cellDancer (Li et al., Nat Biotechnol 2023)
Li, Zhang, Chen, Ye, Brannan, Le, Abe, Cooke & Wang. Nature Biotechnology (2023). Houston Methodist / Weill Cornell. The “relay velocity model.”
Summary
cellDancer relaxes the universal-kinetics assumption by inferring cell-specific α, β, γ (transcription, splicing, degradation) with a per-gene DNN, then relays locally-inferred velocities across neighborhoods to give single-cell-resolution kinetics. This fixes the inverted directions standard models produce on multi-rate-kinetic (MURK) and multi-lineage genes. For the wiki it is — with GraphVelo — one of the two main constant-rate-relaxing methods, but at the cell-resolution kinetics level; temporally ordinal, and trained on a cosine-similarity (direction) loss that weakly constrains magnitude.
Key Claims
- Cell-specific rates. Per gene, a DNN predicts α(t_i), β(t_i), γ(t_i) from (u_i, s_i); neighbor cells share local rates (“relay”). Cell-specific α, β, γ are proposed fate indicators.
- Resolves MURK genes. On the 89 MURK genes (e.g. Smim1, Hba-x) cellDancer recovers the correct differentiation direction where scVelo, DeepVelo and VeloVAE were inverted; lowest error in transcriptional-boost / multi-lineage simulated regimes.
- Direction-based training. Loss = Σ_cells (1 − cos θ) between predicted and observed velocity — i.e. optimizes direction, leaving magnitude/scale weakly pinned.
- Robust + scalable. Holds up under high dropout, sparse data, imbalanced lineage sizes; modular, parallelized.
Physical-time grounding (standing lens)
- Latent time — ordinal or metric? Ordinal (velocity-based / differentiation pseudotime).
- Scale degeneracy. Inherited and aggravated by the cosine-similarity loss — direction is fit, magnitude/scale is not; no absolute anchor.
- External anchor. None (snapshot scRNA).
- Constant-rate assumptions. Relaxed — cell-specific α, β, γ (the headline), like GraphVelo but via a per-cell DNN rather than a manifold tangent-space argument.
Pairs with GraphVelo as the “let rates vary per cell” answer to the multiple-γ problem flagged all the way back in velocyto-2018. Note the magnitude caveat: a cosine loss makes cellDancer direction-faithful but explicitly not a route to metric speed (see physical-time-grounding).
Connections
- cellDancer — the method entity.
- GraphVelo — the other constant-rate-relaxing method (manifold vs per-cell DNN).
- velocyto-2018 — whose multiple-γ failure mode this targets.
- scVelo / DeepVelo / VeloVAE — the methods it shows inverted on MURK genes.
- velocity-benchmark-17studies — among the 15 (moderate consistency).
- splicing-kinetics-ode / latent-time / physical-time-grounding — cell-specific rates, ordinal time.
- FlowVelo — cautionary: direction-only training does not yield metric speed.
Contradictions
- None. Adds a per-cell-rate exception alongside GraphVelo on splicing-kinetics-ode’s constant-rate caveat.