cellDancer
中文導讀
cellDancer(Li et al., Nat Biotech 2023,“relay velocity model”)用 per-gene DNN 推 cell-specific α、β、γ,再把局部 velocity 在鄰域間 relay,給 single-cell-resolution kinetics。它跟 GraphVelo 是兩個主要「放寬 constant-rate」的方法,專治 MURK genes(scVelo/DeepVelo/VeloVAE 在這些基因上方向會 反)。注意:它的 loss 是 cosine similarity(只管方向),所以 magnitude/scale 沒被 pin 住—— direction-faithful 但拿不到 metric speed。physical time 上 ordinal。
What it is
A deep-neural-network RNA-velocity method (source celldancer-2023) inferring cell-specific α, β, γ per gene and relaying local velocities across neighbors for single-cell-resolution kinetics. Resolves multiple-rate-kinetic (MURK) genes that invert standard models.
Physical-time scorecard
| Axis | cellDancer |
|---|---|
| Latent time | ordinal (velocity-based / pseudotime) |
| Rate scale | inherited + weakly pinned (cosine-similarity loss optimizes direction, not magnitude) |
| External anchor | none (snapshot) |
| Constant rates | relaxed — cell-specific α, β, γ (per-cell DNN) |
| Verdict | cell-resolution kinetics; constant-rate relaxed; direction-only, ordinal |
Relation to other methods
- The per-cell-DNN counterpart to GraphVelo’s manifold-based rate relaxation; both fix the multiple-γ problem from velocyto-2018.
- Shows scVelo / DeepVelo / VeloVAE inverted on MURK genes.
- One of the 15 in velocity-benchmark-17studies.
Related
celldancer-2023 · GraphVelo · velocyto-2018 · scVelo · DeepVelo · VeloVAE · splicing-kinetics-ode · velocity-benchmark-17studies · latent-time · physical-time-grounding · FlowVelo