acceptodds
Under review as a conference paper at ICLR 2027

Sparse RNA Prompts Calibrate Spatial Gene Expression from Histology

Abstract

Histology-based models of spatial gene expression may capture within-slide variation while remaining miscalibrated across donors. We evaluated whether sparse same-slide RNA measurements reduce this cross-donor error and whether any improvement reflects local spatial adaptation or slide-level calibration. A frozen UNI2-h–Ridge predictor was evaluated on 70 slides from 47 donors across nine HEST cohorts. For each slide, 20 deterministic spatial partitions defined immutable target fields of view (FOVs) and disjoint prompt FOVs. We compared a no-prompt baseline with scalar, gene-wise, and low-rank empirical-Bayes offsets. At requested budget (mean achieved fraction, 5.94%; range, 4.97–14.48%), the prespecified low-rank comparison reduced donor-macro variance-normalized log-MSE from 1.441 to 1.239 (absolute difference, -0.202; paired bootstrap 95% interval, ), with improvement for 40 of 47 donors. Gene-wise offsets performed best, reducing normalized log-MSE to 1.200 ( relative to the no-prompt baseline) and mean Poisson deviance from 31.90 to 26.33 (), with improvement for all 47 donors. A scalar offset achieved normalized log-MSE of 1.206 and therefore recovered 97.3% of the gene-wise reduction relative to the no-prompt baseline; the low-rank estimator had 2.7% higher normalized log-MSE than the scalar estimator. Within-slide gene and program correlations were unchanged because each estimator adds a spot-invariant vector. These findings support same-slide calibration of expression levels, but not recovery of donor-specific spatial structure or identification of a uniquely technical source of the residual offset.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Spatial Transcriptomics, Computational Pathology, Model Calibration, Active Acquisition, Multimodal Learning

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_ZvD9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work evaluates the calibration of deep learning frameworks for the task of virtual ST prediction: predicting ST expression from H&E histology. The study proposes an evaluation protocol where each slide is partitioned into a number of disjoint FOV and selected FOVs are used to learn the calibration needed in the target FOVs. The strategy is evaluated on the HEST1k dataset across 70 slides from 47 donors and tested using several calibration ladders: (1) no prompt, (2) a scalar offset, (3) gene-wise empirical-Bayes offsets, (4) low-rank empirical-Bayes offsets. Gene-wise calibration reduces normalized log-MSE by 16.7%, while a single scalar offset recovers 97.3% of this gain.

Critical strengths
  • The study evaluates a well-known feature of ST that is that batch effects are present across slides.
  • Inducing prompt-based calibration reduces NMSE while preserving Pearson and Spearman correlations, as explained in Proposition 1.
Critical weaknesses
  • Unrealistic acquisition scenario. Prompting is simulated by masking FOVs within a fully profiled slide. Current sequencing- and imaging-based platforms (e.g., Visium) do not allow selected groups of ~50 spots to be profiled on demand, as the authors acknowledge in their Limitations. A more realistic setting would estimate the calibration on one fully profiled section and apply it to other H&E-only sections from the same donor, or use the profiled capture area to calibrate predictions on the rest of the whole slide image (e.g; Novae, Nat. Methods 2025; iSCALE, Nat. Methods 2025).
  • Limited methodological contribution. The correction is a spot-invariant offset, and a single scalar captures 97.3% of the gain, so the method is effectively slide-level normalization.
  • Missing baselines. The method is only evaluated on spot-based technologies on a single foundation model (UNI2-h) and a single prediction strategy (Ridge predictor). ST (both, sequencing-based and imaging-based technologies) has evolved towards single-cell resolution in the recent years, and virtual ST models have now transitioned to this higher spatial resolution (GHIST, Nat.Methods 2025; iSTAR, Nature Biotechnology, 2024; sCellST, Nat. Communications, 2025).
Questions for authors
  • Q1. Can the calibration mechanism be estimated for example on the leftmost portion of a whole-slide image and applied to the right-most section? Similarly, could the top portion of the FOV be masked and used for calibration of the bottom part? This would be a more realistic scenario than randomly selecting FOV of ~50 spots and masking them.

  • Q2. Can the calibration ladder be evaluated across other ST prediction mechanisms? Currently, it is only assessed on a single foundation model (UNI2-h) and a single prediction strategy (Ridge predictor). How the calibration strategy adapts to virtual ST prediction methods such as TRIPLEX (CVPR, 2024), STNet (Nat. BME, 2020) or BLEEP (NeurIPS, 2023)?

  • Q3. Can the proposed calibration strategy be extended to single-cell ST modalities where virtual ST is currently more common?

Reviewer_Re1EOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates whether a small fraction of RNA measurements from an unseen slide can improve a frozen histology-based predictor via slide-level calibration. Using disjoint prompt and target regions across multiple HEST cohorts, the authors compare scalar, gene-wise, and low-rank empirical-Bayes offsets. They show that a modest sampling budget with gene-wise calibration meaningfully reduces prediction error, though by construction such slide-level shifts leave within-slide spatial rankings unchanged.

Critical strengths
  1. The paper is refreshingly simple and isolates an overlooked phenomenon: how much cross-slide prediction error stems purely from baseline shifts. In more complex, trainable models, this effect is easily absorbed and confounded with spatial adaptation, so disentangling it cleanly is a neat and useful insight.
  2. The evaluation carefully separates prompt and target regions using donors as the statistical unit, and clearly establishes the boundary of its result—showing that constant offsets improve level-sensitive errors without improving within-slide spatial rankings.
Critical weaknesses
  1. Methodologically, the paper is almost too simple. The statistical estimator is routine, and while the post-hoc calibration effect is a nice observation, it offers limited conceptual depth or practical guidance for designing next-generation spatial transcriptomics models.
  2. Empirical generality is narrow: all experiments rely on a single frozen UNI2-h Ridge predictor and 50-gene panels across cohorts that were already part of the broader methodological investigation, leaving open whether the same behavior holds for stronger predictors or untouched external data.
  3. The scalar estimator uses a nonstandard variance divisor, as noted in the limitations, yet its impact on the central claim that a single scalar offset recovers nearly all of the gene-wise gain is not quantified.
Questions for authors
  1. What happens when scalar calibration uses the conventional variance scaling described in the limitations? Does the main conclusion comparing scalar and gene-wise offsets remain unchanged?
  2. Does this calibration behavior persist with stronger, non-linear or fine-tuned predictors, and how do the authors see this finding informing future model design beyond post-hoc offset correction?
Reviewer_pDerOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The major idea behind this paper is that existing H&E to transcriptome prediction algorithms are sub optimal on new slides and are slightly biased, therefore having elevated overall error. The proposed solution is to perform a small set of transcriptomic profiling on a subset of areas on the new slide for which predictions would like to be made. Then, an emprical bayes-inspired procedure is introduced to provide slide level corrections to shrink the new predicted estimates. This is evaluated by essentially retrospective simulations involving non-overlapping tiles from existing spatial transcriptomic data, leading to around a 15-20% reduction in overall error on new slides.

Critical strengths

The major strength of this paper is the novelty of introducing relatively formal and "classic" statistical methodology (empirical bayes style procedure) to sit on top of an output from foundation model embeddings. This is the first time I have seen the two integrated.

The limitation section is comprehensive and honest which is nice to see.

Critical weaknesses

I think the critical issue with this work is I cannot see how the setup is ever likely to be a problem in practice. The idea is one would acquire spatial transcriptomics on a focussed set of regions on a new slide to get a correction factor to improve predictive performance on the remaining regions. But all existing technologies to my knowledge (e.g. Visium, Xenium) measure the full slide at once, so you would have to somehow alter the overall protocol to make it worse for this setup to be realistic. There are some technologies like GeoMx that do profile one region at a time, but these are not close to single cell regions so it is not clear if the proposed methodology would work in practice and this is also not addressed in the paper. Then there is the overall practical consideration that prediction from H&E is far from perfect, and so the idea you would go half way profiling a slide to get a ~16% reduction in error on the remaining predictions that are very far from perfect just doesn't seem realistic. And even if this were all addressed, there is the limitation (that to be fair the authors address head on) that it's an overall slide adjustment, which would have no impact on calling expression high/low regions of a slide which would be the typical output of such a workflow.

Relatedly, there is a very large set of literature predicting transcriptomics from h&e (see top recommendations here: https://arxiv.org/abs/2609.05956). It is unclear the extent to which picking something state-of-the-art rather than ridge regression on UNI embeddings would change the conclusion.

The paper is all over deeply confusing to read, peppered with LLM lingo yet no AI use statement. What is a "prompt set P"? What actually are A_0, A_1, A_2 etc? These are never defined.

It's unclear what the actual expression values predicted are. A section of the related work covers count normalization, but then the prediction model is set up on "raw counts". Given raw count variation is determined not just by biology but also sequencing depth and variation in assay specific chemistry, it could be the slide to slide bias being corrected is just sequencing depth. What happens if one uses log CPM? Or scaled log CPM? Do the issues go away?

Questions for authors

See the main comments above, but largely:

  • what happens if you use logCPM or scaled logCPM instead of raw counts?
  • can you include other prediction model baselines with which to assess the magnitude of the correction?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.