CellRank

中文導讀

CellRank 不是 velocity estimator,而是 velocity 的 downstream 框架:把一個 velocity vector field(或 pseudotime kernel)轉成 cell–cell transition 的 Markov chain,算 terminal states(macrostates)、fate probabilities、driver genes。RegVelo 的 in silico perturbation 就是接 CellRank——perturb 後重算 fate probability,用 depletion likelihood 量 lineage 影響。它本身不碰 physical time,inherit 上游 velocity 的 grounding 狀態。

What it is

A downstream framework that converts a velocity field (or other directional signal, via pluggable “kernels” like PseudotimeKernel) into a Markov chain over cells, from which it computes macrostates / terminal states, terminal-state identification (TSI), fate probabilities, and lineage-driver rankings via gene–fate correlation.

Role with the velocity methods

• Consumes velocities from scVelo, veloVI, RegVelo, dynamo.
• For RegVelo: the substrate for in silico perturbation — re-solve a masked regulon’s velocity, push through CellRank, and read off the change in fate probabilities (CellRank 2’s “depletion likelihood”, 0–1).

Physical-time note

CellRank does not itself address physical time; the Markov transitions are built from whatever directional/temporal signal the upstream method provides. It inherits the ordinal-vs-metric status of its input (see physical-time-grounding).

Related

RNA velocity · RegVelo · scVelo · veloVI · dynamo · latent-time