GraphVelo allows for accurate inference of multimodal velocities and molecular mechanisms for single cells

GraphVelo: multimodal velocities and molecular mechanisms for single cells

Chen, Zhang, Gan, Ni, Chen, Bahar & Xing. Nature Communications 16, 7831 (2025). https://doi.org/10.1038/s41467-025-62784-w (Xing lab, Pitt/Zhejiang; Bahar).

Summary

GraphVelo is a graph-based, dynamical-systems post-processing plugin that takes the RNA velocity produced by any existing method (scVelo, dynamo, veloVI) and refines it by projecting onto the tangent space of the data manifold (manifold-consistent-velocity). Unlike La Manno’s cosine kernel — which gives the right direction but discards magnitude — GraphVelo preserves both direction and magnitude, so velocity norm becomes a usable “cell speed.” It also transforms velocities across data representations (Whitney embedding), extending velocity to multimodal data (chromatin, spatial, host–virus). For this wiki the two notable moves are: (1) it keeps relative speed instead of throwing it away, and (2) it relaxes the constant-rate assumption, recovering cell-context-specific α and γ for multiple-rate-kinetics (MURK) genes that mislead scVelo.

Key Claims

• Tangent-space projection (TSP). The true velocity of a cell on a manifold M must lie in the tangent space T_pM. GraphVelo builds a kNN graph, forms local basis vectors from neighbor displacements, and projects the input velocity onto T_pM by minimizing a TSP loss L(φ) = a·‖v∥ − v_i‖² − b·cos(φ, φ^corr) + λ‖φ‖² — the first term recovers magnitude, the second pins direction (cosine-kernel limit).
• Magnitude is preserved. Benchmarks show GraphVelo retains velocity magnitude (and thus cell speed) where the cosine kernel fails; it improves CBC and RMSE vs ground truth across linear/cyclic/bifurcating dyngen topologies.
• Velocity transformation across representations. Via local linear embedding, velocity can be moved between gene space, PCA space, and other modalities (host–virus interactome, multi-omics, spatial) — a Whitney-embedding argument.
• MacK genes. “Manifold-consistent Kinetics” genes are those whose splicing-based velocity sign agrees with the manifold-derived differentiation direction (MacK score). Whole-genome velocity is inferred from a high-confidence MacK subset.
• Relaxes constant rates (MURK genes). For multiple-rate-kinetics genes (rapid degradation, transcription bursts) that violate scVelo’s constant-γ assumption, GraphVelo recovers cell-context-specific α = u + du/dt and γ = (u − ds/dt)/s, correcting erroneous velocity signs (e.g. Smim1, Hba-x, ANGPT1, RBPMS).
• Downstream. Feeds dynamo vector-field / differential-geometry and Jacobian analyses, and CellRank terminal-state detection; reveals pioneer-TF regulation (Lef1, Hoxc13 → Wnt3) in multi-omics hair-follicle data.

Physical-time grounding (standing lens)

1. Latent time — ordinal or metric? GraphVelo does not infer latent time; it refines velocity vectors. Downstream GraphVelo pseudotime is validated by rank correlation — Spearman ρ = 0.831 vs erythroid embryo time, ρ = 0.601/0.980 vs viral-RNA% — i.e. ordinal. It does yield a relative cell speed (velocity L2-norm), a step most refinement layers discard.
2. Rate–time scale degeneracy. GraphVelo is scale-preserving but scale-inheriting: it keeps the magnitude of the input velocity rather than normalizing it away, but that magnitude carries whatever scale the input method had. It does not itself break the degeneracy of the splicing-kinetics-ode; fed labeling-derived velocities it propagates the absolute scale, fed snapshot velocity it stays relative.
3. External time anchor. None of its own. Validates cell speed against metabolic-labeling data (sci-fate, cycling A549; stratified cell-cycle speed), and can take metabolic-labeling-derived velocity as input — but the plugin adds no anchor by itself.
4. Constant-rate assumptions. Explicitly relaxed — the headline mechanistic advance. Recovers cell-context-specific α and γ for MURK genes, where scVelo’s constant degradation rate gives wrong velocity signs. This directly addresses the “constant β,γ biases the time mapping” critique at the velocity level (though not by anchoring absolute time).

GraphVelo is the first method in this wiki to both retain relative magnitude and let rates vary along the trajectory — partial progress on physical-time realism from the velocity side, without an absolute-time anchor. Contrast with RegVelo (mechanism via GRN) and dynamo (absolute scale via labels).

Key Quotes

“a visually correct vector field does not necessarily imply accurate high-dimensional velocity estimation. La Manno et al. proposed a cosine kernel method … [which] asymptotically gives the correct direction of a velocity vector in the large sampling limit, but the magnitude information is completely lost due to a normalization procedure.”

“GraphVelo revealed a cell context-specific transcription rate α = u + du/dt and degradation constant γ = (u − ds/dt)/s, thus a degradation wave along the differentiation path” — relaxing the constant-rate assumption for MURK genes.

Connections

• GraphVelo — the method this source defines.
• manifold-consistent-velocity — its core mechanism (TSP, MacK genes, transformation).
• scVelo / veloVI / dynamo — input velocity providers it refines.
• splicing-kinetics-ode — whose constant-rate assumption GraphVelo relaxes (MURK).
• dynamo — downstream vector-field / differential-geometry engine.
• CellRank — downstream terminal-state framework.
• metabolic-labeling — used to validate cell speed; possible input.
• physical-time-grounding — scale-preserving but scale-inheriting; rates relaxed.
• FlowVelo — our work; GraphVelo’s tangent-space/manifold framing is methodologically adjacent.

Contradictions

• No conflict with existing pages. GraphVelo sharpens the splicing-kinetics-ode page’s standing caveat that “nearly every method holds β, γ constant”: GraphVelo is the exception that lets them vary — but at the velocity-correction level, still without an absolute-time anchor, so it does not contradict physical-time-grounding.