acceptodds
Under review as a conference paper at ICLR 2027

Physics-Guided Graph Networks for Gene Expression Recovery in Spatially Overlapping Cells

Abstract

Overlapping segmented cell footprints can mix gene expression and obscure source profiles in Stereo-cell sequencing. We form paired mixtures by aligning two measured PBMC chips and translating one chip map relative to the other; directional polygon overlap sets donor contributions, while true overlap edges and source profiles remain hidden at inference. We present (Physics-Guided Graph Recovery), a four-block network with a learned circle-overlap surrogate, residual backprojection, adaptive increment modulation, and graph-context correction. Identity-disjoint evaluation uses three technically matched chips. On the fixed test split, achieves the lowest Count-MAE, Log-MAE, and Rel. among evaluated methods; its Log-MAE is versus for GATv2 ( lower, based on three-seed means). Under strict receiver leave-one-chip-out evaluation within this chip series, it again has the lowest three errors, with Log-MAE of versus for GraphSAGE. Component removals support all three modules within the evaluated stack. These results support source recovery under controlled geometric crosstalk.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-guided machine learning, graph neural networks, inverse problems, gene expression demixing, crosstalk correction, spatial transcriptomics, single-cell transcriptomics.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8keyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Authors present PhyGRec, a new method for deconvolving gene expression data into individual cells by incorporating spatial data and refinements over multiple iteration. Contributions include a new architecture that includes iterative refinement through backprojection and the authors demonstrated greater performance with respect to SOTA methods that use graph attention and similar techniques

Critical strengths

Strengths include a well-calibrated architecture design

  • The overall problem is highly impactful, spatial data is very expensive to generate and still exists only at a limited scale, which is relevant for several disease classes such as cancer micro-environment
  • A more advanced architecture for deconvolving expression data
  • Performance shows clear improvements in deconvolutions
Critical weaknesses

Overall, the paper is confusing and hard to follow for a human. The introduction does not properly introduce the problem

  • It is not clear for the average reader what stereo-seq is, not what a stereo cell is
  • Problem and overall universe are not clearly defined
  • For example, “measured chip’s” not clear what chip is referring to? Chip-seq maybe?
  • Past contributions and novelty are not properly flowing in the introduction
  •  Ablations in table 3 benchmark show modest contributions of the physics addition
    
Questions for authors
  • The contribution here is clearly on the domain application. Would it be possible to expand to other cell types?
  • Would it be possible to test on histological section data from spatial RNA-seq/proteomic data?
1 comment

Thank you for your careful review and constructive suggestions. We have revised the Introduction to clarify the task and contribution, explained the component ablations, and expanded the discussion of future validation. We address the weaknesses and questions below.

W1. The Introduction does not clearly establish Stereo-seq, Stereo-cell, chip, or the recovery problem.

“It is not clear for the average reader what stereo-seq is, not what a stereo cell is” ““measured chip’s” not clear what chip is referring to? Chip-seq maybe?”

A1. The submitted Introduction cites Chen et al. (2022) and Wei et al. (2022) for Stereo-seq's DNA-nanoball arrays, and Liao et al. (2025) for Stereo-cell profiling of dissociated cells.

We have revised the opening of Section 1 to define a capture chip as the physical array, a footprint as a segmented cell region, and Reference as the receiver's measured profile before controlled mixing. We also clarify directional donor contributions and the model inputs: mixed expression and coordinate-derived candidates, with source profiles and transfer coefficients hidden.

W2. The prior contributions and novelty do not flow clearly in the Introduction.

“Past contributions and novelty are not properly flowing in the introduction”

A2. SoupX and DecontX address ambient RNA; SpotClean addresses spot swapping. Our task recovers cell-specific profiles from directional footprint mixing with hidden overlap edges and coefficients. PhyGRec couples a learned mixing surrogate with residual backprojection, adaptive increment modulation and graph-context refinement for this recovery task. We have revised Section 1 to make this task-to-method connection explicit and added matched graph comparisons and recent inverse-method adaptations in Section 4.6.

W3. The ablations show a modest contribution from the physics component.

“Ablations in table 3 benchmark show modest contributions of the physics addition”

A3. RB uses the circle-overlap surrogate to backproject measurement residuals (Sections 3.3–3.4). Its ablation retains the circle operator and candidate graph and sets the backprojection increment to zero. Removing RB increases mean Log-MAE from 0.00369 to 0.00657. Table 3 reports three-seed mean ± sample SD:

ConfigurationParams (M)Log-MAE ↓
Full2.8330.00369 ± 0.00021
Without RB2.8330.00657 ± 0.00148
Without AIM2.6200.00453 ± 0.00007
Without GCC0.2120.01075 ± 0.00013

Parameter counts include all registered trainable parameters in each evaluated architecture. AIM modulates the RB increment through learned gain and momentum, and GCC adds a graph-context correction. Removing AIM disables both gain and momentum; AIM and GCC removals also reduce the parameter count. All nine paired-seed Log-MAE comparisons favor the full model. We have clarified the RB removal and the AIM/GCC ablations in Section 4.5. Table 21 separately evaluates the circle coefficients against the generator.

Q1. Can the evaluation extend to other cell types?

“Would it be possible to expand to other cell types?”

A1. The recovery benchmark covers the cell populations represented in dissociated PBMCs. We have revised Section 5 to clarify this scope and discuss validation on other cell types with independent references as future work.

Q2. Can the method be tested on histological sections or proteomic data?

“Would it be possible to test on histological section data from spatial RNA-seq/proteomic data?”

A2. We have expanded Section 5 to discuss histological sections and spatial proteomics as future research directions, using assay-specific mixing models and independent references for evaluation.

Reviewer_KaC3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper recover gene expression under spatial crosstalk in Stereo-cell data. The authors map segmented footprints on to three technically match PBMC chips and pair the donor with the extent of overlap between the pixels of the PBMC chips. Four blocks of PhyGRec are used to recover gene expression in this situation: a learned circle-overlap operator, a residual backprojection operator, a modulation of increments, and a graph context correction.

Critical strengths

The work targets a practically important and often neglected error mode in single-cell-resolution spatial assays: mixing from overlapping segmented footprints. By combining physics-guided modeling with graph-based refinement, the approach aligns with broader trends in physics-aware deep learning and algorithm unrolling, potentially informing future methods in spatial transcriptomics and other inverse problems.

The model integrates an explicit, parametrized forward operator with residual backprojection and graph-neural refinement in a multi-block, unrolling-inspired architecture.

The data-splitting protocol addresses an important source of potential leakage. Cell identities are assigned to splits before scene generation, and translated appearances remain within their assigned split. This is particularly important because the same measured source profile can occur in multiple synthetic scenes.

The benchmark is carefully designed: true overlap edges and coefficients are hidden at inference; candidate donors are reconstructed from coordinates, and identity-based splits ensure no leakage across scenes.

Critical weaknesses

All experiments use three technically identical PBMC chips from the same series. Despite receiver-LOCO, the number of assay conditions and tissues in each experiment is small.

Beyond the GNN and VAE baselines, a comparison of PhyGRec with recent physics-informed or unrolling-style methods in related modes would contextualize the architectural choices.

Questions for authors

How much of PhyGRec's advantage is still there when GATv2 or MPNN have the same candidate distances and validity information, with comparable capacity and validation tuning?

How are the robust receiver weights Wb calculated in RB and how sensitive is the result to their hyperparameters or function?

Why do you consider radii based on expression totals in X instead of geometric priors? What are the cases where you found this choice to be superior to area based radii?

In Appendix D.1, the post-GCC data consistency step is set at 0.1, but Equation (10) and Algorithm 1 show that there is only one expected update after addition of RB, AIM, and GCC. Was an additional step used in Tables 1-3?

1 comment

Thank you for your careful review and constructive suggestions. We respond to your questions below.

Q1. How much advantage remains with matched information and capacity?

“the same candidate distances and validity information, with comparable capacity and validation tuning?”

A1. We tuned learning rate, weight decay and dropout for both GATv2/MPNN arms, plus GATv2 scheduling, on validation (Appendix E.4). Runs use 100 epochs and 3 seeds. Checkpoints maximize combined recovery: the mean relative validation Log-MAE/Rel. L1 reduction from Mixed (Appendix D.1). Arms within each family share features, topology, architecture, parameter count and supervision; edge attributes add distance and validity. Table 4: tuned baselines and original PhyGRec (mean ± sample SD):

ModelParams (M)Log-MAE ↓
GATv2 topology-only2.840.009207 ± 0.000130
GATv2 edge-attributed2.840.008617 ± 0.000536
MPNN topology-only2.830.017858 ± 0.001741
MPNN edge-attributed2.830.010588 ± 0.000765
PhyGRec2.830.003687 ± 0.000214

Original PhyGRec has 57.2% and 65.2% lower mean Log-MAE than tuned edge-attributed GATv2/MPNN. Table 1's 6.777M MPNN uses a fixed distance kernel.

Q2. How does PhyGRec compare with recent physics-informed or inverse methods?

“recent physics-informed or unrolling-style methods”

A2. Table 5 adds six adaptations from five recent papers on the fixed test, sharing training-fitted circle parameters per seed (Log-MAE, three-seed mean ± SD):

MethodLog-MAE ↓
Physical PGD0.030560 ± 0.000635
DeMUN0.013876 ± 0.001646
Learned RESESOP0.017405 ± 0.000995
ProxUnroll-HQS0.022127 ± 0.001318
ProxUnroll-ADMM0.023791 ± 0.001417
RAM0.432651 ± 0.011705
Vo–Tachella0.559068 ± 0.007873
PhyGRec0.003687 ± 0.000214

Appendix E.4/Table 20: DeMUN: official memory/DnCNN; RESESOP: Algorithm 2 with receiver-row suboperators; ProxUnroll: official HQS/ADMM networks and stage memory; RAM: official operator conditioning, 32-gene groups; Vo–Tachella: exact-operator partitioning, PGD/DRUNet (one stage, patch32/stride16/grid40). Losses and capacities differ. DeMUN, RESESOP and ProxUnroll use a training-only small-positive penalty; test scoring uses no extra output threshold. PGD SD: 3 fitted operators. Mixed Log-MAE: 0.05117 (Table 1).

Q3. How is WbW_b computed, and how sensitive is it?

“How are the robust receiver weights Wb calculated in RB and how sensitive is the result to their hyperparameters or function?”

A3. For receiver mm, r=∥Rb[m]∥1/max⁡(∥X[m]∥1,1)r=\|R_b[m]\|_1/\max(\|X[m]\|_1,1). The gene-shared weight is wm=(1+(r/σb)2/∣νb−2∣)νb/2−1w_m=(1+(r/\sigma_b)^2/|\nu_b-2|)^{\nu_b/2-1}, with Wb=diag⁡(wm)W_b=\operatorname{diag}(w_m). Learned σb>0\sigma_b>0 and νb<2\nu_b<2 start at 11; σb\sigma_b and ∣νb−2∣|\nu_b-2| are floored at 10−410^{-4}. Table 22 tests uniform weights and a joint scale/shape reset on frozen checkpoints:

WeightsLog-MAE ↓Count-MAE ↓
Learned Barron0.0036870.001327
Uniform0.0050500.001479
Initial Barron (scale=shape=1)0.0036640.001451

Uniform weights raise mean Log-MAE 37.0% over learned weights. Initial Barron gives slightly lower Log-MAE but higher Count-MAE.

Q4. Why use expression totals, and when do they outperform area?

“radii based on expression totals in X instead of geometric priors” “superior to area based radii”

A4. Library size uses fixed X totals; area supplies a segmentation scalar. Table 23 compares the same radius mapping with matched-seed retraining:

Radius inputLog-MAE ↓Block-4 positive-edge coefficient MAE ↓
Library size0.0035760.2174
Constant0.0039080.2071
Segmentation area0.0039610.1906

Library size gives 9.7% lower mean Log-MAE than area in this PBMC protocol (2/3 seeds); area gives lower positive-edge coefficient MAE. These are separate retrainings. Section 3.3 clarifies packed-batch medians.

Q5. Was an additional post-GCC step used?

“Was an additional step used in Tables 1-3?”

A5. Runs use one projected update per block (Eq. (10), Algorithm 1). Appendix D.1 gives original learning rates and 0.1 as the graph-correction strength initialization. Tables 1–3 use original validation-selected checkpoints.

Q6. What do the same-series experiments establish?

“All experiments use three technically identical PBMC chips from the same series.”

A6. Receiver-LOCO holds out receivers within the matched PBMC series. Fig. 4(b) shows Y40628PA12 expression structure using a separate frozen 27,162-gene checkpoint (Appendix F.2). Section 5 discusses future independent-reference validation on natural overlaps, other populations/chip series, histological sections and spatial proteomics.

Reviewer_wQpoOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhyGRec, a physics-guided graph network for recovering gene expression from spatially overlapping cells. The proposed method combines a learned circle-overlap operator with residual backprojection, adaptive increment modulation, and graph-context correction to reconstruct cell-specific gene expression from mixed observations.

Critical strengths
  1. The paper addresses an important problem in spatial transcriptomics by incorporating physical mixing information into graph-based gene expression recovery.

  2. It provides extensive experimental comparisons and ablation studies, demonstrating substantial improvements in gene expression reconstruction accuracy over GNN and physics-based baselines. The leave-one-chip-out evaluation further supports the effectiveness of the proposed method.

Critical weaknesses
  1. The proposed circle-overlap operator relies on heuristic assumptions about RNA transfer between cells. It remains unclear whether the learned overlap coefficients accurately reflect the underlying mixing process. Since the benchmark provides ground-truth polygon-overlap coefficients, directly comparing them with the learned coefficients would help validate the proposed physical model.

  2. The experimental validation is limited to artificially generated mixtures. Although the experiments use measured gene expression profiles from three PBMC chips, the overlapping mixtures are generated through controlled geometric transformations. Moreover, all three chips belong to the same technical series. Therefore, it remains unclear whether the proposed method generalizes to naturally occurring RNA contamination or different biological conditions. Additional evaluations on real overlapping-cell data and more diverse datasets would strengthen the experimental validation.

Questions for authors

Please see the concerns detailed in the Weaknesses.

1 comment

Thank you for your careful review and constructive feedback. We have added a direct comparison of the learned and polygon overlap coefficients and clarified the evaluation scope and future validation directions. We address your two concerns below.

W1. Do the learned overlap coefficients agree with the polygon coefficients?

“Since the benchmark provides ground-truth polygon-overlap coefficients, directly comparing them with the learned coefficients would help validate the proposed physical model.”

A1. We compared all four blocks of three validation-selected main checkpoints on all 363 fixed-test scenes, using polygon coefficients for scoring. Table 21 reports mean ± sample SD:

OperatorPositive-edge coefficient MAE ↓All-edge coefficient MAE ↓
Block 10.2100 ± 0.04600.0814 ± 0.0562
Block 20.1660 ± 0.00520.0380 ± 0.0072
Block 30.1518 ± 0.00920.0360 ± 0.0040
Block 40.1935 ± 0.00210.0306 ± 0.0014
Zero transfer0.24870.0324

Positive-edge MAE scores generator coefficients > 0; all-edge MAE also includes zero-coefficient evaluated edges. Block 3 has the lowest positive-edge MAE, while Block 4 has the lowest all-edge MAE. With predicted coefficients > 0 defining predicted support, Block 4 has 98.68% recall and 42.07% precision. It detects most positive generator edges and assigns positive transfer to some zero-coefficient edges.

We have added this comparison to Section 4.6 and Appendix E.5. The learned coefficients parameterize the circle-overlap surrogate for directional mixing. Table 21 shows that coefficient accuracy and expression recovery differ across blocks: the final supervised block has the lowest Count-MAE, Log-MAE and Rel. L1 among the four blocks. Calibration against naturally occurring RNA transfer is a future direction.

W2. What evidence addresses naturally occurring contamination and different biological conditions?

“Therefore, it remains unclear whether the proposed method generalizes to naturally occurring RNA contamination or different biological conditions.”

A2. Our quantitative evaluation covers controlled mixtures and receiver-LOCO within the matched PBMC series, using measured pre-mixing targets. Under receiver-LOCO, Log-MAE is 0.00801 ± 0.00065 for PhyGRec and 0.01305 ± 0.00012 for GraphSAGE (Table 2).

Submitted Fig. 4(b) shows measured input and output on unseen Y40176KE (8,821 cells). We have updated the panel to Y40628PA12 (28,685 cells) using a separate frozen 27,162-gene checkpoint. In Appendix F.2, we clarify that these panels explore PBMC expression structure and that the revised display settings were selected after inference. In Section 5, we have expanded the discussion of future work to cover independent-reference validation on naturally overlapping cells, other populations and chip series, and extensions to histological sections and spatial proteomics.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by authors

We thank the Area Chair and reviewers for their careful reading and specific suggestions on the physical model, baseline comparisons, and presentation. PhyGRec recovers cell-specific expression from directional footprint mixtures using mixed expression and coordinate-derived donor candidates. We have expanded the experiments and revised the manuscript around four shared concerns:

  • Matched graph comparisons (Table 4; Appendix E.4). We tuned GATv2 and MPNN with matched candidate information and comparable parameter counts using validation-based selection. The original PhyGRec checkpoints achieve 57.2% and 65.2% lower mean Log-MAE than the tuned edge-attributed GATv2 and MPNN, respectively. Table 4 reports all eight metrics as three-seed means and sample standard deviations.
  • Recent physics-informed and unrolling methods (Table 5; Table 20). We added six task adaptations from five recent papers: DeMUN, Learned RESESOP, ProxUnroll-HQS/ADMM, RAM, and Vo–Tachella. Appendix E.4 describes the original code components and task-specific adaptations; Table 5 reports their fixed-test results alongside Physical PGD and PhyGRec.
  • Overlap-operator analysis (Table 21; Appendix E.5). We compared the learned circle-overlap coefficients with the polygon coefficients on all 363 fixed-test scenes, using three original validation-selected checkpoints. The analysis reports positive-edge and all-edge coefficient errors, support recall and precision, and expression recovery across the four blocks.
  • Weight and radius sensitivity (Tables 22–23). With checkpoints frozen, uniform receiver weights increase mean Log-MAE by 37.0% relative to learned weights. In matched-seed radius retraining, library size gives 9.7% lower mean Log-MAE than segmentation area in the controlled PBMC protocol, with lower error in two of three seeds. The tables also report the initial Barron-weight and constant-radius comparisons.

We have revised Section 1 to introduce Stereo-seq, Stereo-cell, capture chips, footprints, and the recovery target together. Section 3.3 and Appendix D.1 clarify the radius normalization and original execution settings; Section 4.5 explains the component removals. The measured-chip illustration is updated in Figure 4(b) and detailed in Appendix F.2. We have expanded Section 5 to discuss the evaluation scope and future validation on naturally overlapping cells, other populations and chip series, histological sections, and spatial proteomics with independent references.

The new experiments, analyses, and revised text are marked in blue in the revised manuscript. We thank the reviewers and Area Chair again for their time and welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.