acceptodds
Withdrawn from ICLR 2027

HyBoxST: Coupling Hyperbolic Spatial Representations with Molecular Program Boxes for Transcriptomics Prediction from Histology

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

13 views

Abstract

Predicting spatial gene expression from histology provides a scalable way to profile tissue molecular states. Although existing methods increasingly capture tissue morphology and spatial context, they rarely model the pathway structure among predicted genes explicitly. This leaves known relationships among functionally related genes underused when learning their expression from tissue morphology. This makes it difficult for these models to accurately recover coordinated expression changes among genes. Here we propose HyBoxST, which incorporates gene-to-pathway memberships into expression prediction through a pathway-guided structured decoder. HyBoxST represents pathways as learnable boxes and genes as points in a shared molecular latent space, so that pathway membership can be expressed through containment and shared membership through overlap between pathway boxes. For each spot, features from the spot image and its surrounding niche generate a query box. Its overlap with pathway boxes is used to adjust directly predicted pathway expression scores according to the local tissue context. For each gene, the decoder aggregates the corrected scores from pathways to which the gene belongs, while a residual branch captures expression variation beyond the selected pathways. Across four tissue datasets, HyBoxST consistently improves gene expression prediction. Additional controls with randomized gene-pathway memberships and spatial analyses of pathway signals show that the improvement relies on the biological pathway structure rather than the mask structure alone.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics prediction, Pathway-guided decoding, Box embeddings, Histology images

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_g9V4OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces HyBoxST, a method for predicting spatial gene expression from H&E images. Built upon HyperST, HyBoxST incorporates a pathway-guided decoder. Specifically, the method involves representations for both pathways and genes. Reactome/Hallmark pathways are represented as learnable axis-aligned boxes, while genes are represented as points, with a membership constraint encouraging genes to fall within the boxes of their respective pathways. To generate a correction for pathway scores, a "query box" is derived from spot and niche image features; the smooth overlap between this query box and each pathway box is mapped to an additive correction applied to the directly predicted pathway score (trained to match the mean log-expression of member genes). These corrected pathway scores are aggregated into gene-level predictions using non-negative weights constrained by membership masks, followed by gated fusion with a branch that directly predicts gene expression. Experiments were conducted on four tissue types from the HEST-1K dataset (kidney, colorectum, skin, and lung) using 200 genes per tissue and five random data splits. Performance was evaluated using metrics such as PCC@200, MSE, and MAE, with comparisons made against baseline results from the original HyperST paper; additional analyses included component ablation studies, pathway mask shuffling experiments, and an examination of pathway scores in kidney tissue.

Critical strengths

The inductive bias is natural, and the internal control experiments show consistent trends. Using box embeddings to represent the concept that a single gene can belong to multiple overlapping pathways, and then matching this against queries generated from image features, is a natural and somewhat novel design; furthermore, component ablation and mask-shuffling control experiments yield consistent results across four tissue types. This is likely the paper's only methodological innovation that could justify its acceptance.

The code is relatively complete, allowing key details to be verified. Based on the released code, inference-time predictions rely solely on image features, while the measured expression of test spots enters only the evaluation, ensuring no label leakage occurs; additionally, checkpoints are selected based strictly on the validation set. This rules out test-label leakage during training and inference, a common pitfall in this type of work, and the results are verifiable within the authors' own pipeline.

Critical weaknesses
  1. The contributions are meager, and the distinction from existing work is unclear. The three claimed contributions (the overall method, a specific module within it, and the experiments) effectively boil down to one.

The combination of "pathway scores as intermediate variables + membership-masked decoding + residual branches" follows the same logic as MSGR and PEaRL, yet neither was used as a baseline for comparison; the hyperbolic alignment is inherited from HyperST. The only truly novel elements are the box query correction and the learned static weights WW. Consequently, the vast majority of components are unremarkable.

The impact of the new components is minimal. In Appendix D.4, the box geometry outperforms an MLP query with matched parameter counts by only 0.006–0.013 (PCC@200). No ablation study compares WW against uniform averaging. Both WW and the gating mechanism are global parameters, which contradicts the stated motivation of "allowing the contributions of functional groups to gene prediction to reflect local tissue context."

The relationship with HyperST is inadequately explained, potentially misleading readers. The main text contains only a single sentence stating "follow the spatial alignment framework of Zhang et al. (2026)," and Appendix A.1 mentions "inherited" in two places, yet it never explicitly clarifies that this framework is the very same HyperST listed as a baseline in Table 1. Neither the abstract nor the introduction states that "HyBoxST = HyperST framework + new decoder," even though the title foregrounds "Hyperbolic Spatial Representations" and no ablation study is provided for the hyperbolic component. The code and README also fail to mention HyperST.

Appendix B contains mathematical redundancies. The 21 equations mostly cover elementary facts, such as the invariance of Pearson correlation to positive affine transformations, Jensen's inequality, expansion of squares, the triangle inequality, and operation counts; none of these support the core claims. Conversely, the analysis actually needed regarding the inductive bias introduced by boxes is missing.

For a top-tier deep learning/representation learning conference like ICLR, the methodological contribution here is incremental and lacks a clear distinction from existing work. While the paper possesses some novelty with experiments, I believe it would be more suitable for submission to journals such as BIB or Bioinformatics.

  1. The main conclusions lack controlled comparisons, making it impossible to attribute the performance gains to the proposed method.

HyBoxST employs UNI (ViT-L, ~300M parameters, pre-trained on large-scale pathology tiles) with LoRA fine-tuning; in contrast, the original papers for StNet, BLEEP, and TRIPLEX utilized ImageNet-pre-trained DenseNet-121, ResNet-50, and self-supervised ResNet-18, respectively. The paper fails to specify the encoders used for the baselines in Table 1, whereas the HEST-Benchmark has demonstrated that simply changing the encoder can lead to significant score improvements.

Relying solely on the UNI foundation model (FM) makes it impossible to determine whether the gains from the decoder are independent of the encoder. The only evidence from a controlled setting is weak; gains observed in the internal ablation study range from approximately 0 to 0.05, falling within the margin of standard deviation. Each split contains only 1–2 test slides (5 for the Skin dataset), and the five random splits overlap with one another. Furthermore, the 200-gene panel was selected using the entire dataset, including the test slides.

On the Colorectum dataset, HyBoxST underperforms compared to HyperST and TRIPLEX in PCC@200, contradicting the abstract's claim that HyBoxST "consistently improves gene expression prediction."

Based on the information currently presented in the paper, I cannot determine whether the improvements stem from the pathway decoder or from the stronger backbone and distinct training pipeline.

Questions for authors

All baseline numbers in Table 1 are quoted from Zhang et al. (2026). The HyBoxST pipeline uses EMA, data augmentation, additional loss terms and post-hoc calibration, and the paper acknowledges that the quoted baselines are not guaranteed to follow the same pipeline. However, Figure 4 (Appendix D.6) shows per-spot predictions of HyperST, StNet, BLEEP, Stem and TRIPLEX on NCBI704, which indicates that the authors are able to run these baselines. Please report the following results under an identical pipeline (same gene panel and data splits, LoRA-tuned UNI, EMA, data augmentation, loss terms and post-hoc calibration).

  • (a) HyperST;
  • (b) the hyboxst_base variant in the released code, which does not include the pathway decoder;
  • (c) TRIPLEX, BLEEP and StNet using UNI features, or at least a UNI + MLP/ridge baseline following the HEST-Benchmark protocol.

Does the gain of HyBoxST over hyboxst_base persist when UNI is replaced by at least two other encoders (e.g., an ImageNet-pretrained ResNet-50 and another pathology foundation model), or when UNI is frozen?

Under the same setting, how does HyBoxST compare with MSGR and PEaRL? What happens if the learned weights W are replaced by uniform averaging over the pathways a gene belongs to? Since W and the gene gates are shared across all spots, how does the model realize the stated motivation that pathway contributions vary with local tissue context, beyond the additive correction to the pathway scores?

Please state explicitly in the abstract and introduction which of the following components are inherited from HyperST, namely the image encoder and LoRA setup, hyperbolic alignment, niche construction, gene selection, data splits, and the training and evaluation pipeline. Please also report results with the hyperbolic alignment removed; if it is not essential, the title and the framing of the contributions should be revised accordingly.

Since the current five random splits overlap with each other, disjoint folds or leave-one-slide-out evaluation would be preferable. Please also report results with the gene panel selected from training slides only. Finally, the abstract claims consistent improvement, but HyBoxST underperforms HyperST and TRIPLEX in PCC@200 on Colorectum; please revise this claim or explain the result.

If the authors address these questions thoroughly in the rebuttal, I will consider raising my score to 2.

Reviewer_jBVNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors of this paper introduces HyBoxST, a method to predict spatial gene expression from H&E data. The central motivation of this paper is valid as the author mention since genes within a pathway shows coordinated expression changes, they try to incorporate pathway structure (gene-gene) into the model to improve the spatial expression reconstruction from H&E. Pathways are represented as learnable boxes and genes as points, and training encourages genes to fall inside the boxes of their pathways. Experiments conducted on HEST1K data set on 4 tissues and authors include ablations. I found the central idea of the paper clean and elegant.

Critical strengths
  1. The problem formulation is interesting (representing pathways as learnable boxes while allowing the membership overlap) and suite well to biology than formulating one vector per pathway.

  2. The writing is clean and concise

Critical weaknesses
  1. All the baselines in table 1 is copied from Zhang et al. (2026), not rerun.

  2. The author argues that current models fail to recover coordinated expression changes, and they mention that pathway coherence (HistoPrism) can be good alternative metric, but in their own evaluation uses only pergene PCC, MSE, and MAE. There is no pathway coherence comparison against baselines.

Questions for authors
  1. Could you please rerun the all the baselines with the same splits of the data, gene panel, preprocessing and report the reults?

  2. Could you please report the pathway coherence for all the methods?

  3. Could you please add significance test on the results (several seeds, random shuffled masks, etc )?

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

This paper proposes HyBoxST, which uses pathway expression scores as a shared intermediate representation between histology images and gene expression prediction. Annotated gene–pathway memberships determine how these scores are propagated to individual genes, and a direct gene prediction branch compensates for incomplete pathway coverage. The paper reports improved gene expression prediction across four tissue datasets. Randomized membership controls and a spatial analysis of pathway signals further suggest that the gains depend on biological pathway structure, not merely on the structure of the membership mask.

Critical strengths

1.Most existing H&E→ST methods regress each gene independently and make little use of functional structure among genes. Using pathways as an intermediate representation between morphology and gene expression is more biologically meaningful. 2. The experiments are thorough, with comprehensive ablations and controls.

Critical weaknesses
  1. All baseline results in Table 1 are quoted directly from Zhang et al. (2026) rather than rerun under the paper's pipeline. HyBoxST uses LoRA adaptation of UNI, EMA, training-time augmentation, and an additional post-hoc per-gene calibrator. The paper does not clarify whether the cited baselines use the same settings, and Appendix C.2 explicitly states that external baselines are not guaranteed to follow the same inference pipeline. The current results therefore cannot separate the contribution of the proposed method from differences in training and inference configuration.

  2. HyBoxST inherits its visual backbone and hyperbolic alignment module from HyperST, but the ablations mainly test the incremental contribution of components within the structured decoder. For example, "w/o query box" still keeps pathway prediction, membership-constrained decoding, and gene prediction fusion. It therefore only isolates the contribution of the query correction. It does not show the net benefit of the full pathway-guided decoder over direct gene prediction.

  3. The abstract states that "Across four tissue datasets, HyBoxST consistently improves gene expression prediction," which readers may take to mean it outperforms existing methods on all four datasets. In Table 1, however, PCC@200, the primary metric, drops on Colorectum from 0.477 (HyperST) to 0.438, a relative decrease of about 8.2%. MAE on Kidney and Skin is also slightly worse than HyperST.

  4. The pathway boxes and query matching live in a separate Euclidean space, and the decoder uses visual features taken before the Lorentz mapping. The hyperbolic component comes from the inherited HyperST alignment framework. The two are therefore linked only indirectly, through shared visual representations and joint training, rather than through explicit geometric coupling. The paper neither removes LalignL_{align} to test the contribution of alignment supervision, nor compares against a matched Euclidean alignment to test the value of hyperbolic geometry itself. The current evidence mainly supports the Euclidean box decoder design. It does not justify the title's emphasis on "coupling hyperbolic spatial representations with molecular program boxes."

Questions for authors
  1. Under the same encoder, alignment objective, training, and post-processing pipeline (including LoRA, prediction losses, EMA, the affine calibration layer, and the post-hoc calibrator), could the authors compare HyperST, a model with only a direct gene prediction head, the pathway decoder without the query box, and the full HyBoxST? This would separate the contributions of the training pipeline, pathway decoding, and the box query. Please also confirm whether all methods in Table 1 use exactly the same 200-gene panel and five slide splits.

  2. The gene panel and pathway subset are currently selected using expression statistics that include the test slides. If selection used only the training slides, with all methods predicting the same panel, would the main conclusions still hold? In addition, do the splits keep patients or specimens independent, so that slides from the same source or adjacent sections never appear on both sides of a split?

  3. For the key comparisons (full model vs. w/o query box, MLP query, and the two randomized membership controls), could the authors report paired performance differences? Please also add multiple training seeds and multiple independent random masks (e.g., 10 per randomization scheme) to separate training randomness from mask randomness. Confidence intervals or statistical tests should account for clustering by patient or slide, and for test slides that recur across runs.

  4. Could the authors report membership containment and non-member exclusion rates before and after training, together with the across-spot variation of the overlap scores and the fraction saturated at the clipping bounds? How does performance change when the image-conditioned overlap is replaced by a fixed mean overlap? These analyses would help determine whether the gains actually come from the learned membership geometry and the local image-conditioned query.

  5. Beyond PCC/MSE, could the authors evaluate how well co-expression among genes within a pathway is recovered, e.g., agreement between predicted and ground-truth gene–gene correlations? Could they also include representative pathway-guided methods, and a simple control that replaces the learned pathway-to-gene weights with uniform weights? This would show whether the decoder has an advantage over existing methods and over simple aggregation.

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

Related papers

Discussion (0)

Sign in to comment.