acceptodds
Under review as a conference paper at ICLR 2027

Self-Supervised Graph-Based Refinement of Spatial Transcriptomics at Subspot Resolution

Abstract

Spatial transcriptomics enables the study of gene expression within tissue context, but current sequencing-based technologies operate at a spatial resolution where each measurement aggregates signals from multiple cells. While deconvolution methods aim to estimate each spot's cell-type composition, they do not provide fine-grained spatial structure. We propose \methodname, a self-supervised framework based on graph attention networks to infer plausible subspot-level gene expression patterns from spatial transcriptomics data. The method leverages spatial neighborhood structure through message passing and introduces virtual nodes that represent latent subspot locations, enabling the redistribution of observed transcriptomic signals into finer spatial units. Importantly, our approach operates without requiring external references such as single-cell RNA-seq data or histological images. Rather than recovering true cellular-resolution measurements, Square aims to infer spatially consistent fine-grained structure that is compatible with the observed data and underlying spatial dependencies. Across synthetic, ST, Visium and Visium HD samples, we show that the inferred subspots exhibit improved spatial coherence, reduced mixing of cell-type signals, and enhanced alignment with known biological organization, while mitigating sparsity. These results suggest that graph-based modeling of spatial context can effectively refine coarse spatial transcriptomics data, providing a principled approach to studying fine-scale tissue organization when direct cellular-resolution measurements are unavailable.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Spatial transcriptomics, distribution shift, graph, attention, entropy.

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_J9TFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Square, a self-supervised framework for inferring subspot-level spatial transcriptomic profiles without requiring histology images or single-cell reference data as model inputs. Training examples are constructed by pooling neighboring spots into super-spots and using the original spots as reconstruction targets. A graph attention network introduces virtual nodes representing latent subspot locations, while a jointly trained variational autoencoder is intended to mitigate sparsity and the distribution shift between training and inference resolutions.

Critical strengths

A useful self-supervised formulation with limited input requirements. Learning from artificially pooled measurements provides supervision without requiring externally paired high-resolution data. Combining this procedure with virtual-node graph prediction offers a practical approach to spatial refinement when additional modalities are unavailable.

Critical weaknesses
  1. The paper does not sufficiently explain why the proposed method can address subspot-level spatial inference. Although the architecture is described, the underlying assumptions and the mechanisms through which virtual nodes, graph attention, and latent representations enable meaningful fine-grained predictions remain insufficiently justified.

  2. The motivation and effectiveness of the training objective are not adequately explained. In particular, the paper does not clearly establish why the reconstruction, enhancement, and KL regularization terms are combined in the proposed manner, how their roles complement each other, and why optimizing this objective should produce reliable subspot-level representations.

Questions for authors
  1. What assumptions about gene expression and spatial organization allow the proposed framework to infer meaningful within-spot structure from aggregated measurements? How do these assumptions guide the model toward biologically plausible refinements?

  2. Why should learning to reconstruct original spots from pooled super-spots generalize to decomposing original spots into unobserved subspots? Could you explain the conditions under which this transfer across resolutions is expected to work?

  3. What are the distinct roles of the reconstruction and enhancement likelihood terms in Eq. (5)? Please clarify their respective inputs and targets, justify the likelihood models, and explain why both terms are necessary.

  4. How does KL regularization help mitigate the training–inference distribution shift while preserving biological differences? Could you provide evidence that its benefit goes beyond generic latent-space smoothing?

  5. How are the relative contributions of the three objective terms balanced? Can controlled loss-term ablations demonstrate that their combination improves fine-grained prediction, rather than merely reconstruction accuracy or spatial smoothness?

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

The authors of this work address the inference of high-resolution gene expression data from coarse-grained spatial transcriptomics data, where each measurement captures the gene expression of multiple cells. To this end, they propose a graph-based learning model Square, which constructs a small graph for each measurement (so-called spots) connected to its neighboring spots. They then add additional virtual nodes that represent the subspots for which they seek to interpolate gene expression. They use a graph attention network (GAT) that is trained in a self-supervised fashion, by coarse-graining groups of real spots into super-spots and then learning to recover the original fine-grained expressions. A variational autoencoder is used to summarize gene profiles into shared representations and the GAT model is used to predice gene expressions of sub-spots based on the neighbors. The method is qualitatively evaluated in real data, where no ground truth is available, by checking the consistency of results with known biological mechanisms.

Critical strengths
  • The authors propose a potentially interesting method for a practical problem that can be applied to gene expression data and measurement spot coordinates alone, and does not require external references. This makes it suitable for scenarios where no single-cell spatial transcriptomics data is available.

  • The approach to train the model in a self-supervised setting by coarse-graining the available data and then learning to recover the original resolution is interesting and potentially meaningful.

  • Results on simulated data show good performance compared to three baseline methods.

  • The authors perform a qualitative evaluation that is grounded in knowledge about biological mechanisms.

  • The authors honestly discuss the limitations of the approach, arguing that the method relies on spatial dependencies in the graph structure, which might break in tissue regions with highly heterogeneous cell composition.

Critical weaknesses
  • My biggest concern is that the experimental evaluation of the method only partially supports the authors claims. In particular, no ground truth was used for the evaluation (even though this might be available in one of the data sets), the evaluation in real-world data is based on a self-referential argument, I have questions regarding the choice of baselines, and there are issues in the hyperparameter settings that question the fairness of the comparison, see my questions Q1 - Q4 below.

  • Another concern is that - considering the self-supervised training - the method seems to assume a scale-invariance in the patterns that may not be justified, especially as this depends on the resolution of the measurement and the specific size of biological structures in the tissue that it is applied to. See my question Q5 below.

  • Related to the previous issue, another strong assumption of the method is that the spatial autocorrelation is sufficiently strong for neighboring spots to provide meaningful supervision signal for the training, which is bound to fail in highly heterogeneous tissue where the tssue structure contains local niches with specific compositions of cell types. While the authors are honet about this limitation, this is arguably the most interesting practical setting, which limits the relevance and practical appeal of the approach.

  • I finally had some issues to understand specific details of the methods and found some inconsistencies in the reported results that require clarification, see Q6, Q7 and Q8 below.

Questions for authors

Q1: While I acknowledge that it is not easy to evaluate the performance of the method against ground truth given the available data, I still think more effort could have been made to support the claims in the paper. First of all a quick check seems to suggest that one of the data sets (Mouse Small Intestine) contains high-resolution data, while the evaluation has only been performend on a coarse-grained version of the data set. This seems like a missed opportunity. Could the authors explain this?

Q2: I think the result that the proposed method yields higher spatial coherence in the glioblastoma data set, which is taken as evidence that it produces biologically meaningful results, is somewhat self-referential as this seems to simply be grounded in how the method interpolates gene expression based on neighbors. Also, for this analysis, there is no comparison of biological plausibility with the baseline methods.

Q3: The method is compared to three baselines BayesSpace, STAGE and CARD. However, given the approach to interpolate gene expression based on its neighbors, a comparison to a trivial baseline that simply copies a spot's value to all sub-spots, or a simple averaging across neighbors would have been interesting to evaluate how much the actual learned model contributes to the performance. This experiment is also warranted by the abaltion result, that a model variant without virtual nodes (which seems to be close to my proposal to simply average neighbors) actually achieves comparable performance.

Q4: Regarding the comparison to baselines, I am concern that the comparison is not fair in the sense that the authors write that the hyperparameters of their method have been tuned in the simulated data set, while apparently no hyperparameter tuning has been performed for the baseline. This approach would favor the proposed method. I kindly ask the authors to clarify this important aspect of the evaluation setting.

Q5: Refering to the weakness mentioned above, it seems that the method relies on assumed scale-invariance of patterns that may not be present in real data. Specifically, could the authors clarify why they expect a decomposition that is learned at the super-spot level to transfer to a sub-spot level? Could this assumption be directly tested in some of the data sets, e.g. by training and evaluation across different spatial scales?

Q6: Regading the reported results, it seems that some of the numbers are inconsistent. From what I understood, the results for Square with mean-based neighbor initialization should match between table 1 and table 10, however, they strongly differ without explanation. Could the auhohrs explain this discrepancy? Was this based on a different simulated setting?

Q7: It seems that the hyperparameter values in different tables do not match. In table 8, the author list the final hyperparameter values chosen after the Optuna optimization, which does not seem to match any of the the dropout values in the top 10 trials reported in table 4. Moreover, some of the reported values fal outside the brackets reported for the search grid in table 3, e.g. the optimal dimension ratio was found to be 0.55 when table 3 reports that values 0.3, 0.4, ... 0.8 have been tested. This may all seem like minor details, but it does not really convey trust in the experimental evaluation.

Q8: The parameter k=9 which determines the number of subspots per spot seems rather arbitrarily chosen and further does not adapt to differences in local cell density. Did the authors try different values? How sensitive are the results to this choice?

Reviewer_KP7gOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces a self-supervised graph-based framework called "Square," designed to refine spatial transcriptomics data into higher-resolution, sub-spot-level gene expression representations—without requiring external single-cell reference data or histological images.

Critical strengths
  1. Inferring finer-grained spatial organization from measurements that aggregate multiple cells is a significant challenge. Reference-free methods are particularly attractive when matched single-cell RNA sequencing (scRNA-seq) data or histological images are unavailable.

  2. By aggregating neighboring spots to construct training samples, the framework learns spatial refinement without the need for experimentally measured sub-spot-level ground truth. The use of "virtual nodes" allows the model to explicitly represent latent positions within each original spot.

  3. Beyond synthetic benchmarks with known ground-truth structures, the experiments examine cell-type composition in mouse olfactory bulb data, as well as biologically relevant spatial patterns in glioblastoma and mouse intestinal tissue. These analyses provide compelling evidence for the potential applications of this method.

Critical weaknesses
  1. During training, Square aggregates nine original spots into a "super-spot" and learns to reconstruct their individual expression profiles. However, during inference, the model decomposes each original spot into nine smaller sub-spots. The resulting discrepancies in transcript abundance, sparsity, spatial scale, and neighborhood composition can be substantial. Although the authors employ a variational autoencoder (VAE) to mitigate this shift, the ablation studies do not definitively prove that the learned decomposition remains reliable when extrapolated to a higher resolution than that used during training. A comparative evaluation using held-out, high-resolution experimental data would provide a more robust validation of this hypothesis.
  2. While the simulated benchmarks offer quantitative assessments against known ground truths, the analyses of glioblastoma and mouse intestinal tissue primarily demonstrate spatial consistency, cell-type purity, and alignment with established biological patterns. While these observations are encouraging, the presence of biologically plausible or smoother spatial patterns does not necessarily guarantee the accurate reconstruction of the underlying high-resolution gene expression. In particular, the Visium HD experiment begins with measurements at 2 µm intervals, which are then aggregated into 8 µm intervals; this provides a natural opportunity to compare the reconstructed expression with the original high-resolution measurements. A more direct quantitative assessment would significantly strengthen the main argument.
  3. It is not explicitly guaranteed that the observed expression levels are preserved following subspot decomposition. The paper treats the inferred subspots as a partition of the original measurements; however, the training objective in Equation (5)—which combines VAE reconstruction, graph-based prediction, and KL regularization—does not explicitly require that the aggregated expression levels of the inferred subspots recover the observed expression levels of the original spots. This raises the question: does this refinement process preserve the total expression levels and the relative gene abundances present at the original resolution? Through reconstruction consistency analysis or the introduction of explicit aggregation constraints, one could verify whether the inferred expression profiles remain quantitatively consistent with the measured data.
  4. Comparisons with other existing methods are complicated by differences in output resolution and evaluation protocols. The Square method partitions each original spot into 9 subspots, whereas the STAGE method partitions each into 4. The authors address this mismatch for certain metrics by averaging the ground-truth expression profiles of neighboring spots, while omitting the Adjusted Rand Index (ARI) when evaluating STAGE. Although these adjustments are understandable, they hinder a direct comparison of spatial reconstruction quality. Furthermore, the simulation benchmarks used are relatively simplified; thus, comparisons with methods like CARD may not fully capture the differences in their respective original design objectives. Evaluating all methods at a unified spatial resolution would allow for a clearer assessment of their relative performance.
  5. The contributions of the VAE and the "virtual nodes" appear inconsistent across different evaluation metrics. Table 1 shows that removing the virtual nodes reduces the MSE from 0.39 to 0.29; conversely, the full model performs better in terms of ARI and several spatial distribution metrics. The authors correctly note that MSE alone is insufficient to comprehensively measure spatial reconstruction quality. Nevertheless, this trade-off suggests that the proposed architecture may prioritize spatial organizational structure over the quantitative accuracy of expression levels. A more in-depth analysis of the relationship between expression reconstruction error, spatial coherence, and cell-type discriminability would help to provide a more comprehensive understanding of the model’s practical strengths and limitations.
  6. The scope of biological validation is relatively limited, and the results may, to some extent, reflect the spatial priors incorporated into the model. For instance, the analysis of glioblastoma is based on a single tissue section, while the evaluation using Visium HD focuses on a specific region containing three intestinal villi. Given that the model explicitly leverages spatial neighborhood information, the observed improvements in spatial coherence and the recovery of expected gradients may, in part, stem from the smoothness assumptions underlying the model. Further validation using independent samples and heterogeneous tissue regions—particularly those near sharp boundaries or containing rare cell populations—will be essential to determine whether the method truly preserves authentic, fine-scale heterogeneity or merely tends to produce spatially coherent patterns.
Questions for authors

Weaknesses

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

in this paper, authors introduces a new method, a self-supervised method for refining spatial transcriptomics into latent subspots without requiring scRNA-seq or histology as model inputs. It creates training examples by pooling observed spots, then combines a variational autoencoder with graph attention and virtual subspot nodes. Evaluation includes synthetic data, simulated mouse olfactory bulb mixtures, glioblastoma, and Visium HD intestine data. Results suggest improved synthetic reconstruction and biologically plausible spatial patterns. The authors appropriately describe the outputs as inferred refinements, rather than directly recovered cellular measurements.

Critical strengths

The paper addresses a practical problem where complementary measurements may be unavailable. Its self-supervised formulation makes effective use of the observed data. The following strengths support the potential value of the contribution. 1 Pooling observed spots provides training targets without requiring paired low- and high-resolution measurements or external references. 2 Virtual nodes provide an explicit mechanism for spatially indexed outputs, with applications using both nine and sixteen subspots. 3 Combining a VAE with graph attention is a coherent response to sparsity and the distribution shift between pooled training inputs and inference inputs. 4 Synthetic results are encouraging: Square reports MSE 0.39 and ARI 0.88, compared with 2.33 and 0.67 for BayesSpace, respectively. and lastly, The evaluation includes component ablations, sparsity analyses, and biological examples in two tissues. The explicit qualification that subspots are model-based estimates also strengthens the paper’s framing.

Critical weaknesses

My main concern is whether Square provides reliable fine-scale spatial detail beyond smoothing. MOB entropy and proportion metrics cannot validate within-spot placement because they are unchanged by rearranging subspots. The Visium HD experiment provides aligned 8-µm and 2-µm measurements, but lacks a direct quantitative recovery comparison against interpolation or smoothing baselines.

The benchmark also needs stronger controls. Algorithm 2 appears to reuse one expression profile and dropout pattern per cell type, limiting biological variability, while independence between the synthetic benchmark and the 500-trial hyperparameter search is unclear. STAGE is evaluated at four locations against nine-subspot ground truth, although its decoder supports arbitrary coordinates. Independently generated simulations and comparisons at identical target locations would strengthen the conclusions.

The method’s consistency with observed measurements also requires clarification. Equation 5 does not establish how predicted subspots reaggregate to their parent measurements, and Algorithm 1 appears to pool already log-normalised profiles. The VAE’s handling of resolution shift needs evidence beyond PCA overlap, while the use of edge weights and whether the decoders share parameters remain insufficiently specified. These gaps limit confidence in both the inferred expression values and reproducibility.

Questions for authors

What is the quantitative recovery performance on the 8-to-2-µm Visium HD experiment, compared with measured fine-resolution profiles and simple interpolation or smoothing on the same grid?

Are spots pooled before or after normalisation and log transformation? How closely do predicted subspots reproduce the parent measurements when reaggregated?

Does Algorithm 2 generate independent expression variability and dropout at each location? Were the final synthetic benchmarks generated independently of hyperparameter tuning?

Why was STAGE not evaluated at the same target coordinates, given its arbitrary-coordinate decoder? Do the conclusions persist when output locations and expression scales are aligned?

and my last question is that how do the adjacency weights enter message passing, and are the VAE and enhancement decoders shared? Please specify the likelihoods used in Equation 5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.