by a sheaf (connection) Laplacian into near-parallel coordinates — avoids the geometric distortion that UMAP velocity-embedding introduces (density equalization destroys lengths/volumes). Methodologically a more complete cousin of GraphVelo’s tangent-space projection.
• Potential = ordering (pseudotime). The gradient component’s potential gives the temporal order of cell states; correlates with developmental stage. Caveat: where the divergence-free (cyclic) flow dominates, the potential cannot provide a faithful temporal ordering (shown on FUCCI cell cycle: 49% cyclic).
• Divergence = stability / differentiation potency. Negative divergence = convergence (canalization, stable); positive = divergence (plastic, unstable). Mouse embryogenesis: early cells (E6.5–7.5) positive divergence (plastic) → later (E8.5–9.5) negative (canalize). Quantifies potency without a separate model.
• Curl / harmonic = oscillation / cycle (e.g. cell cycle), independent of the gradient.
• Acceleration & richer indicators. Recovers acceleration (deceleration/ acceleration along the path), Grassmann distance (change of gene modules / dimensional change), and Schur vectors (cross-sections selecting driver genes).
• Empirical Waddington. Embryogenesis is ~88% gradient / ~12% rotational — i.e. development is mainly a potential-landscape (gradient) system, validated with real data for the first time (vs FUCCI’s ~50/50, cyclic-dominated).
• Benchmarks. More accurate divergence/curl/Jacobian reconstruction than DEC (finite element) and than dynamo’s SparseVFC, with favorable scaling (sparse matrices, Krylov solvers).

Physical-time grounding (standing lens)

1. Latent time — ordinal or metric? Ordinal. The “potential” orders cell states (pseudotime), validated by correlation with developmental stage — a rank-type check. Notably it is conditionally valid: only where dynamics are gradient-dominated; in cyclic regions the potential gives no faithful ordering. ddHodge adds a richer temporal structure than mere ordering — it measures acceleration/deceleration (whether the state moves at constant speed) — but still in relative, not metric, units.
2. Rate–time scale degeneracy. Inherited. ddHodge consumes upstream velocity and reconstructs the field’s geometry; it works in scaled velocity units (noise sims use γ̄ = γ/β = 1) and adds no absolute scale. The scale degeneracy of the splicing-kinetics-ode passes through untouched.
3. External time anchor. None. Embryogenesis analysis uses VASA-seq (high intron coverage → accurate velocity) — but VASA-seq is still snapshot, not metabolic-labeling; no real-time anchor.
4. Constant-rate assumptions. ddHodge makes none of its own (it does not model splicing kinetics); it inherits the upstream method’s assumptions. It can surface rate-change events — in a simulation with cell-type-specific transcription-rate changes at t=3, 7, ddHodge detected the rate changes via divergence shifts.

ddHodge is the first method in this wiki to extract acceleration and to recast the temporal question as a potential landscape (potential-landscape) plus a potency axis (divergence). This enriches the structure of inferred time but does not make it metric — potential ordering is ordinal and conditional on gradient dominance. Contrast: dynamo (absolute via labels), RegVelo (regulatory), GraphVelo (magnitude-preserving manifold projection).

Key Quotes

“the gene expression dynamics during development follow a gradient system shaped by potential landscapes, which has not previously been validated with real data.”

“The shape of the potential indicates whether the cell state traverses the landscape at a constant speed (flat slope) … or experiences acceleration … or deceleration (changing slope).” — divergence ≠ 0 ⇒ non-constant speed.

“In regions where the divergence-free flow dominates, potential cannot provide a faithful temporal ordering.” — the honest limit of potential-as-pseudotime.

Connections

• ddHodge — the method this source defines.
• hodge-decomposition — its mathematical core (grad/curl/harmonic).
• potential-landscape — the Waddington gradient picture it validates with real data.
• manifold-consistent-velocity — shared geometry-preserving stance with GraphVelo.
• dynamo — the vector-field-reconstruction method it benchmarks against and beats on div/curl/Jacobian.
• GraphVelo — grouped with ddHodge as downstream velocity-field reconstruction (vs estimation).
• splicing-kinetics-ode — upstream; ddHodge inherits its scale degeneracy.
• latent-time — potential = ordinal pseudotime, plus acceleration as richer structure.
• physical-time-grounding — ordinal, scale-inheriting; adds acceleration/potency axes.
• FlowVelo — our work; the potential-landscape / geometry framing is directly relevant.

Contradictions

• No conflict with existing pages. ddHodge reinforces physical-time-grounding: its potential is explicitly an ordering, faithful only in gradient-dominated regions, and it inherits the splicing ODE’s scale degeneracy. It enriches (does not contradict) the temporal picture by adding acceleration and a potency (divergence) axis.