Jianhua Xing (邢建華)

中文導讀

Jianhua Xing(邢建華,Univ. Pittsburgh)是 dynamo 與 GraphVelo 背後的 PI,理論物理 出身,主張單細胞 dynamics 要建在嚴格的 chemical-physics / dynamical-systems theory 上 (chemical master equation、Langevin、Koopman operator、Zwanzig–Mori projection),而不是 靠 intuition-driven ML。他是本 wiki physical-time 軸上最直接的 velocity-skeptic:他認為 latent time 根本是個 geometric reaction coordinate、跟 physical time 無關,scVelo 把 2018 的 physical velocity 偷換成 pseudo-velocity(見 xing-hu-regvelo-debate、 velocity-skepticism)。

Who

PI at the University of Pittsburgh (Computational & Systems Biology / Physics). Senior author of GraphVelo (Nat Commun 2025) and a lead of dynamo (Cell 2022). Theoretical-physics background; built a projection-operator / dynamical-systems framework for reconstructing cellular dynamics from 2005–2011, applied once the big-data era arrived. Also runs live-cell-imaging dynamics reconstruction (Wang et al. Sci Adv 2020; eLife 2022).

Position (the velocity critique)

Xing is the wiki’s sharpest internal critic on the temporal axis (see xing-hu-regvelo-debate, velocity-skepticism):

• Latent time is a geometric reaction coordinate with no intrinsic relation to physical time — mislabeling it “time” is a concept-swap; pseudotime/latent time = chemical-kinetics reaction coordinate.
• Most “velocity” is the derivative of cell state w.r.t. that coordinate, not a physical velocity vector field. The silent switch happened from 2018 (RNA velocity, physically defined) to scVelo’s pseudo-velocity.
• dynamo is a new paradigm (systems-biology simulation + metabolic-labeling physical time), not an incremental velocity fix; GraphVelo is the rigorous, minimal fix to scVelo’s steady-state problem that preserves the 2018 biophysical model.
• Deep-learning velocity models add no new physics — on his view a neural network is a basis-function approximator, not new dynamics; so RegVelo does not, in his reading, go beyond the dynamo framework, and MultiVelo is internally inconsistent.

Methods / theory

dynamo · GraphVelo · dynamical-systems-formulation (Koopman / Zwanzig–Mori projection; Galerkin/NN basis-function vector-field reconstruction; manifold assumption; Whitney embedding for cross-modality velocity).

Relation to others

• Lab line: dynamo → GraphVelo (and the planned “dynamo2” interpretable-NN note, 2022).
• Critiques RegVelo & FabianTheis’s line and MultiVelo.
• Critique cluster with LiorPachter and GennadyGorin (velocity-skepticism).
• Cordial exchange with HuYizhou (2018 RNA-velocity preprocessing contributor) in xing-hu-regvelo-debate.

Related

dynamo · GraphVelo · dynamical-systems-formulation · HuYizhou · velocity-skepticism · physical-time-grounding · latent-time · RNA velocity · RegVelo · MultiVelo · metabolic-labeling · FlowVelo