acceptodds
Under review as a conference paper at ICLR 2027

SpatialDataAgent: Autonomous Curation of a Decade of Public Spatial Omics at Scale

Abstract

Spatial transcriptomics maps gene activity within tissue sections. When paired with tissue images, these measurements connect molecular activity to visible tissue structures, providing a basis for studying cellular organization and developing image-based molecular analyses. Yet public spatial omics data often remain difficult to reuse—stored across non-standard archives where coordinates, histology images, and molecular counts frequently diverge, use proprietary containers, or are mixed with dissociated single-cell data. Here, we present SpatialDataAgent, an autonomous system that converts public deposits into standardized multimodal datasets at scale. SpatialDataAgent is designed to preserve source-deposited measurements through source-grounded coordinate transformations and explicit data-preservation constraints: it grounds coordinate mappings in deposited metadata, uses a catalog of 26 specialized domain tools for spatial formats, and checks candidate alignment through automated visual diagnostics. Evaluated on 400 reference studies, our screening module identifies eligible datasets with 91.75% classification accuracy and a 0.913 macro-F1 score. In a controlled evaluation across 40 diverse GEO studies (331 samples per condition), domain workflow support increases human-reviewed visual alignment from 54.7% to 77.6% (+23.0 percentage points; 95% bootstrap CI: 8.6–45.3), reaching 81.0% when coupled with visual quality feedback. Using SpatialDataAgent, we construct and release HESRT (Histology-paired Expression atlas of Spatially Resolved Transcriptomics), an open atlas of 4,468 human-reviewed multimodal samples comprising 37.05 million spatial observations across multiple species and technologies. HESRT makes matched images and gene measurements available for studies of tissue organization and image-based molecular modeling.

Primary area:
datasets and benchmarks
Keywords:
autonomous agents, scientific data curation, spatial transcriptomics, multimodal datasets, tool-augmented code generation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_3mFYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents SpatialDataAgent, an LLM coding agent that turns public GEO spatial transcriptomics deposits into paired histology, coordinate and expression data. A screening module first classifies studies by the evidence for H&E pairing. A reconstruction agent then writes and runs study-specific code using 26 domain tools and an optional visual QC checker. On 40 GEO studies, the domain tools raise human-reviewed alignment from 54.7% to 77.6%, and to 81.0% with QC.

Critical strengths
  1. The problem is a real practical bottleneck for anyone working with public spatial data. The released HESRT collection is considerably larger than HEST-1k and STimage-1K4M, so it could be useful to the community.
  2. The controlled comparison is set up carefully. The model, inputs and budgets are the same across conditions, failed and missing outputs are kept in the denominator.
Critical weaknesses
  1. My main concern is the contribution from an ML perspective. The system is an existing LLM coding agent (a Codex-based scaffold running Muse Spark 1.2) with 26 hand-written domain tools and a rule-based image checker. The main finding is that giving the agent domain-specific tools raises the pass rate, which is expected. The gain is also concentrated where the tools encode the file format: standard Visium goes from 60% to 99%, while containers and alternative arrays stay at around 50% (Fig. 2d). I could not find a new agent method, or a lesson that would transfer beyond this specific curation task. Also, the value of HESRT for ML is not shown. There is no downstream experiment, and Fig. 3b compares the resources only by the number of samples that pass four basic geometric checks (unique IDs, finite coordinates, unique centers, centers inside the image), which says little about alignment quality.

  2. The headline results also overstate how well the data integrity claims are verified. The 81.0% in the abstract is a human visual PASS on overlays, which cannot catch small offsets, scale errors or altered expression values. However, the automated source audit, which checks exactly what the paper claims to preserve, passes only 133 of 331 samples for the full system. Depending on how unresolved cases are counted, the pass rate lies between 40% and 77% (App. C, D), and none of these numbers appear in the main text. Moreover, QC is presented as one of the three design principles, but its effect is not significant in either comparison (Table 2). This is despite Sec. 5.2 being titled "Domain Tools and Quality Checks Drive Successful Reconstruction". The screening reference was built with AI cross-review and sampled using the same model's earlier predictions (Sec. 4.1), so the 91.75% agreement is not fully independent either.

Besides, every released sample still went through human review (App. E), so "autonomous" seems too strong a description of the curation pipeline as a whole.

Questions for authors
  1. Given the paper's emphasis on preserving deposited measurements, why is the main evaluation based on visual PASS rather than on the source audit? How should readers interpret the gap between 81.0% (visual) and 133/331 (source audit) for the full system?
  2. How exactly were the 400 reference labels produced, and in how many cases did the AI cross-review change the initial human label?
Reviewer_9dy8OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SpatialDataAgent, an autonomous system for curating public spatial omics data into paired tissue images and gene expression measurements. It combines study screening, 26 domain-specific tools, and automated quality checks while preserving source measurements and provenance. Domain workflow support improves alignment pass rates from 54.7% to 77.6%, reaching 81.0% with quality feedback. The authors also construct HESRT, a resource containing 4,468 human-reviewed samples across 507 studies.

Critical strengths
  1. The paper addresses a practical problem in reusing public spatial omics data, with careful attention to data integrity and provenance.

  2. Human evaluation shows a substantial improvement in alignment quality, and the resulting HESRT dataset is a useful resource for future research.

Critical weaknesses
  1. My first concern is that the experiments do not clearly isolate the contribution of the proposed workflow. Domain tools, guidance, and evidence interfaces are introduced together, so a tools-only (or guidance/ecidence interfaces only) baseline using the same coding agent would help establish what the workflow itself adds.

  2. The evaluation uses only Muse Spark 1.2, leaving the framework’s robustness across models untested. In addition, both comparisons measuring the incremental benefit of QC in Table 2 have confidence intervals that include zero, providing limited evidence that QC improves performance within the framework.

  3. While HESRT has reported quality checks, its value for downstream modeling is not demonstrated through comparisons with established curated datasets.

Questions for authors

Please see Weaknesses.

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

The authors propose SpatialDataAgent, an LLM-based agent for autonomously curating public spatial omics datasets. The system screens GEO studies and uses a coding agent with 26 specialized tools and visual QC to reconstruct paired histology images, spatial coordinates, and expression matrices. The authors evaluate screening on 400 studies and reconstruction on 40 studies, showing improvements in visual alignment from 54.7% to 81.0%. They further release HESRT, a collection of 4,468 human-reviewed samples with over 37 million spatial observations.

Critical strengths
  • I think the problem is quite important. There is a large amount of publicly available spatial omics data, but curating it into usable image-expression pairs requires substantial manual effort. Automating this process is valuable, especially for training spatial omics foundation models.

  • The experiments show substantial improvements over the base coding agent, particularly with the domain-specific tools.

  • The HESRT dataset is potentially a useful contribution by itself. Having over 4,400 paired samples across different platforms and species could be valuable for downstream spatial biology research.

Critical weaknesses
  • I am not fully convinced by the methodological novelty of the agent. The system essentially uses a coding agent with specialized tools and workflow instructions. While the improvements are substantial, I feel much of this comes from providing the right domain-specific tools rather than the agent itself. For eg, in Figure 2d, standard Visium reconstruction improves from 60% to 99%, but these formats already have well-defined parsers. What happens with a deterministic pipeline using the same tools? Moreover, the ablation combines tools, instructions, and evidence interfaces, so it is difficult to isolate which components actually contribute.

  • I have some concerns about the data validation. The main results report 77.6-81.0% human visual alignment, but Appendix D reports only 134/331 and 133/331 samples passing the stricter audit. What fraction of the released HESRT dataset is fully verified for coordinate and expression preservation?

  • Although HESRT is a substantial resource, there is limited evaluation of its downstream utility. The comparison with HEST-1k and STimage-1K4M is primarily based on dataset size. Does the additional curated data actually improve image-to-expression prediction or spatial representation learning? Can we run the same downstream experiments as HEST as a comparison?

Overall I think the paper addresses an important practical problem, and the dataset could be valuable for spatial omics research. However, I am not fully convinced by the methodological contribution, and I feel more rigorous source verification and downstream evaluation are needed to support the broader claims about the released dataset. I also want to point out that the paper, especially the appendix, was extremely difficult to read, a major rewrite is needed. Moreover, the dataset hosted is gated, and I was unable to access it. I also went through the code, and it was extremely difficult to follow. I lean to weak reject in the current state.

Questions for authors

Mentioned along with weaknesses.

Reviewer_H6VhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SpatialDataAgent is an LLM-based framework for automatically curating public spatial transcriptomics datasets. It combines agentic code generation with domain-specific tools for automated spatial alignment diagnostics and reconstruction of paired histology images and gene expression measurements from public repositories.

The authors evaluate the framework on 40 GEO studies, showing that domain-specific workflow support improves human-labelled spatial alignment. They use the proposed framework to create HESRT, a curated resource comprising 4,468 samples from 507 studies -- publicly available on HuggingFace.

Critical strengths
  • The paper addresses an important, timely, and practical problem. Public spatial transcriptomics datasets are highly heterogeneous; thus, their harmonisation can require considerable manual effort. The proposed framework is reasonably well motivated. Combining general-purpose coding agents with specialised domain tools and diagnostic feedback is a sensible approach.

  • The improvements in visual alignment are meaningful, particularly considering the heterogeneous nature of studies.

  • The released HESRT resource is an interesting and potentially valuable contribution in itself.

Critical weaknesses
  • SpatialDataAgent mainly integrates existing LLM coding capabilities with processing tools and relatively simple quality-control procedures. While this is a legitimate systems contribution (one of the accepted kinds of contribution for this conference), I am not entirely convinced that the experiments sufficiently establish its advantages over alternative approaches. For instance, all configurations use the same underlying agent, and no ablations test other agentic systems or conventional data-processing pipelines.

  • Each study is processed only once under each experimental condition. Although the authors appropriately use study-level bootstrapping, this does not capture variability across independent executions of the stochastic agent.

  • I think that HESRT is potentially one of the --if not the-- strongest contributions of the paper. However, its evaluation mainly focuses on scale, coverage, and geometric consistency. It would be interesting to see stronger evidence of its quality and usefulness for downstream applications.

  • While the main paper is generally readable, the appendix is considerably less polished and, in several places, reads more like automatically generated documentation than a scientific supplement. It contains repetitive descriptions, undefined terminology, and frequent references to supplementary files in place of sufficiently detailed methodological explanations. Given that much of the experimental and implementation evidence is reported in the appendix, I believe this is a significant weakness rather than a purely editorial issue.

Questions for authors

-Q1: Could the authors repeat a representative subset of the reconstruction experiments across multiple independent runs?

-Q2: Could the authors compare SpatialDataAgent against at least one additional baseline, such as a coding agent equipped with standard spatial transcriptomics libraries and documentation, or an alternative agentic framework?

-Q3: Could the authors demonstrate the practical utility of HESRT on a representative downstream task, such as histology-to-expression prediction or cross-study representation learning?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.