acceptodds
Under review as a conference paper at ICLR 2027

Generating Cellular Spatial Organization with Denoising Papangelou Score Matching

Abstract

Biological tissues are organized not only by which cell types are present, but also by how they form cellular neighborhoods, compartments, and interfaces. Generative models of spatial tissues should therefore capture not only cell counts and composition, but also spatial organization. We introduce Denoising Papangelou Score Matching (DPSM), a framework for generating cellular spatial configurations consisting of cell locations and types. DPSM uses time dependent Papangelou fields to model where cells of different types are likely to be inserted given the current configuration. Learned along a corruption path to a Poisson reference, these fields determine the reverse generative processes. We show that noisy pseudolikelihood performs population field matching, derive a denoising identity for the corrupted fields, and relate field estimation error to reverse generation error. Experiments on controlled interacting point processes, synthetic cellular patterns, and real world spatial omics data show that DPSM recovers local cell type associations and spatial organization while preserving cell counts and composition.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial omics, generative model, score matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dKDLOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduced a sophicasticated framework named Denoising Papangelou Score Matching (DPSM) for generating spatial patterns of multiple elements. The authors presented the mathematical development of the method, and used a series of geometrical patterns to illustrated the properties of the proposed framework. They also claimed that experiments were done on real biological data and showed some benchmarking metrics, but they didn't really present what were the biological data and what were the generated results. The proposed framework may have some potential in certain spatial generation tasks. But the biological scenario is poorly defined and the claims and conclusions on the biology side of the story are unjustified. Given that the title, abstract, introduction and discussion are all toward a paper that aimed at solving biological tasks, the submission cannot be a strong candidate for publication.

Critical strengths

The mathematical framework seems to be strong with the series of theorems proving its properties.

Critical weaknesses

The link of the mathematical framework will real biological situations is weak or absent. The method may be used for generating toy simulation data, but its use on real biological problems is absent.

Questions for authors

What is the definition of the biological task of "generating cellular spatial organization"? What is the use or potential use of such generation? What are the assumptions you need to make on the biological data to justify the principle of DPSM, and how well they are met in different biological organs/tissues? How do you evaluate the generated results in the biological sense? What are the biological significances of the metrics used in the experiments?

Reviewer_63yeOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies the generation of marked cellular point configurations, where the aim is to reproduce not only cell counts and types but also their spatial arrangement. DPSM corrupts observed configurations by thinning and Poisson superposition, then learns time-dependent Papangelou conditional intensities with noisy pseudolikelihood. The same learned field determines insertion and removal rates in reverse-time generation. The authors establish population-risk and denoising identities and, under stated assumptions, relate field estimation error to generation error. Experiments on interacting point processes, synthetic tissue patterns, and patient-held-out spatial omics datasets evaluate the recovery of spatial interactions and organization.

Critical strengths
  1. The paper tackles an important problem that cannot be reduced to matching cell counts or marginal spatial density. The synthetic experiments are particularly useful here: they test whether models recover coherent mark organization even when simpler spatial summaries are similar. This makes the motivation concrete rather than merely intuitive.
  2. The main methodological contribution is the connection between Papangelou field estimation and reverse-time generation. A single time-dependent field determines both insertion and removal rates, while the theoretical analysis links the learning objective to field estimation and generation error. This gives the proposed framework a clear probabilistic justification. The experimental evaluation is also well structured, with complementary studies on interaction recovery, synthetic spatial organization, and real biological data. Together with the interaction ablation and direct MCMC comparisons, these experiments provide convincing support for the main claims.
Critical weaknesses
  1. Although the interaction ablation in Appendix G.4 demonstrates the importance of configuration-dependent interactions, it does not distinguish the contributions of the global and local energy terms. A more targeted ablation would help determine how much each component contributes to the final generation quality.
  2. The real-data evaluation relies mainly on population-level spatial statistics. While these metrics are appropriate and the reported results are strong, it remains unclear how well the generated configurations reproduce specific cellular neighborhoods and tissue interfaces. A more focused analysis on one biological dataset would help address this question.
Questions for authors
  1. How much of the improvement comes from the global configuration energy versus the local neural energy? Can the authors provide a focused comparison to clarify their respective contributions?
  2. Can the authors show whether the generated samples preserve characteristic cell-type neighborhoods or tissue interfaces in one of the real datasets, beyond the aggregate statistics already reported?
Reviewer_LWaZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This submission introduces Denoising Papangelou Score Matching (DPSM) approach that generates cellular spatial configurations with cell locations and cell types by learning time-dependennt Papangeleou fields using noisy negative psuedolikelihood.

Critical strengths

Authors evaluate DPSM through four analyses. This includes testing whether DPSM recovers known interactions and spatial statistics. The authors also use synthetic cellular patterns to isolate configuration-level organization from marginal geometry and mark frequencies. For real-world evaluation, authors use patient-held-out spatial omics datasets to test whether generated windows reproduce spatial statistics beyond training patients. FOr model analysis, authors compare DPSM with MCMC under the same fitted energy examine the role of learned noise path.

Critical weaknesses
  • The paper is hard to read and follow, and the algorithm is unclear. It is unclear what model hyperparameters are, how they are set, what training versus inference efficiency is.

  • The analysis of the learned noise path is useful, but the paper has no ablations of the other design choices.

  • Evaluations compare DPSM with MCMC under the same fitted energy but do not include recent spatial generative tissue models, such as Steamboat, SPATIA, CONCERT, etc. (1) SPATIA multimodal model integrates cell morphology, gene expression and spatial context at the cell, niche and tissue levels. It includes a spatially conditioned flow matching generator that predicts changes in cell morphology across tissue transitions. (2) Steamboat decomposes each cell's profile into intrinsic, local and sample wide components and supports in silico transplantation and knockout experiments. (3) Celcomen is a generative model that separates intracellular from intercellular gene regulation and generates counterfactual spatial transcriptomics. (4) CONCERT is a niche aware generative model that learns spatial kernels with a Gaussian process variational autoencoder. It generates counterfactual spatial transcriptomes, for example after relocating or resizing a perturbed patch of cells, at unmeasured time points, or in missing tissue regions. (5) Many existing studies also study dissociated single-cell and perturbation models (scGEN, Biolord and GEARS) and a spatial kNN baseline.

  • Authors should cite, discuss and evaluate their models in the context of these existing lines of research. These models generate expression or morphology for cells at given positions, while DPSM generates the positions and types themselves. A comparison is still feasible on the cell type component conditioned on observed locations, or by deriving cell types from generated expression.

Questions for authors

Please address pointed raised as critical weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.