PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference

Chenglei Yu, Chuanrui Wang, Bangyan Liao, Tailin Wu

TL;DR

PACE (Geometry-Aware Bridge Transport) is a new framework for reconstructing continuous biological trajectories from "destructive" single-cell snapshots. By replacing standard Euclidean distances with a geometry-aware Riemannian metric that penalizes motion off the cellular manifold, PACE significantly improves the accuracy of inferred cell transitions, especially in complex, asynchronous developmental processes.

What Problem Does This Paper Address?

In single-cell biology, researchers often want to know how a cell evolves from one state to another (e.g., during stem cell differentiation). However, measuring a cell’s gene expression usually destroys it, meaning we only get "snapshots" of different populations at different time points rather than a continuous movie of a single cell.

Reconstructing these paths is an "ill-posed" problem: there are infinite ways to connect dots between Time A and Time B. Most existing methods rely on Optimal Transport (OT) using Euclidean distance, assuming cells move to their nearest neighbors in expression space. This fails in two major ways:

  1. Asynchrony: Cells sampled at the same time might be at different developmental stages. A "fast" cell at Day 2 might actually correspond to a "slow" cell at Day 5.
  2. Manifold Geometry: Biological transitions follow specific "valleys" in a high-dimensional landscape. Euclidean distance is a straight line that often "cuts corners," suggesting biologically impossible shortcuts across the manifold.

Methods

PACE introduces a three-stage pipeline to solve these issues:

  1. Anisotropic Riemannian Metric: Instead of treating all directions of motion equally, PACE uses local Principal Component Analysis (PCA) to identify "tangent" directions (plausible biological change) and "normal" directions (implausible shortcuts). It constructs a metric that makes moving in the "normal" direction much more "expensive" than moving along the manifold.
  2. Iterative Bridge Transport: The framework uses a "bootstrap" approach. It alternates between:
  3. Coupling Refinement: Deciding which cells at Time A correspond to which at Time B based on the current geometric cost.
  4. Neural Bridge Fitting: Training a neural network to find the optimal, geometry-consistent path between those coupled endpoints.
  5. Velocity Distillation: Finally, the learned specific paths are distilled into a global, continuous-time velocity field. This allows researchers to simulate the trajectory of any cell from any starting point.

Datasets

The authors evaluated PACE across seven datasets, including both synthetic controls and real biological systems:

  • Controlled 2D: "Ocean" (a rotational non-gradient flow) and "Two-Branch" (a simple bifurcation).
  • Biological Time-Courses: Embryoid body differentiation (EB PHATE), human iPSC reprogramming (iPSC-Liu), and mouse reprogramming (Schiebinger2019).
  • High-Dimensional Benchmarks: 100D CITE-seq and Multiome data from the NeurIPS Open Problems challenge.

Key Insights and Ideas

  • Geometry as a Prior: The core "aha moment" is that while we don't know the exact path, the local density of cells provides a "weak prior." If cells are clustered along a curve, motion along that curve is more likely than motion perpendicular to it.
  • Overcoming the "Curse of Dimensionality": In high-dimensional spaces, Euclidean distances tend to concentrate (everything looks roughly the same distance away). By focusing on local anisotropy (the direction of variation), PACE maintains a signal even when global distances become uninformative.
  • No Supervision Needed: Remarkably, PACE achieves these results without requiring RNA velocity data, lineage tracing, or explicit cell pairing during training. It learns the "rules" of the manifold purely from the distribution of the snapshots.

Key Results

  • Superior Reconstruction: Across nine held-out experiments, PACE reduced distributional error (measured by MMD and Wasserstein distances) by an average of 23.7% compared to the strongest baselines.
  • Biological Realism: On the embryoid body benchmark, PACE improved alignment with RNA velocity (a proxy for the true direction of cell change) by 15.4%, despite not seeing that data during training.
  • Robustness: Ablation studies showed that all components—the anisotropic metric, the local neighborhood construction, and the iterative rematching—are necessary for these gains.

Limitations and Future Work

The authors identify a few key areas for improvement:

  • Sensitivity to Sparsity: The local metric relies on nearest-neighbor estimates. If the data is extremely sparse or noisy, the estimated "tangent" directions might become unstable.
  • Closed Populations: The current model assumes a "closed" population. In reality, cells divide and die. The authors plan to incorporate birth-death or growth terms in future versions.
  • Scalability: While PACE performs well in 100D, the computational cost of iterative matching is higher than simpler, geometry-agnostic methods. Future work may look at more efficient ways to estimate these metrics in even higher-dimensional latent spaces.

Summary updated

Abstract

Single-cell trajectory inference from destructive time-course snapshots is fundamentally ill-posed: neither cross-time cell correspondences nor the continuous paths between snapshots are observed, so the observed snapshot distributions alone do not uniquely determine the underlying dynamics. Existing optimal transport and flow-based methods typically couple cells by Euclidean proximity at observed clock times, which can misalign trajectories when development is asynchronous and cells sampled at the same experimental time occupy different latent pseudotime stages. We propose PACE, a trajectory inference framework that selects geometry-consistent continuous transport dynamics from destructive time-course snapshots through three coupled components. First, PACE constructs a state- and time-dependent anisotropic Riemannian metric that preserves low cost along locally supported tangent directions while penalizing normal velocity components. Second, it alternates between refining cross-time couplings under the induced path-action cost and fitting endpoint-preserving neural bridges between adjacent snapshots. Third, it distills the learned bridge dynamics into a global continuous-time velocity field over cellular states. Across seven controlled and biological datasets covering nine held-out reconstruction experiments, PACE achieves the strongest overall reconstruction performance, reducing MMD, $\mathcal{W}_1$, and $\mathcal{W}_2$ by 23.7\% on average relative to the strongest competing baseline. PACE also improves RNA-velocity alignment by 15.4\% on an embryoid body differentiation benchmark, without requiring explicit cell pairing, lineage tracing, or RNA velocity supervision during training. Code is available at \url{https://github.com/AI4Science-WestlakeU/PACE}.
BibTeX
@inproceedings{yu2026pace,
  author = {Chenglei Yu and Chuanrui Wang and Bangyan Liao and Tailin Wu},
  title = {{PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference}},
  booktitle = {GenBio @ ICML},
  year = {2026},
  url = {https://openreview.net/forum?id=tMu0zdfnqp},
}
Figure