acceptodds
Under review as a conference paper at ICLR 2027

TissueField: Reconstructing Spatial Gene Expression with Reusable Conditional Neural Fields

Abstract

Spatial transcriptomics samples gene expression at discrete locations, creating a need to reconstruct spatial expression patterns from partial and irregular observations. Implicit neural representations offer a continuous model of these patterns, but fitting a separate field for each tissue section limits the reuse of learned structure. We present TissueField, a conditional neural field that learns from source sections to reconstruct expression on a new section from partial observations. TissueField constructs local context–query episodes and uses a Transformer hypernetwork to infer window-specific modulation of a shared coordinate network. The network predicts compact expression embeddings, which a shared decoder maps to gene expression. On a new section, conditioning and reconstruction require only forward computation, with modulation reused across queries within each window. Across Stereo-seq MOSTA, MERFISH, Visium, and Slide-seqV2, within-platform cross-section evaluations show consistent reconstruction gains over directly transferred baselines. Spatial analyses show preservation of characteristic expression patterns and tissue organization. These findings support reusing learned reconstruction structure to recover expression on new sections without repeated model fitting.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial Transcriptomics, implicit neural representation, Hypernetwork

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3TPLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a conditional neural field for reconstructing spatial gene expression across tissue sections without target-specific optimization.

Critical strengths

Addresses a practical problem in spatial transcriptomics: efficient reconstruction of missing gene expression across tissue sections. Demonstrates promising reconstruction performance across multiple ST platforms.

Critical weaknesses
  1. Limited scientific novelty. It applies established hypernetwork and neural-field techniques to spatial transcriptomics without introducing a substantially new modeling principle.
  2. Limited biological impact. Reconstructing gene expression at missing spatial coordinates is useful, but the paper does not demonstrate that this enables new biological discoveries, improves downstream scientific conclusions, or addresses a major limitation of existing ST analyses.
Questions for authors
  1. Can the authors demonstrate improvements over simpler spatial interpolation or conditional regression models? Although the paper includes several coordinate-based regression baselines (FFN, SIREN, and SUICA variants), the main comparisons are not sufficiently controlled. TissueField conditions on observed target-section expression, whereas the baselines are evaluated through direct parameter transfer. Simple but relevant alternatives, such as kNN-based spatial interpolation and context-conditioned MLP regression using the same target observations, are missing. Without these comparisons, it is difficult to establish whether the proposed architecture provides meaningful advantages over simpler methods.
  2. How well does the method preserve biologically meaningful expression patterns and generalize across independent biological samples?
Reviewer_bLpYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TissueField, a conditional neural field framework for reconstructing spatial gene expression from partial observations. Building upon existing implicit neural representation methods such as SUICA, TissueField uses a Transformer hypernetwork to predict window-specific modulation parameters for a shared SIREN network. This enables reconstruction on new tissue sections without target-specific parameter optimization.

The method is evaluated on four spatial transcriptomics platforms, including Stereo-seq MOSTA, MERFISH, Visium, and Slide-seqV2. The authors report improved reconstruction accuracy over existing INR-based methods under cross-section transfer, along with substantially reduced target-section processing time.

Critical strengths
  • The paper addresses a meaningful limitation of existing INR-based spatial transcriptomics reconstruction methods, which typically require separate optimization for each tissue section. Learning a reusable conditional model that adapts to new sections through forward inference is a reasonable and practically valuable direction. The reported reduction in target-section processing time is particularly encouraging.
  • The method achieves strong reconstruction results across several spatial transcriptomics datasets. The in-section experiments and component ablations provide additional evidence that the proposed conditional reconstruction framework can effectively model spatial gene-expression patterns.
Critical weaknesses
  • My main concern is that the primary cross-section comparisons do not provide equivalent information to all methods. TissueField uses observed gene expression from the target section as conditioning input, while the baseline INR methods are directly transferred from source sections without access to the same target observations. Therefore, the substantial performance improvements cannot be attributed solely to the proposed architecture. Although the paper also reports in-section comparisons, these do not fully resolve the issue. More appropriate comparisons with alternative conditional reconstruction architectures using the same target observations would be necessary to demonstrate the specific advantages of TissueField.
  • The method mainly combines established components, including Transformer hypernetworks, conditional parameter modulation, SIREN, and latent expression decoding. While the overall framework is reasonable, the methodological novelty beyond adapting existing techniques to spatial transcriptomics appears limited.
  • The claimed reusability is evaluated primarily under same-platform transfer, with several experiments training on only one source section. It remains unclear how broadly a source-trained model can generalize across substantially different tissue structures and biological conditions. For example, the MOSTA developmental stages exhibit considerable morphological and molecular variation, but the paper provides limited stage-specific analysis of this transfer behavior.
Questions for authors
  • Can the authors provide comparisons with conditional reconstruction methods that use the same target-section observations and require no target-specific optimization? This would help isolate the contribution of the proposed hypernetwork-modulated INR.
  • What is the total training time of TissueField for each dataset? How many new sections must be processed before its amortized inference becomes computationally more efficient than per-section INR fitting?
  • How does reconstruction accuracy vary across different MOSTA developmental stages when training only on E15.5? Does performance systematically decrease as the biological differences between source and target sections increase?
  • Could the authors compare the proposed hypernetwork-based modulation against simpler conditioning mechanisms under matched target observations and comparable model capacity?
1 comment

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns, and these additions strengthen the experimental validation and demonstrate the practical value of amortized reconstruction.

However, my main concern regarding methodological novelty remains. The improvements of hypernetwork modulation over direct context injection are relatively modest, and this controlled comparison is limited to MERFISH. Moreover, the primary cross-section comparisons still involve unequal access to target observations, making it difficult to attribute the reported gains specifically to the proposed architecture.

While I acknowledge the substantial efficiency benefits and the improved experimental evidence, I remain unconvinced that the work establishes a sufficiently strong methodological contribution beyond existing conditional neural field techniques. Therefore, I will maintain my original rating.

Reviewer_56WSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on the spatial gene expression reconstruction from partial, irregularly distributed measurements in the ST. The method TissueField learns a conditional neural field on source sectiosn that can be reused on a new section without any target specific parameter optimization. The context tokens combine projected latestn space feature with a projected local coordinates. Traiing is 2 staged, with latent space warm up and expression space fine tuning. The evaluation is conducted with four resources under cross section transfer with fixed parameters. Results show promising sensitivity and accuracy.

Critical strengths
  1. The evaluation description is clear and comprehensive, making this paper a well evalauted contribution.
  2. The results and ablations studies conducted on 4 major datasets are effective and convincing.
  3. TissueField solves a challenging new problem on reconstructs that compress time-consuming cost. This method improves the cross section recon over all evaluated baselines on the four resources.
Critical weaknesses
  1. The evaluation of TissueField that sees the target data compared to baselines needs some clarity. For example, TissueField gets 512 observed spots per window from the test section, does baselines sees the same data?
  2. The paper ensembled a few exisitng work such as the coordinated network and modules, the latesnt expression space from SUICA, and the modulation FiLM. The authors can highlight the key adaptation on the motivation of gene expression domain.
Questions for authors
  1. If the filled in helout sspots by averaging in nearby spots, how close it is got to TissueField?
  2. How does baseline methods evaluated get the target data on TissueField sees? This might need a study.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.