DeepVelo (Chen et al. 2022)
中文導讀
DeepVelo(Chen et al., Sci Adv 2022,Yale/Gerstein lab)用 neural ODE 學一個 velocity field dx/dt = f_A(x)(VAE),抓 nonlinear gene interaction,再用 ODE solver 積分往未來/過去外推,做 trajectory、cell criticality index、in silico perturbation。屬於 dynamo/GraphVelo 那種 post-velocity vector-field 家族,physical time 上是 ordinal。注意:跟 2024 Genome Biology 的 另一個 DeepVelo(Cui et al., GCN)同名不同 paper。
What it is
A neural-ODE velocity-field learner (source deepvelo-2022): a VAE predicts dx/dt = f_A(x); integrating it extrapolates cell states over a longer horizon than instantaneous velocity. Name collision: a separate 2024 Genome Biology DeepVelo (Cui et al., GCN, cell-specific kinetics) shares the name.
Physical-time scorecard
| Axis | DeepVelo (2022) |
|---|---|
| Latent time | ordinal (velocity-based + cell criticality index) |
| Rate scale | inherited; ODE integration extends range, not units |
| External anchor | none (snapshot) |
| Constant rates | learned nonlinear field f_A(x) (no explicit α/β/γ) |
| Verdict | neural vector-field layer; temporally ordinal |
Relation to other methods
Vector-field family with dynamo, GraphVelo, ddHodge. Among the stronger methods in velocity-benchmark-17studies. Neural-ODE precedent relevant to FlowVelo.
Related
deepvelo-2022 · dynamo · GraphVelo · ddHodge · cellDancer · velocity-benchmark-17studies · latent-time · physical-time-grounding · FlowVelo