Jiaqi Gan

GraphVelo

Nature Communications · 2025

Open-source package for multimodal single-cell velocity.

Problem

RNA velocity estimates each cell's direction and speed of change, but raw estimates are noisy, common projections lose the speed, and velocities from one measurement type do not carry over to another.

My contribution

  • Co-developed the package (one of three equal-contribution authors); implemented local tangent-space projection to refine noisy velocities and map them between RNA and chromatin accessibility.
  • Ran the ground-truth benchmarks and the sensitivity analysis.

Approach

Tangent-space projection
Velocities are projected onto the local tangent space of the cell-state manifold.
Known-answer benchmarks
Linear, cyclic and branching simulations, scored by cosine similarity, RMSE and sign accuracy.
Sensitivity analysis
Robustness to noise and to the choice of embedding.

Results

  • 3 simulated trajectory types; speed information kept that cosine-kernel projection loses
  • PCA > UMAP speed agreement with the known dynamics

Paper

2025 Published

GraphVelo allows for accurate inference of multimodal velocities and molecular mechanisms for single cells

Yuhao Chen*, Yan Zhang*, Jiaqi Gan*, Ke Ni, Ming Chen†, Ivet Bahar†, Jianhua Xing†

Nature Communications 16, 7831 (2025) · DOI 10.1038/s41467-025-62784-w

* These authors contributed equally. † Corresponding authors.