acceptodds
Under review as a conference paper at ICLR 2027

Gene-Amortized Fields for Held-Out Gene Imputation in Spatial Transcriptomics

Abstract

Spatial transcriptomics profiles gene expression across intact tissue but measures only a bounded gene panel. We study gene-level generalization within a tissue domain, where a gene that is described by domain-level side information but withheld entirely from model supervision must have its spatial expression reconstructed at every location from the remaining genes. A gene held out in full has no measurements of its own to fit, which makes the task out-of-column matrix completion. Completion is feasible only if genes carry side information relating masked columns to observed ones. Existing methods take that side information from an external single-cell reference in which the masked gene is measured, and none of the eight tissue domains studied here has a paired one. We instead learn a single gene-independent coordinate field, the shared field, and decode every gene from it through a map amortized from a fixed gene embedding. The embedding is computed from the panel's own co-expression rather than fitted to the gene it describes, so a gene the model was never supervised on is decoded by the same map with no code fitted to it. It need not appear in a single-cell atlas, though it must have been measured in spatial data of the domain, and a cell-type reference still anchors the count likelihood. The imputation is therefore reference-light rather than reference-free. Expression is structured at multiple non-stationary scales, so the field is multi-resolution, instantiated as a gene-amortized multi-scale wavelet representation. On eight HEST-1K tissue domains spanning two species and five organs under gene-level cross-validation, the model surpasses the strongest of ten reproduced baselines on average ( against ) and on seven of eight domains. The per-domain differences are significant under a paired signed-rank test (). A single-scale variant is worse on all eight. Over the two hundred genes whose measured expression is most spatially autocorrelated, the mean correlation is . A low-rank spectral coupling added to the readout leaves accuracy unchanged but exposes a factorization over genes whose modes align with measured co-expression in all eight domains once the input embedding is controlled for, and recover programs recurring across organs and species.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial Transcriptomics, Gene Imputation, Implicit Neural Representation, Out-of-column Matrix Completion

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 2 reviews, above 27% of the others.
Reviewer_UquaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies held-out gene imputation in spatial transcriptomics, where part of the gene panel is masked during training and has to be predicted at every spot. The model is a multi-scale wavelet INR shared across genes. Each gene is decoded through weights produced by an amortizer from a fixed 68-d gene embedding, which consists of a truncated-SVD co-expression code plus Moran's I features. A negative-binomial deconvolution path ties the latent to a cell-type signature matrix. On eight HEST-1K domains, the method reports a mean Pearson correlation of 0.300, compared with 0.265 for SpaGE, the best of ten reproduced baselines.

Critical strengths
  1. The idea of generating all per-gene decoding weights from a fixed gene embedding is simple and makes sense. The paper also explains clearly why existing INR methods such as STINR and SUICA cannot give a prediction for an unseen gene.

  2. The multi-scale design is well supported by the ablation. The single-scale variant is worse on all eight domains (Table 2), and Fig. 4b shows that the loss is larger for the more spatially structured genes.

  3. The evaluation is broad (8 domains, 90 Visium sections, two species), and all baselines are run through one driver with the same panels, splits and evaluator.

Critical weaknesses
  1. My main concern is that the held-out genes are not really held out. In App. B, the code of gene gg is the gg-th row of VΣV\Sigma from a truncated SVD over all slides and all genes, which includes the held-out columns and the scored slides. However, (VΣ)g=x~g⊤U(V\Sigma)_g = \tilde{x}_g^\top U. In other words, the code is the held-out gene's own expression projected onto the 64 leading spot factors, so U(VΣ)g⊤U(V\Sigma)_g^\top is already a rank-64 approximation of the column being predicted. Since UU is mostly determined by the observed genes, the model can largely recover the gene from its own measurements. The embedding also contains the gene's measured Moran's I (Sec. 3.1), which Sec. 4.3 shows is the main driver of accuracy. So I do not agree with the claim in Sec. 4.1 that the side information is "no more informative" than what the baselines receive, and this is not the cold-start setting described in Sec. 1 either. I don't see how this can be fixed without changing the problem itself.

  2. The motivation and the baseline setup also do not fit well together. The paper argues that no paired single-cell data is available (Sec. 1). Yet BB is built from a single-cell reference of the same tissue (App. A), which is exactly the kind of unpaired reference that SpaGE, Tangram etc. are designed for. However, these baselines were given the remaining Visium slides instead. SpaDiT is negative on all eight domains (Table 1), which suggests that some of them do not work properly in this setup.

Besides, the ML novelty seems limited, since the method is mostly a wavelet INR combined with hypernetwork-style conditioning and STINR-style deconvolution.

Questions for authors
  1. Can the authors confirm that the held-out gene's column, including its values on the scored slides, enters the SVD that produces its code? If so, how does the full model compare with simply reading out the rank-64 reconstruction U(VΣ)g⊤U(V\Sigma)_g^\top?
  2. Why were the reference-based baselines not given the same-tissue single-cell reference that is used to build BB?
Reviewer_UwyFOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses reconstructing gene expression values within a tissue domain of spatial transcriptomics data, framing it as a matrix-completion task in which entire columns (corresponding to specific genes) are missing. The model transforms 64-dimensional truncated-SVD ​​co-expression codes and 4-dimensional features related to Moran’s I into gene-specific readout parameters. This approach enables predicting expression values for held-out genes without learning separate parameters for each gene. The model outperformed other methods in 7 out of the 8 HEST-1K domains.

Critical strengths

S1. The design concept for addressing the problem is clear. Instead of directly learning weights, for each gene, the model generates them from a gene code.

S2. The experimental scope is extensive. The study reproduces several baselines across eight tissue domains and conducts comparisons using consistent gene splits and evaluation metrics.

S3. This work analyzes performance variations across individual genes in detail. Beyond reporting average performance, the study examines performance in relation to spatial autocorrelation.

Critical weaknesses

W1. This work does not compare against a simpler model that uses the same gene code as input. This method takes an SVD code as input that already encapsulates genetic relationships. Consequently, the model begins with highly significant, pre-existing information.

Therefore, the first thing that should be verified is: “To what extent can the data be reconstructed by simply applying a linear decoder or ridge regression to this code?” Without this comparison, I cannot assess how much the proposed representation contributes to the reported performance. Even if the authors define the problem as "inductive matrix completion," they omit a simple low-rank baseline that directly matches this definition.

W2. The effect of multi-scale itself has not been isolated.

Switching to a single-scale approach also alters the number of spatial features and the auxiliary learning scheme. Hence, it is difficult to determine whether performance improvements stem from using multiple resolutions or from including more features or additional learning mechanisms. A comparison that controls for model size and training conditions is required.

Questions for authors

Q1. Would you please compute the SVD codes and Moran’s I features using only the other slides, as you mentioned in Appendix B. How does performance change when the evaluation slide’s observed genes remain available for training?

Q2. Please evaluate a linear or ridge regression baseline using the same gene codes, trained only on the observed genes. How does its performance compare with that of the proposed spatial model?

Comments on the paper

1 comment
Area_Chair_zEFjOpenReview ↗

Please check authors' rebuttal and participate in the discussion

Dear reviewer,

The discussion phase is coming to an end. Please read the authors’ rebuttal and post a brief follow-up in the discussion thread. Let us know which of your concerns the response addresses and which remain open.

Please post any follow-up questions in the thread visible to the author as soon as possible, so that the author has time to respond and provide clarification. I also encourage you to engage with one another’s assessments—especially where your perspectives differ—and to note whether any counterarguments or discussions lead you to revise your own assessment.

Thank you for your effort and thoughtful review. Thank you again!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.