DeepVelo (Chen et al., Sci Adv 2022)

Chen, King, Hwang, Gerstein & Zhang. Science Advances 8:eabq3745 (2022). Yale / Microsoft / UCI. (Distinct from the 2024 Genome Biology “DeepVelo”, Cui et al., a GCN method — name collision; the benchmark velocity-benchmark-17studies lists DeepVelo twice.)

Summary

DeepVelo learns a neural-ODE velocity field: a VAE f_A(x) maps a cell’s expression state to its rate of change dx/dt = f_A(x), capturing nonlinear gene interactions (vs the linear dx/dt = Ax assumption), reducing dimensionality and denoising the field. Integrating f_A with a black-box ODE solver extrapolates each cell’s future/past over a longer period than the hours-scale instantaneous velocity allows, and supports trajectory simulation, a cell criticality index (instability), and in silico perturbation. For the wiki it is a neural vector-field method in the dynamo / GraphVelo family (post-velocity field modeling), temporally ordinal.

Key Claims

  • Neural-ODE field. dx/dt = f_A(x) learned by a VAE; nonlinear regulatory cascade; integrate to predict future/past states (Euler), extrapolating beyond the instantaneous hours-scale.
  • Cell criticality index (CCI). Velocity-based latent time + an instability score flagging fate-commitment regions; driver genes via CCI correlation (e.g. Neurog3).
  • In silico perturbation. Perturb initial conditions → shift cell-fate bifurcation proportions (alpha vs beta). Notes single-cell dynamical systems can be chaotic.
  • Benchmarks favorably on out-of-sample velocity prediction; in the Luo benchmark DeepVelo is among the better-balanced accuracy/usability methods (GPU, fast, low memory).

Physical-time grounding (standing lens)

  1. Latent time — ordinal or metric? Ordinal (velocity-based latent time + CCI); ODE integration extends the range of extrapolation but not the units.
  2. Scale degeneracy. Inherited — integrates a learned field in arbitrary step units; relative.
  3. External anchor. None (snapshot scRNA).
  4. Constant-rate assumptions. Abandons the explicit α/β/γ parameterization for a learned nonlinear field f_A(x); inherits the scale of the input/measured velocity it is trained on.

Family resemblance: like dynamo and GraphVelo, DeepVelo is a vector-field layer on top of velocity, not a new time anchor — fits the geometric/field axis of physical-time-grounding-across-methods, not the temporal one.

Connections

Contradictions

  • None. Reinforces that neural field-learning extends range/denoises but does not anchor time.