acceptodds
Under review as a conference paper at ICLR 2027

SCOR: A Section-Centered Objective for Gene Expression Prediction from Histology

Abstract

Predicting spatial gene expression from hematoxylin and eosin (H&E) histology is typically optimized with spot-wise regression, even though many spatial observations are grouped within the same tissue section. Viewing the task as grouped regression, conventional mean squared error (MSE) decomposes exactly into within-section residual variation and section-mean residual error. This exposes an implicit objective-design choice: MSE assigns equal weight to error components defined at different levels of the grouped data structure. We introduce the Section-Centered Objective for Regression (SCOR), which retains full weight on within-section residual variation while explicitly controlling the contribution of the section-mean residual component, without changing the histology representation, predictor architecture, or inference procedure. SCOR further admits a profiled regularized section-offset interpretation that gives its weighting parameter a direct shrinkage meaning. Controlled diagnostics show that preserving the section-defined grouping is important for reproducing the benefit: pairwise residual differencing alone is insufficient, and the effect persists beyond local spatial proximity. A fixed-sample analysis further shows that training-section replication modulates its magnitude. With a single fixed objective setting, SCOR improves over MSE on the majority of eight HEST-1K tasks, transfers the same setting without retuning to HER2ST and STNet, and extends to ridge regression. These results identify grouped-error weighting as an overlooked design choice in histology-to-transcriptomics regression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computational pathology, Error Decomposition, Gene Expression, Spatial Transcriptomics

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

This paper frames the prediction of spatial gene expression from H&E images as a grouped regression problem. The spot-wise Mean Squared Error (MSE) can be precisely decomposed into two equally weighted components, namely the intra-slice residual variance (VV) and the slice-mean residual (MM). The authors propose SCOR, utilizing V+cMV + cM as the training objective, where the weight cc controls the contribution of the slice-mean residual, and demonstrate that this is mathematically equivalent to applying a ridge penalty to the bias term of each slice before eliminating it; they also provide formulations for mini-batch training and intra-slice pairwise representations. Experiments employed frozen UNI2-h features alongside MLP or ridge prediction heads; the hyperparameter cc was set to 1/651/65 based on the ccRCC dataset and subsequently kept fixed. The method was compared against MSE, Direct PCC, MISO, and STRank across eight tasks in the HEST-1K benchmark, as well as on the HER2ST and STNet datasets. Additionally, the underlying mechanism was analyzed through experiments involving intra-slice and cross-slice pairings, pairwise distance binning, and varying the number of training slices.

Critical strengths
  1. Motivated by a specific problem, the proposed method is concise and theoretically grounded. The paper points out that spot-wise MSE implicitly assigns equal weight to intra-slice errors and slice-mean errors. SCOR explicitly controls the latter using a weighting parameter; it is equivalent to applying a ridge penalty to slice biases without altering the representation, model architecture, or inference process. This offers a clear characterization of a training objective choice that is often overlooked, and the method is "plug-and-play."

  2. The empirical evidence is rigorously controlled and goes beyond merely optimizing for evaluation metrics. All objective functions were compared using identical frozen UNI2-h features, the same prediction head, and consistent data splits and training budgets. The hyperparameter cc was selected based on the ccRCC dataset alone; transferring this setting to HER2ST and STNet without re-tuning still yielded performance improvements. Diagnostic experiments (examining intra-slice versus cross-slice pairings, pairing distances, and the number of training slices) validated the underlying mechanism. Total MSE decreased in 7 out of 10 tasks, while slice-mean errors remained largely unchanged near the selected cc value (Table 5); this indicates that the gains did not simply stem from reducing error components to which PCC is inherently insensitive. The experimental work is robust.

Critical weaknesses

The scope of validation is limited and insufficient to support the conclusion that this is a universal design choice that has been overlooked.

The experiments utilized only a single frozen encoder (UNI2-h, reduced to 256 dimensions via PCA) and two types of prediction heads (MLP and ridge regression); HER2ST and STNet served merely as external datasets rather than as additional models for comparison. While leaving representations and model architectures unchanged is a key selling point of the paper, SCOR was not validated on existing MSE-based training methods, such as ST-Net, HisToGene, Hist2ST, or TRIPLEX, nor was it tested in an end-to-end encoder fine-tuning setting. End-to-end training typically necessitates smaller batch sizes, resulting in few spots per slice within a single batch. According to the analysis in Appendix A.1, the effective coefficient for intra-slice terms drops significantly under these conditions; in the limiting case of a single spot per slice, SCOR degenerates into a scaled version of standard MSE.

The scope of comparison is also limited. The prediction targets in HEST-1K were not normalized for library size (Section 3), meaning that variations in sequencing depth across slices could introduce technical mean shifts. However, the paper fails to compare SCOR against the simplest alternative approaches, namely MSE with slice-wise centering or standardization of targets, and PCC loss calculated on a per-slice basis. Given the current validation scope, the average pooled PCC on HEST-1K improved by only 0.0046 (with gains in 5 tasks, no change in 1, and declines in 2), and the performance gap relative to MISO remains marginal.

Consequently, the existing evidence demonstrates only a modest improvement for SCOR under this specific setting; it does not yet establish the method as a universal design choice. Furthermore, the authors have made no commitments or statements regarding code release; while not strictly obligatory, I would strongly prefer that the authors provide code to allow for verification.

Questions for authors

Please compare SCOR with the following under identical settings, namely MSE calculated after slice-wise centering or standardization of targets, and slice-wise PCC loss. Additionally, please clarify whether Direct PCC is calculated across slices within a minibatch.

Please evaluate SCOR on several existing MSE-based training methods, such as ST-Net, HisToGene, Hist2ST, or TRIPLEX, including at least one setup involving end-to-end encoder fine-tuning. End-to-end training typically employs smaller batch sizes; as noted in Appendix A.1, this results in fewer spots per slice within a batch, significantly reducing the effective coefficients for intra-slice terms. Please report the performance gains achieved by SCOR in this setting and indicate whether batch organization by slice is required. Also, please test other FM embeddings using the HEST-1K framework.

If SCOR’s advantages remain evident after the authors provide the aforementioned validations, I will consider raising the score to 3.

Reviewer_kM2WOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies spatial gene expression prediction from hematoxylin and eosin (H&E) histology and proposes the Section-Centered Objective for Regression (SCOR). The method decomposes the conventional mean squared error into within-section residual variation and section-mean residual error, and explicitly controls the relative weight of the latter component. The paper provides a profiled section-offset interpretation and analyzes the minibatch approximation of the objective.

Critical strengths

The mathematical formulation of SCOR is clearly presented, including the within-section/section-mean decomposition, the profiled offset interpretation, and the minibatch approximation analysis.

The empirical study spans multiple HEST-1K tasks and two external datasets, and compares SCOR with several objective-level baselines under both nonlinear and linear predictors.

The paper provides relatively complete implementation details, including the datasets, preprocessing procedures, model configurations, training hyperparameters, and evaluation settings.

Critical weaknesses

Incomplete characterization of minibatch dependence. SCOR’s effective weighting depends on within-batch section occupancy msm_s, but its distribution and induced weights are not reported for the main experiments. Controlled comparisons of batch compositions, while preserving overall objective weighting, would clarify sensitivity to the chosen sampling scheme.  Limited statistical support for modest benchmark gains. On HEST-1K, average pooled PCC improves by only 0.0046 over MSE and 0.0032 over MISO. Paired effect estimates and confidence intervals accounting for repeated seeds and fold dependence are needed to assess the robustness of these gains.  Limited isolation of the grouping mechanism. Although the matched diagnostics control several training factors, section and patient identities coincide in ccRCC. Increasing section count also changes patient diversity, spots per section, and minibatch occupancy. Size-matched random-group controls and section-versus-patient grouping comparisons would help identify which grouping properties drive the benefit.  Limited validation across predictor architectures. SCOR is presented as an objective that requires no architectural changes, yet evaluation is restricted to MLP and ridge regression. Applying it to established spatial-context regression models, such as Hist2ST or TRIPLEX, under matched MSE-versus-SCOR comparisons would test whether its gains extend beyond simple predictors.

Questions for authors

see weakness

Reviewer_BrjcOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a new training objective for the H&E-to-ST prediction task. In section-grouped data, traditional MSE naturally incorporates two equal-weighted components: (i) M: the average section-level error, and (ii) V: the relative pattern error within each section. In the context of ST data, M dominates training and pushes the model to ignore the within-section patterns. This can easily lead to pitfalls such as the model trying to learn technical noise or to learn shortcuts like stain intensity which might identify a whole slide rather than biology. The paper introduces the Section-Centered Objective for Regression (SCOR), L = V + c$\times$M, with c ∈\in [0,1]. SCOR down-weights the average offset across the whole section. In SCOR’s formulation, c = 1 recovers MSE and c = 0 yields a fully offset-invariant loss. Experiments show that SCOR is equivalent to profiling out ridge-penalized per-section offsets c=α1+αc = \frac{\alpha}{1 +\alpha}. A minibatch estimator is provided, whose expected weight for M stays c. Experiments also include a same-section pairwise representation which recovers SCOR. A single setting of c=165c = \frac{1}{65} is selected on ccRCC and an MLP or ridge regressor on frozen UNI2-h features improves PCC on eight Hest1k tasks. It is also evident that performance transfers to HER2ST and STNet datasets without re-tuning.

Critical strengths
  1. The criticism of MSE as an objective is well-motivated. The fact that ST spots are nested within tissue sections, and that this might lead to training being dominated by over-penalized section-level offsets or shifts, is a real and often-ignored issue. It might be difficult to ascertain batch effects, sequencing depth and other variables from the H&E. This would lead the model astray, where it tries to learn noise or take cheap and incorrect shortcuts like tissue stains or other artefacts that identify sections.

  2. SCOR has a clear decomposition and is an easy plug-in. The V+M identity is mathematically exact. It is also easy to follow and converts an underlying modeling choice into a tunable parameter. Implementing SCOR would be equivalent to adding a few lines of training code to any standard training pipeline.

  3. Zero re-tuning transfer to external datasets is demonstrated. The SCOR setting fixed on the ccRCC cohort from Hest1k is shown to be transferable to HER2ST and STNet datasets. This is good generalization evidence.

  4. Stochastic implementation is thoroughly analyzed. Appendix A.1 details how stochastic implementation of the section-offset weight is a closer approximation of the full-section objective. Given γs=1−(1−c)(ns−ms)ms(ns−1)\gamma_s = 1 - \frac{(1-c)(n_s - m_s)}{m_s(n_s-1)}, when a mini-batch samples ms=1m_s = 1 spot, ys=cy_s = c and when it samples ms=nsm_s = n_s or the whole set of spots from the section, ys=1y_s = 1. That means, as more and more observations from the same section are sampled in minibatch, the weight approaches 1 and approximates full-section SCOR.

  5. Section-offset is significant. Fig. 3a shows that RMR_M varies from 20 to 63% across tasks under MSE. This demonstrates that the component of the loss that is reweighted has a significant contribution during training.

  6. Transparency is commendable. Authors clearly report their losses on the SKCM, LUAD and LYMPH tasks, as well as MISO’s section-level win on HER2ST. They also transparently present the large increase in held-out M at c = 0. Section/patient confounding is acknowledged and the c-sweep is labeled as post-hoc.

Critical weaknesses

W1. The training objective is not novel. Credit is due.
The objective is Eigen et al’s grouped scale-invariant loss (NeurIPS 2014). It is equivalently related to the MiDaS style scale and shift-invariant loss. These predecessors should be attributed in the manuscript. However, this does not take away the novel application of the objective to the specific context of H&E-to-ST prediction and interpreting section-level errors in this fashion.

W2. The training objective is aligned with the evaluation metrics.
Section-level gene-wise PCC ignores per-section additive offsets and scaling. Downweighting M cannot reduce PCC by introducing mean-errors anyway. This leads me to question what chunk of the reported gains are purely mechanical by design, and inevitable with the alignment of the training objective with the evaluation metric.

W3. Some simple baselines need to be added for validation.
Since the single setting of c=165c = \frac{1}{65} is fixed across all experiments, SCOR is effectively close to full centering, with a small anchor. Some obvious baselines I can think of in this context are:
- MSE on per-section mean-centered or z-scored transcriptomics
- Explicitly learned per-section offsets, which would be dropped at test time
- Batch correction using ComBat/Harmony and then training on MSE

W4. Gains exist, but are small and untested for significance.
Hest1k average improvement over MISO is 0.0032 in PCC, and HER2ST pooled margin is 0.0006. These are not very big improvements when it comes to Pearson Correlation. Several of the Hest1k tasks seem to be in one seed's worth of standard deviation from the next best. For example, PRAD yields 0.3973±0.00880.3973 \pm 0.0088 vs 0.3872±0.00920.3872 \pm 0.0092, READ yields 0.2511±0.01530.2511 \pm 0.0153 vs. 0.2453±0.00980.2453 \pm 0.0098. Many ridge deltas are 0.0000-0.0028. It is difficult to buy into these values being stated as evidence for improvement.

W5. Hyperparameter tuning and checkpoint selection.
The value of c is selected on a validation-sweep on ccRCC, but ccRCC is also reported as a Hest1k task where SCOR wins. This can be classified as leakage. Furthermore, checkpoint selection is based on the highest pooled validation PCC. W2 already discusses how pooled PCC favors SCOR because of the alignment between the objective and the evaluation metric. This feeds into SCOR’s bias.

W6. Architecture scope seems narrow.
Since SCOR is a loss-function paper, I would expect it to demonstrate robustness even when combined with stronger models, specially the ones that have context-aware architectures (BLEEP, TRIPLEX, MERGE, Stem, STFlow, etc.)

W7. Section-level shifts are not always technical noise.
While the proposition that section-level offsets or shifts are technical noise that dominates the training objective in traditional MSE, it is not always noise. There could be a number of biologically salient reasons for which the shifts can occur, such as tumor content, immune infiltration, cancer subtypes, etc. These traits matter for downstream objectives such as cross-sample comparison, biomarker thresholds, pseudo-bulk prediction, outcome prediction, etc. The authors do acknowledge this in principle, but do not test it. Meaningful additions would include pseudo-bulk prediction, outcome prediction or some other tasks.

Questions for authors
  1. Add the obvious simple baselines, add stronger context-aware model baselines and use other foundation model patch encoders.
  2. Report MSE, MAE, R2R^2, V and M separately and calibrate for all tasks and methods. Trade off should explain how much absolute accuracy SCOR sacrifices to boost correlation.
  3. Add significance tests and perform patient-level cross validation.
  4. Don’t claim a win on ccRCC since it guided the selection of c=165c=\frac{1}{65}.
  5. Present analysis on whether SCOR sacrifices section-level offsets or shifts that are biologically salient features of the samples and could be important for downstream use cases.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.