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

Bangyan Liao; Chenglei Yu; Chuanrui Wang; Tailin Wu

arxiv: 2605.18587 · v2 · pith:YQ22GJTZnew · submitted 2026-05-18 · 🧬 q-bio.GN · cs.LG

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

Chenglei Yu , Chuanrui Wang , Bangyan Liao , Tailin Wu This is my paper

Pith reviewed 2026-05-20 01:20 UTC · model grok-4.3

classification 🧬 q-bio.GN cs.LG

keywords single-cell trajectory inferenceoptimal transportRiemannian metricneural bridgesvelocity fielddevelopmental dynamicsasynchronous trajectories

0 comments

The pith

PACE recovers continuous cell trajectories from snapshots by building a time-varying anisotropic metric that favors local developmental directions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Single-cell trajectory inference from destructive snapshots is ill-posed because neither cell correspondences nor continuous paths are observed. Existing methods couple cells by Euclidean proximity, which misaligns trajectories when development proceeds asynchronously. PACE instead constructs a state- and time-dependent anisotropic Riemannian metric that lowers transport cost along locally supported tangent directions and raises it for normal components. It alternates between refining cross-time couplings under the induced path cost and fitting endpoint-preserving neural bridges, then distills the result into a global continuous velocity field. On seven datasets and nine held-out reconstructions, this yields lower MMD and Wasserstein distances than prior baselines while also improving RNA-velocity alignment.

Core claim

PACE shows that a state- and time-dependent anisotropic Riemannian metric can be used to define path-action costs that enforce geometry-consistent couplings between snapshots; alternating optimization between these costs and neural bridge fitting then produces a distilled global velocity field that reconstructs held-out trajectories more accurately than Euclidean optimal transport or flow-based baselines.

What carries the argument

A state- and time-dependent anisotropic Riemannian metric that assigns low transport cost along locally supported tangent directions while penalizing normal velocity components.

If this is right

Reduces MMD, Wasserstein-1, and Wasserstein-2 distances by 23.7 percent on average across nine reconstruction experiments on seven datasets.
Improves alignment with measured RNA velocity by 15.4 percent on an embryoid body differentiation benchmark.
Recovers continuous dynamics without requiring explicit cell pairing, lineage tracing, or RNA-velocity supervision during training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same metric construction could be tested on other asynchronous dynamical systems outside single-cell biology where local geometry is known to matter.
Datasets that supply ground-truth continuous trajectories would allow direct measurement of how much the anisotropy assumption improves path accuracy versus endpoint matching alone.
Combining the distilled velocity field with multi-omics measurements could yield joint trajectory models across transcriptomic and proteomic layers.

Load-bearing premise

That a suitable anisotropic metric reflecting local tangent directions of development can be constructed from the observed snapshots alone.

What would settle it

A controlled simulation in which true trajectories follow known curved paths but the learned metric forces straighter couplings would produce higher reconstruction errors than Euclidean baselines on the same data.

Figures

Figures reproduced from arXiv: 2605.18587 by Bangyan Liao, Chenglei Yu, Chuanrui Wang, Tailin Wu.

**Figure 1.** Figure 1: Overview of PACE. PACE uses local PCA to construct an anisotropic metric Gk(x, t) = I + αC(k) N (x, t), trains endpoint-preserving neural bridges under the corresponding path-action cost, iteratively refines cross-time couplings, and distills the learned bridge dynamics into a global velocity field for trajectory inference from unpaired snapshots. to non-gradient dynamics using approximate velocity informa… view at source ↗

**Figure 2.** Figure 2: Overview of the 2D benchmark datasets. Points are colored by observed time for Ocean [ [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Velocity-alignment diagnostics on Ocean [ [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: High-dimensional concentration diagnostics. Dashed lines mark norm CV = 0.3 and inter-time/withintime ratio = 1.0. 6 11 16 Test Timepoint 0.1 0.2 0.3 0.4 0.5 MMD 6 11 16 Test Timepoint 0.15 0.20 0.25 0.30 0.35 W1 6 11 16 Test Timepoint 0.20 0.25 0.30 0.35 0.40 0.45 W2 PACE No Rematch =0 (Euclidean) KNN = All Spatial Kernel Temporal Kernel [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

read the original abstract

Single-cell trajectory inference from destructive time-course snapshots is fundamentally ill-posed: neither cross-time cell correspondences nor continuous trajectories are observed, so the 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 recovers 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 assigns low transport 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, Wasserstein-1 distance, and Wasserstein-2 distance 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 https://github.com/AI4Science-WestlakeU/PACE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PACE combines an anisotropic metric with neural bridges for single-cell trajectories and reports clear reconstruction gains, but the geometry part needs tighter validation.

read the letter

The main thing to know about this PACE paper is that it adds an anisotropic Riemannian metric to a neural bridge framework for single-cell trajectory inference and reports average improvements of 23.7% on reconstruction distances. What stands out as new is the three-component pipeline. They construct a state- and time-dependent metric that favors transport along local tangent directions. Then they alternate between refining couplings under the path cost and fitting neural bridges that preserve endpoints. Finally they distill the dynamics into a global velocity field. This setup aims to handle asynchronous development better than standard Euclidean optimal transport. The paper does well on the empirical side. It tests across seven datasets with nine held-out experiments, beats the strongest baselines on MMD, Wasserstein-1, and Wasserstein-2, and shows gains in RNA-velocity alignment on a differentiation benchmark. No explicit cell pairing or supervision is needed, and the code is released. The softer part is the metric construction step. The performance edge depends on that metric actually recovering the right tangent directions from the snapshots. If local estimation is noisy or off, the method falls back toward Euclidean behavior and the gains might not come from the geometry awareness. The abstract lacks the explicit formula or any ablation that isolates this component, so the attribution remains unclear. This paper is for computational biologists working on trajectory inference from time-course single-cell data. A reader interested in optimal transport applications in biology would find the specific technical choices and the benchmark results worth looking at. I would send it to peer review. The core approach is coherent and the results are competitive, though the methods would benefit from more detail on the metric and controls to let referees verify the claims.

Referee Report

2 major / 2 minor

Summary. The paper proposes PACE, a trajectory inference method for single-cell data from destructive time-course snapshots. It constructs a state- and time-dependent anisotropic Riemannian metric that favors transport along locally supported tangent directions, alternates between optimizing cross-time couplings under the induced path cost and fitting endpoint-preserving neural bridges, and distills the dynamics into a global continuous-time velocity field. On seven controlled and biological datasets with nine held-out reconstruction experiments, it reports a 23.7% average reduction in MMD, Wasserstein-1, and Wasserstein-2 distances relative to the strongest baseline, plus a 15.4% improvement in RNA-velocity alignment on an embryoid-body benchmark, without requiring cell pairing or velocity supervision.

Significance. If the central geometry-consistency claim holds, PACE would address a key limitation of Euclidean OT and flow-based methods in asynchronous developmental settings by penalizing normal velocity components. The open-source code at the provided GitHub link is a clear strength for reproducibility and further testing. The empirical gains on held-out reconstruction tasks suggest practical utility, but only if the performance can be attributed to the Riemannian metric rather than the neural-bridge or distillation components alone.

major comments (2)

[Abstract and §3] Abstract and §3 (metric construction): the headline 23.7% average improvement is load-bearing for the geometry-awareness claim, yet the manuscript provides no explicit formula, algorithm, or pseudocode for estimating local tangent directions from snapshots (e.g., via local PCA or similar). Without this, it is impossible to verify that the anisotropic metric reliably penalizes normal components rather than reverting to near-Euclidean behavior under noise.
[§4 and Table 2] §4 (experiments) and Table 2: the nine held-out reconstruction experiments report average percentage reductions but omit per-experiment error bars, standard deviations across random seeds, and an ablation that isolates the Riemannian metric from the neural-bridge fitting and distillation steps. This prevents assessment of whether the reported superiority is robust or driven by the geometry component.

minor comments (2)

[§2] Notation for the path-action cost and the time-dependent metric tensor is introduced without a clear summary table relating symbols to their definitions, which would aid readability.
[§4.3] The RNA-velocity alignment experiment on the embryoid-body benchmark is described only in the abstract; a dedicated subsection with the exact alignment metric and baseline details would strengthen the biological validation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments on our manuscript. We address each of the major comments below and have revised the manuscript to incorporate the suggested improvements for greater clarity and rigor.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (metric construction): the headline 23.7% average improvement is load-bearing for the geometry-awareness claim, yet the manuscript provides no explicit formula, algorithm, or pseudocode for estimating local tangent directions from snapshots (e.g., via local PCA or similar). Without this, it is impossible to verify that the anisotropic metric reliably penalizes normal components rather than reverting to near-Euclidean behavior under noise.

Authors: We appreciate this observation. Section 3 describes the use of local PCA on cell neighborhoods within each snapshot to approximate tangent directions for the state- and time-dependent anisotropic metric. To improve verifiability and address the concern directly, we have added explicit pseudocode and the precise mathematical formula for the metric tensor construction in a new algorithm box in the revised Section 3. This addition clarifies how normal components are penalized and allows readers to assess behavior under noise. revision: yes
Referee: [§4 and Table 2] §4 (experiments) and Table 2: the nine held-out reconstruction experiments report average percentage reductions but omit per-experiment error bars, standard deviations across random seeds, and an ablation that isolates the Riemannian metric from the neural-bridge fitting and distillation steps. This prevents assessment of whether the reported superiority is robust or driven by the geometry component.

Authors: We agree that these details are important for assessing robustness. In the revised manuscript, we have updated Table 2 to report per-experiment means with error bars and standard deviations across five random seeds. We have also added an ablation study in Section 4.3 (with corresponding results in the supplement) that isolates the Riemannian metric by comparing full PACE against variants that disable the anisotropic component while retaining the neural-bridge fitting and distillation steps. The ablation confirms the geometry component drives a substantial portion of the gains. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The provided abstract and description outline a three-component framework: construction of a state- and time-dependent anisotropic Riemannian metric, alternation between cross-time couplings and neural bridge fitting, and distillation to a global velocity field. Performance is evaluated empirically via MMD, Wasserstein-1, and Wasserstein-2 distances against external baselines on held-out experiments, with no equations or self-referential definitions that reduce claimed gains to inputs by construction. No load-bearing self-citations, fitted parameters renamed as predictions, or ansatzes smuggled via prior work appear in the text. The chain remains self-contained with independent empirical support.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the ability to define and optimize under a custom geometry that is not derived from first principles in the abstract; neural bridge fitting introduces many implicit parameters whose values are learned from data.

free parameters (2)

neural bridge network parameters
Endpoint-preserving neural bridges are fitted to data and therefore contain numerous learned weights and biases.
metric anisotropy parameters
The state- and time-dependent anisotropic Riemannian metric requires choices or fitting of local tangent directions and penalty weights.

axioms (1)

domain assumption A state- and time-dependent anisotropic Riemannian metric exists that correctly captures locally supported tangent directions for cellular transport.
Invoked in the first component of the framework as the basis for path-action cost.

pith-pipeline@v0.9.0 · 5804 in / 1401 out tokens · 51666 ms · 2026-05-20T01:20:07.309358+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

PACE constructs a state- and time-dependent anisotropic Riemannian metric that assigns low transport cost along locally supported tangent directions while penalizing normal velocity components... Local PCA... Gaussian-kernel weighted covariance... normal projector P_N^{(r)}
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The metric turns the qualitative principle of geometry-aligned cellular motion into an explicit action functional... v⊤ G v = ||v_T||² + (1+α)||v_N||²

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.