acceptodds
Under review as a conference paper at ICLR 2027

MoSPR: Histology-to-Gene Expression Prediction with Morpho-Spatial Macrostates and Low-Rank Molecular Programs

Abstract

Predicting molecular profiles from histopathology remains challenging because whole-slide images contain spatially organized, heterogeneous tissue patterns, while gene expression comprises thousands of correlated targets. We introduce MoSPR (Morpho-Spatial Program Regression), a linear framework that couples an adjacency-informed histology representation with a low-rank molecular basis. MoSPR clusters frozen patch embeddings into morphology microstates, aggregates their spatial adjacencies across the training cohort, and groups microstates with similar adjacency patterns into shared macrostates. Each slide is then represented by global morphology and macrostate-specific deviations, which are linearly mapped to coefficients of a training-derived low-rank gene-expression basis. Across three cancer cohorts from The Cancer Genome Atlas, MoSPR achieves the highest mean gene-expression prediction scores among all evaluated methods. Without pathway-level supervision, pathway scores derived from its predicted expression profiles rank first in eight of nine comparisons across three pathway collections. Ablation studies on the breast cancer cohort show complementary gains from adjacency-derived macrostate representation and low-rank molecular prediction. Moreover, with half of the training data on this cohort, MoSPR exceeds the full-data gene-prediction score of the strongest competing baseline. Finally, its linear formulation enables exact decomposition of each predicted expression profile into global and macrostate-specific molecular contributions, providing an interpretable link between spatially coherent macrostate regions and their associated molecular programs. Our code is available at https://anonymous.4open.science/r/MoSPR-C957/README.md.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
computational pathology, whole-slide images, gene expression prediction, morpho-spatial representation learning, low-rank regression, data-efficient learning, spectral clustering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_TZSrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces MoSPR, a lightweight framework for predicting bulk gene expression from whole-slide histopathology images. The method first clusters frozen CONCH patch embeddings into morphology “microstates”, summarizes their spatial co-occurrence through patient-level and cohort-level adjacency matrices. Each WSI is represented using a global mean feature together with abundance-weighted macrostate-specific deviations. On the molecular side, the target gene-expression matrix is projected onto a PCA-derived low-rank basis, and ridge regression maps the WSI representation to the corresponding molecular coefficients. Owing to the block-structured linear formulation, predictions can be decomposed into global and macrostate-specific contributions. The method achieves strong performance across three TCGA cohorts and shows favorable data efficiency.

Critical strengths

The method is simple, computationally efficient, and appears well suited to the relatively limited sample sizes available for paired WSI–RNA data. The experiments are reasonably comprehensive, including multiple TCGA cancer cohorts, pathway-level evaluation, training-set scaling, computational complexity, and component ablations.

Critical weaknesses
  1. The methodological novelty appears limited and should be positioned more carefully. Most individual components of MoSPR are well-established techniques: frozen pathology foundation-model features, k-means-based morphology prototypes, spatial patch adjacency, diffusion-map/spectral clustering, PCA-based dimensionality reduction, and ridge regression.
  2. The low-rank molecular component is largely standard PCA regression. The second major contribution claimed by the paper is a “training-derived low-rank expression basis”. However, the proposed procedure consists of fitting PCA to the standardized gene-expression matrix, regressing histology features onto the PCA coefficients using ridge regression, and reconstructing the output through the PCA basis.
Questions for authors
  1. The manuscript suffers from substantial writing and presentation issues. Overall, it reads more like a quickly assembled technical report than a carefully polished conference paper. The abstract is poorly balanced. Roughly two-thirds of the abstract is devoted to experimental results, while the motivation, research problem, and core methodological idea receive comparatively limited explanation. The introduction lacks a clear problem-driven narrative. It should better explain why H&E-to-bulk gene expression prediction is important, what the key challenges are, and where existing methods fall short, rather than mainly presenting a sequential description of the proposed pipeline. The introduction is overly technical for its purpose. As a computational pathology paper, the introduction should emphasize motivation, challenges, intuition, and introducing detailed mathematical formulations at this stage unnecessarily hurts readability. The overall readability is weak. Many sections are written in a technical-report style, with insufficient high-level explanation and transitions between methodological components.

  2. Equation (20) is poorly formatted. The block partition of the regression matrix would be clearer.

  3. The Related Work section is incomplete. It does not adequately review the broader literature on H&E and bulk gene expression, including works on multimodal pathology-gene fusion and H&E-gene contrastive learning.

  4. The relationship to prior histology-to-gene-expression methods is insufficiently discussed. The manuscript mainly lists several prediction baselines but does not clearly position MoSPR relative to different methodological paradigms in this area.

  5. The figure quality is insufficient. Several figures are difficult to read because of low resolution and small visual elements, which significantly affects presentation quality.

  6. The methodological novelty of the low-rank molecular prediction component is limited. Projecting high-dimensional and strongly correlated gene-expression targets into a low-dimensional space before regression is a natural and widely used strategy.

  7. The actual methodological novelty should be stated more narrowly. The most distinctive part seems to be the construction of cohort-level adjacency-informed morphology macrostates, rather than the overall low-rank regression framework.

3 comments
AuthorsOpenReview ↗

Response to Reviewer TZSr (1/3): Methodological novelty

We thank the reviewer for the candid and detailed critique, which has been very helpful in identifying weaknesses in our presentation and clarifying the scope of our methodological contributions.

W1. The methodological novelty appears limited and should be positioned more carefully. Most individual components of MoSPR are well-established techniques: frozen pathology foundation-model features, k-means-based morphology prototypes, spatial patch adjacency, diffusion-map/spectral clustering, PCA-based dimensionality reduction, and ridge regression.

We thank the reviewer for this comment, and we agree that each component of MoSPR is well established. Our intention was not to introduce a new learning algorithm, but to assemble these components into a design suited to a specific setting: predicting thousands of correlated genes from whole-slide images when paired histology–transcriptomic cohorts contain only hundreds of patients.

The central design choice is to describe all patients in one shared frame of reference rather than learning a representation for each patient. MoSPR first builds, without supervision, a single cohort-level vocabulary of recurring tissue states. Every slide is then summarized by the abundance of each state and by how that state's morphology departs from the slide average. Because this structure is fixed before supervision, the only part learned from expression labels is a single linear map onto a low-rank expression basis.

We chose this design to address three practical limitations of patient-specific approaches such as MIL, Transformer- and graph-based models:

  1. Data efficiency. These models must learn how to aggregate each slide's patches from slide-level labels alone, which requires many parameters relative to the available cohorts. In MoSPR, the aggregation is determined without labels, so supervision is spent only on a compact linear mapping.
  2. A common reference across patients. Attention weights or graphs differ from slide to slide, so learned representations are not aligned across patients. In MoSPR, every patient is expressed over the same tissue-state vocabulary, which makes representations directly comparable and lets each prediction be decomposed additively.
  3. Computational cost. MoSPR requires 0.96M parameters and 2.02 GFLOPs per slide, compared with 1.5–20.7M parameters and up to 516 GFLOPs for the baselines (Table S3).

We have revised the Introduction to state this design rationale explicitly and to position MoSPR more carefully relative to existing prototype- and graph-based approaches.

W2. The low-rank molecular component is largely standard PCA regression. The second major contribution claimed by the paper is a “training-derived low-rank expression basis”. However, the proposed procedure consists of fitting PCA to the standardized gene-expression matrix, regressing histology features onto the PCA coefficients using ridge regression, and reconstructing the output through the PCA basis.

We thank the reviewer for this accurate description, and we agree that the low-rank component is a standard PCA-based (reduced-rank) regression. We have revised the manuscript so that it is no longer presented as a separate methodological contribution, but as a deliberate design choice within the overall framework.

Its role follows from the setting described in our response to W1. When thousands of strongly correlated genes must be predicted from only hundreds of patients, restricting the output to the main directions of variation in the training expression matrix regularizes the regression and stabilizes estimation. Within MoSPR, this choice is beneficial: with the same inputs, predicting low-rank coefficients instead of each gene directly improved gene-level SCC on BRCA, both for the global representation (0.348 → 0.380) and for the full MoSPR representation (0.395 → 0.411; Table 3).

At the same time, the low-rank output alone does not explain MoSPR's performance. When we applied the same low-rank output to two representative baselines, either by retraining them on the low-rank targets or by projecting their predictions onto the same basis, they remained well below MoSPR in all three cohorts (Table S4; e.g., BRCA: ABMIL 0.354–0.380 and tRNAformer 0.327–0.338 vs MoSPR 0.411). The benefit therefore arises from combining a compact, linear state-structured representation with a low-rank output, rather than from the low-rank regression itself. We now state this explicitly and refer to the PCA components as a low-rank expression basis rather than as molecular programs.

AuthorsOpenReview ↗

Response to Reviewer TZSr (2/3): Presentation and related work

Q1. The manuscript suffers from substantial writing and presentation issues. Overall, it reads more like a quickly assembled technical report than a carefully polished conference paper. The abstract is poorly balanced. Roughly two-thirds of the abstract is devoted to experimental results, while the motivation, research problem, and core methodological idea receive comparatively limited explanation. The introduction lacks a clear problem-driven narrative. It should better explain why H&E-to-bulk gene expression prediction is important, what the key challenges are, and where existing methods fall short, rather than mainly presenting a sequential description of the proposed pipeline. The introduction is overly technical for its purpose. As a computational pathology paper, the introduction should emphasize motivation, challenges, intuition, and introducing detailed mathematical formulations at this stage unnecessarily hurts readability. The overall readability is weak. Many sections are written in a technical-report style, with insufficient high-level explanation and transitions between methodological components.

We thank the reviewer for this detailed and constructive feedback on presentation, and we agree with these points. We have rewritten the Abstract so that it focuses on the motivation, the research problem, and the core idea of MoSPR, with the results summarized in two sentences. We have also rewritten the opening of the Introduction around a problem-driven narrative: why predicting gene expression from H&E slides is useful, what makes it difficult (heterogeneous tissue, thousands of correlated genes, and paired cohorts of only hundreds of patients), and why existing methods are data-hungry in this setting. We have also restated our contributions more concisely. We agree that the remaining technical description in the Introduction should be simplified, and that the Methods section needs more high-level explanation and transitions between components. We will address both in a future revision.

Q2. Equation (20) is poorly formatted. The block partition of the regression matrix would be clearer.

We thank the reviewer for this helpful suggestion. We have rewritten Eq. (20) so that each patient's representation, x_n = [M_n | s_n1 | … | s_nK], and the fitted regression matrix are partitioned into matching blocks. The resulting sum, x_n Wq = M_n W{q,global} + Σ_k s_nk W*_{q,k}, is now shown in the same equation. This makes explicit which block of the regression matrix acts on the global representation and which acts on each macrostate, and it leads directly to the macrostate-level terms in Eq. (21).

Q3. The Related Work section is incomplete. It does not adequately review the broader literature on H&E and bulk gene expression, including works on multimodal pathology-gene fusion and H&E-gene contrastive learning.

We thank the reviewer for pointing this out, and we agree that the Related Work section did not adequately cover the broader literature on histology and gene expression. We have revised it to include both lines of work the reviewer mentions. Multimodal pathology–gene fusion methods, such as MMP (Song et al., 2024b), take both histology and genomic profiles as inputs, typically to predict clinical outcomes such as survival. H&E–gene contrastive methods, such as TANGLE (Jaume et al., 2024), use paired bulk expression as a training signal to learn slide representations, which are then used for slide-level tasks such as classification and retrieval. In both cases, expression serves as an input or a training signal, whereas in MoSPR it is the prediction target, inferred from histology alone. These approaches are therefore complementary to ours; for example, slide representations learned by contrastive pretraining could serve as input features to MoSPR.

AuthorsOpenReview ↗

Response to Reviewer TZSr (3/3): Positioning, figures, and contribution scope

Q4. The relationship to prior histology-to-gene-expression methods is insufficiently discussed. The manuscript mainly lists several prediction baselines but does not clearly position MoSPR relative to different methodological paradigms in this area.

We thank the reviewer for this comment and agree that the previous Related Work section listed methods rather than positioning MoSPR among methodological paradigms. We have reorganized it around how each approach obtains its slide representation.

Most existing histology-to-expression methods learn how to aggregate each slide's patches from slide-level expression labels. They either predict expression per tile and pool the predictions (HE2RNA, MOSBY), or use attention-, transformer- or clustering-based aggregation (AbReg, tRNAformer, SEQUOIA). General WSI aggregators such as ABMIL, TransMIL and Mamba-based models follow the same principle, and graph-based models such as Patch-GCN additionally learn spatial interactions within each patient's patch graph. As described in our response to W1, learning both aggregation and spatial modeling from slide-level labels makes these models parameter-heavy relative to the hundreds of patients typically available. CPNN takes a different route by structuring the output, reconstructing expression from single-cell-derived cell-type prototypes with image-predicted proportions.

MoSPR is closest to unsupervised prototype-based slide representations, which summarize each slide over a set of morphological prototypes without labels. For example, PANTHER (Song et al., 2024a) fits a Gaussian mixture to each slide's patch embeddings and uses its parameters as the slide representation. Within this family, MoSPR differs in three ways:

  1. Its tissue states are defined once at the cohort level by how often morphologies occur next to each other, so spatial context is captured across the cohort rather than through patient-specific graphs.
  2. Each state is represented by its deviation from the slide average.
  3. The only supervised component is a single linear map onto a low-rank expression basis.

This positioning is now stated explicitly in the revised Related Work section.

Q5. The figure quality is insufficient. Several figures are difficult to read because of low resolution and small visual elements, which significantly affects presentation quality.

We thank the reviewer for this comment and agree that several figures are difficult to read in their current form. In a future revision, we will regenerate all figures at higher resolution, preferably in vector format, and enlarge text, labels and other visual elements so that they remain legible at the printed size.

Q6. The methodological novelty of the low-rank molecular prediction component is limited. Projecting high-dimensional and strongly correlated gene-expression targets into a low-dimensional space before regression is a natural and widely used strategy.

We thank the reviewer and agree, as discussed in our response to W2. Projecting correlated expression targets onto a low-dimensional basis before regression is a standard reduced-rank strategy. We have therefore revised the manuscript so that this component is no longer presented as a methodological contribution, and the contributions no longer list it separately.

Q7. The actual methodological novelty should be stated more narrowly. The most distinctive part seems to be the construction of cohort-level adjacency-informed morphology macrostates, rather than the overall low-rank regression framework.

We thank the reviewer for this assessment, with which we agree. The cohort-level, adjacency-informed macrostate dictionary is the most distinctive part of MoSPR, and we have revised the contributions accordingly. The first contribution is now this cohort-level representation, in which all patients are described over a shared dictionary of morpho-spatial tissue states learned once from the training cohort without supervision. The remaining points, data-efficient prediction and a lightweight model, are presented as properties that follow from this design rather than as separate methodological contributions.

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

MoSPR predicts bulk gene expression from H&E WSIs with a fully linear pipeline over frozen CONCH patch features. Patches are k-means quantized into microstates; 8-neighbor adjacency counts between microstates are normalized per patient, averaged over the training cohort, spectrally embedded, and clustered into a small number of macrostates. Each slide is encoded as the mean patch feature concatenated with abundance-weighted per-macrostate mean deviations, and ridge regression maps this to coefficients of a PCA basis fitted to the training expression matrix. On TCGA BRCA, KIRC and LUAD with four-fold patient-level CV, MoSPR outperforms fourteen MIL and transcriptomic-prediction baselines on gene-level and pathway-level correlation with a tiny parameter budget, and the paper adds component ablations with shuffled-state controls, a training-set scaling study, bootstrap partition stability, and a qualitative macrostate example.

Critical strengths
  1. The compact linear formulation is practical and yields consistent reported gains over a broad baseline set. Training-fold preprocessing and patient-level splits are clearly specified.
  2. The ablations include shuffled-state controls, baseline-specific low-rank output controls, and training-set scaling. These strengthen the evidence beyond comparisons against unmodified neural baselines.
Critical weaknesses
  1. The specific value of spatial adjacency is not isolated from morphology grouping. Once the dictionary is fixed, Eqs. (12)–(14) depend on patch features and assignments, rather than coordinates; spatial information enters only through dictionary construction. This is consistent with a cohort-informed representation, but makes a strong appearance-only grouping control essential. Shuffling the learned microstate-to-macrostate map destroys coherent morphology grouping as well as adjacency structure. Comparisons against direct morphology clustering and a dictionary relearned after within-slide coordinate permutation are needed to establish that adjacency, rather than structured feature pooling, explains the improvement.
  2. Exact decomposition does not establish identifiable macrostate-specific molecular programs. The centered blocks satisfy ∑ksnk=0\sum_k s_{nk}=0. Adding the same coefficient matrix to every macrostate block therefore leaves the total prediction unchanged while altering individual contributions Γnk\Gamma_{nk}. Ridge regularization selects a particular decomposition, but that choice is imposed by the penalty rather than identified by bulk expression. Figure 3’s selected slide and pathway enrichments do not resolve this ambiguity. The paper appropriately acknowledges the lack of spatial molecular validation, yet its interpretability contribution still requires evidence that the attributed programs are stable and biologically meaningful beyond the chosen parameterization.
  3. Robustness of the cohort-derived representation remains insufficiently tested. All predictive evaluations use within-cohort TCGA splits. Neither an independent cohort nor tissue-source-site-held-out evaluation tests whether the adjacency dictionary and molecular basis transfer beyond the acquisition and population structure used to construct them. Tables 1–3 report fold averages without uncertainty for paired method differences. The macrostate bootstrap in Table S5 measures partition stability with fixed microstates, rather than predictive robustness or stability of molecular attributions. This limits confidence in the proposed representation as a general histology-to-expression method.
Questions for authors
  1. How does MoSPR compare with morphology-only clustering and coordinate-shuffled dictionary construction under the same rank and regularization search?
  2. How stable are macrostate pathway contributions across centering choices, regularization strengths, and folds, and can any be checked against independent regional measurements?
  3. Do the gains persist under tissue-source-site-held-out evaluation, with patient-level uncertainty estimates?
4 comments
AuthorsOpenReview ↗

Response to Reviewer CqxX (1/3): W1 and Q1

We sincerely thank the reviewer for the careful and constructive review, which helped us identify limitations in our design, and for suggesting exactly the controls needed to isolate the role of adjacency.

W1. The specific value of spatial adjacency is not isolated from morphology grouping.

We thank the reviewer for this precise analysis, and we agree with it. Once the dictionary is fixed, Eqs. (12)–(14) depend only on patch features and state assignments, so spatial information enters MoSPR only through dictionary construction. We also agree that the shuffled-state control in Table 3 disrupts morphological coherence and adjacency at the same time.

We performed both controls the reviewer suggested. In each, only the dictionary construction was changed; the microstate vocabulary, the [M | S] representation, and the rank and regularization search were identical to MoSPR.

  • Appearance-only grouping: instead of grouping microstates by adjacency, we clustered their mean CONCH features directly into the same number of macrostates (K = 8, size-weighted k-means).

  • Coordinate permutation: we randomly permuted patch coordinates within each training slide before computing adjacency, and re-learned the dictionary. This preserves state composition but removes the true spatial arrangement.

Gene SCCBRCAKIRCLUAD
Appearance-only grouping0.4210.3570.386
MoSPR0.4120.3480.373
Coordinate permutation0.4110.3400.359

The appearance-only dictionary performed slightly better than our adjacency-based dictionary (+0.011 on average; higher in 56 of 60 fold–seed pairs), whereas coordinate permutation lowered performance only modestly (−0.008 on average; lower in 49 of 60 pairs). The improvement is therefore explained mainly by structured feature pooling over morphologically coherent states rather than by adjacency itself, and we acknowledge this as a limitation of the current design.

A likely reason is that our adjacency is defined by immediate (8-neighbour) contacts. Because tissue morphology is spatially continuous, immediately adjacent patches are usually morphologically similar. The adjacency graph thus largely overlaps with appearance similarity. This explains both why a dictionary built from appearance alone performs at least as well and why removing the true spatial arrangement costs only a modest amount of performance.

Q1. How does MoSPR compare with morphology-only clustering and coordinate-shuffled dictionary construction under the same rank and regularization search?

We sincerely thank the reviewer for the careful and constructive review, which helped us identify limitations in our design, and for suggesting exactly the controls needed to isolate the role of adjacency. Both comparisons were run under the same protocol as MoSPR. The microstate vocabulary, the [M | S] representation, the rank and regularization grid, and the two-stage selection (train → validation, then refit on train + validation) were identical; only the dictionary construction differed. Gene-level SCC, mean over 4 folds × 5 seeds:

Dictionary constructionBRCAKIRCLUAD
Morphology-only clustering0.4210.3570.386
MoSPR0.4120.3480.373
Coordinate-shuffled adjacency0.4110.3400.359

Morphology-only clustering was higher than MoSPR in 56 of 60 fold–seed pairs, and the coordinate-shuffled dictionary was lower in 49 of 60. Together, these results indicate that the gain of MoSPR comes mainly from summarizing each slide over morphologically coherent states, not from adjacency itself. True spatial arrangement does carry some information, since shuffling coordinates lowers performance, but with immediate (8-neighbour) contacts much of it overlaps with appearance similarity, because neighbouring patches are usually morphologically alike. We acknowledge this as a limitation of the current adjacency definition (see also our response to W1).

AuthorsOpenReview ↗

Response to Reviewer CqxX (2/3): Attribution identifiability and stability (W2/Q2)

W2. Exact decomposition does not establish identifiable macrostate-specific molecular programs.

Please see our response to Q2.

Q2. How stable are macrostate pathway contributions across centering choices, regularization strengths, and folds, and can any be checked against independent regional measurements?

We thank the reviewer for this precise point, and we agree with it. Because the state blocks are centered on the slide mean, they sum to zero across macrostates (∑ksnk=0\sum_k s_{nk}=0). Adding the same matrix to the regression coefficients of every macrostate block therefore leaves the prediction unchanged but changes each macrostate's contribution Γnk\Gamma_{nk}. The least-squares fit alone admits infinitely many such coefficient sets, and ridge regularization selects one of them; the original contributions were thus determined by the penalty rather than by the data.

We therefore performed an additional analysis using an identifiable, ablation-based definition of contribution (Δremove\Delta_{\mathrm{remove}}). For each test slide, we remove the patches assigned to macrostate kk, recompute [M | S] from the remaining patches, and predict again; the difference from the original prediction is the contribution of macrostate kk to each gene. Every slide representation, original or ablated, satisfies ∑ksnk=0\sum_k s_{nk}=0, so all equivalent coefficient sets give identical predictions for both slides. Δremove\Delta_{\mathrm{remove}} is therefore uniquely determined for a fitted model.

As the reviewer asked, we assessed stability across centering choices, regularization strengths, and folds. For each macrostate, we averaged contributions over test slides and summarized them into 50 Hallmark pathway scores. We then measured how well the pathway ranking is preserved when the setting changes (Spearman correlation, median over macrostates; BRCA / KIRC / LUAD). The three centering choices were: (i) the original slide-mean centering, (ii) anchoring all states to a reference macrostate, and (iii) centering on a cohort-level mean.

SettingΔremove\Delta_{\mathrm{remove}}Original Γ\Gamma
Centering0.92 / 0.78 / 0.940.11 / 0.20 / 0.58
Rank and regularization0.94 / 0.95 / 0.970.92 / 0.90 / 0.90
Fold0.83 / 0.75 / 0.900.75 / 0.51 / 0.65

Both definitions were stable across rank and regularization strength. Changing the centering, however, largely scrambled the rankings of Γ\Gamma, whereas Δremove\Delta_{\mathrm{remove}} preserved them. Δremove\Delta_{\mathrm{remove}} was also more stable across folds; the remaining fold-to-fold variation reflects differences in the macrostate partition itself (see our response to Reviewer 1, Q3).

We interpret these contributions only as tendencies of the fitted model within the given cohort, i.e., certain macrostates appear to contribute relatively more to the prediction of certain pathways, rather than as identified molecular programs or causal effects of the corresponding tissue regions. Validation against independent regional measurements, such as spatial transcriptomics, remains a limitation of this work.

AuthorsOpenReview ↗

Response to Reviewer CqxX (3/3): W3 and Q3

W3. Robustness of the cohort-derived representation remains insufficiently tested.

Please see our response to Q3.

Q3. Do the gains persist under tissue-source-site-held-out evaluation, with patient-level uncertainty estimates?

We thank the reviewer for raising the possibility that our gains reflect differences between tissue source sites, and for asking for uncertainty estimates. We address these points with three analyses.

Within-site evaluation. TCGA slides come from many sites that differ in staining, scanning, and patient population, and a model could exploit these differences without learning tissue biology. To test this, we re-evaluated all methods using correlations computed only among test slides from the same tissue source site. We used sites with at least eight test slides, weighted each site by its number of slides, and obtained 95% confidence intervals by patient-level bootstrap. Within a single site, staining and site-level expression differences are shared by all slides, so this evaluation measures only how well each method captures differences between patients from the same site.

BRCAKIRCLUAD
MoSPR, all test slides0.4110.3480.376
MoSPR, within site0.3710.3010.320
Rank of MoSPR within site1st1st1st
Margin over the strongest baseline, all test slides+0.047+0.029+0.039
Margin over the strongest baseline, within site+0.042+0.017+0.026
Predicting from training-site averages only0.1150.1450.144

All methods lost 0.04–0.06 SCC under within-site evaluation, indicating that every model draws partly on site-level signal. Nevertheless, MoSPR remained the best method in all three cohorts. Predicting each test slide from the average expression of its training site reached only 0.115–0.145, so most of MoSPR's accuracy does not come from site identity. The margin over the strongest baseline decreased in KIRC (0.029 → 0.017) and LUAD (0.039 → 0.026), more than in BRCA (0.047 → 0.042), but it remained positive in all three cohorts (e.g., LUAD 95% CI 0.009–0.038).

These results indicate that MoSPR's advantage is not an artifact of differences between tissue source sites. Once staining, scanning, and site-level expression differences are held constant by comparing slides only within the same site, MoSPR still predicts patient-to-patient variation in expression better than all baselines. Its predictions therefore reflect morphology–expression relationships that hold within a site, not merely the ability to recognize where a slide came from. The larger reductions in KIRC and LUAD suggest that site-related signal plays a greater role in these cohorts.

Tissue-source-site-held-out evaluation. The within-site analysis rules out site effects as the source of the gain, but test sites still appear in training. To test transfer to unseen sites directly, we held out entire tissue source sites: the dictionary, expression basis, and model were rebuilt only from the remaining sites and evaluated on slides from the held-out sites. We retrained ABMIL, the most widely used multiple-instance learning baseline and the strongest competing method on BRCA, on the same splits for a paired comparison on identical test patients.

Gene-level SCC difference on held-out sitesBRCAKIRCLUAD
MoSPR - ABMIL+0.042+0.037+0.056
MoSPR - global-only model+0.027+0.014+0.014

MoSPR outperformed ABMIL in all three cohorts on unseen sites, with patient-level bootstrap intervals excluding zero, and retained its advantage over a global-only representation. All methods lost 0.02–0.05 SCC on unseen sites, and the decline of MoSPR was similar to that of the global-only model, indicating that the cohort-level dictionary transfers to new sites as well as a simpler representation does.

Patient-level uncertainty. We have revised Tables 1–3 to report uncertainty for all paired method differences. In Tables 1 and 2, the last rows give MoSPR minus the strongest baseline in each setting, with 95% confidence intervals obtained by resampling test patients (1,000 resamples per fold). At the gene level, MoSPR's margin is positive with intervals excluding zero in all six cohort–metric settings (e.g., SCC: BRCA +0.047 [0.040, 0.055], KIRC +0.029 [0.016, 0.042], LUAD +0.039 [0.028, 0.049]). At the pathway level, the margin is significant for all three collections in BRCA, whereas the intervals include zero in KIRC and LUAD, indicating performance comparable to the strongest baseline in these two cohorts. Table 3 now reports the difference between MoSPR and each ablation variant with fold-level paired t intervals (n = 4), since slide-level predictions of the variants are not available. In particular, the low-rank output improves gene-level SCC over Spatial Direct by +0.016 [−0.004, 0.036], a consistent but not statistically established gain with four folds.

Thank you for the careful rebuttal and additional experiments. Some of my concerns have been targeted. However, morphology-only grouping outperforms MoSPR across all three cohorts, and the more stable attribution does not establish region-specific molecular programs. These findings narrow the central contribution, leaving the methodological novelty beyond morphology-based pooling and low-rank regression insufficiently established. I therefore maintain my score.

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

This paper proposes MoSPR, a framework for predicting bulk gene expression from histopathology whole-slide images. The method first clusters patch embeddings into microstates and then groups them into macrostates based on spatial adjacency patterns. These macrostates are used to construct slide-level representations, which are combined with PCA-based low-rank molecular regression for gene expression prediction.

Critical strengths

Traditional MILs typically view a WSI as a patch collection, and then obtain a representation of the entire image through methods such as attention pooling. MoSPR is: processing WSI to 200 Microstates, and then to 8 Macrostates, then predict Gene Expression.

This hierarchical representation may be better preserves heterogeneity in tissue morphology.

Critical weaknesses

The proposed framework mainly combines existing techniques, including K-means clustering, spectral clustering, PCA, and ridge regression. The main contribution seems to be the construction of a shared morphological dictionary using spatial adjacency information. Although the overall combination is reasonable and produces good results, the novelty of the methodology is somewhat limited.

Another thing is that the contribution of spatial information is not sufficiently isolated. Although spatial adjacency is central to the proposed method, it is primarily used to construct a shared macrostate dictionary for the entire training queue. The final patient-level representation is based on aggregated features and does not explicitly present a patient-specific spatial arrangement. Therefore, two slides with similar macrostate distribution and average characteristics but different spatial organization can produce an objective representation.

Moreover, the existing shuffled-state ablation does not fully establish whether the improvements come from spatial adjacency modeling or simply from grouping patches into different morphological categories. A more direct comparison with feature-based clustering, without spatial adjacency information, would help clarify the contribution of the proposed spatial modeling strategy.

The proposed molecular procedure was obtained using PCA on gene expression profiles. However, PCA identifies directions of maximum variance that do not necessarily correspond to biologically meaningful molecular procedures. These components can capture a mixture of biosignals and technological changes.

Questions for authors
  1. How does the proposed spatial macrostate discovery compare with directly clustering patch embeddings into the same number of groups without using spatial adjacency?

  2. Can the authors provide additional evidence that the PCA-derived molecular programs correspond to biologically meaningful processes?

  3. How stable are the learned macrostates across different training splits and random initializations?

6 comments
AuthorsOpenReview ↗

Response to Reviewer XgSE (1/4): W1

We thank the reviewer for the careful and constructive review and for recognizing the potential of the hierarchical representation. We address the concerns and questions below.

W1. The proposed framework mainly combines existing techniques, and the methodological novelty is somewhat limited.

We thank the reviewer for this comment, and we agree that MoSPR is built from established components (k-means, spectral clustering, PCA, ridge regression). Our contribution lies in how these components are combined for a setting where paired histology-transcriptomic cohorts are small. Instead of learning a separate representation for each patient, MoSPR builds a single cohort-level dictionary of tissue states once, and describes every slide in this shared coordinate system by how much of each state it contains and how each state deviates from the slide average. The result is a compact, efficient linear framework that remains effective with limited data.

In addition, spatial context was intended to enter at the same cohort level. Rather than modeling each patient's own spatial graph, which would add patient-specific parameters that must be learned from limited data, MoSPR uses spatial adjacency once, pooled over the training cohort, to decide which microstates should be grouped into the same tissue state. Spatial information thus shapes the shared dictionary without enlarging the per-patient representation. However, our new analyses show that, with immediate (8-neighbour) contacts, this spatial signal overlaps substantially with morphological similarity; we discuss this limitation and broader neighbourhood definitions in our response to Q1.

This choice addresses three practical limitations of patient-specific approaches such as MIL, Transformer- and graph-based models:

  1. Data efficiency. These models learn how to aggregate each patient's patches from slide-level expression targets alone, which requires many parameters while paired cohorts number only hundreds to about a thousand patients. MoSPR fixes the aggregation structure without supervision, so the supervised part reduces to a single linear map. Accordingly, it reaches the full-data performance of ABMIL with about 44% of the training data on BRCA (Fig. 2).
  2. A shared reference. Patient-specific attention weights or graphs differ from slide to slide, so representations are not aligned across patients. In MoSPR, all patients are described on the same state dictionary, making their representations directly comparable. The linear form also lets each prediction be decomposed additively into global and state-specific components. We use this decomposition descriptively and do not claim that it identifies biological programs.
  3. Computational cost. MoSPR requires 0.96M parameters and 2.02 GFLOPs per slide, compared with 1.5–20.7M parameters and up to 516 GFLOPs for the baselines (Table S3).

We have revised the Introduction to state explicitly how MoSPR differs from existing approaches: rather than learning a patient-specific representation for each slide, it describes all patients over a shared, cohort-level dictionary of tissue states, so that supervision is spent only on a compact linear mapping to expression.

AuthorsOpenReview ↗

Response to Reviewer XgSE (2/4): W2

W2. The contribution of spatial information is not sufficiently isolated, and the final representation does not explicitly encode patient-specific spatial arrangement.

We thank the reviewer for this precise observation, and we agree with it. Once the dictionary is fixed, the slide representation (Eqs. 12–14) depends only on patch features and their state assignments, not on patch coordinates. Two slides with the same state composition and features but different spatial organization therefore receive identical representations.

This is a deliberate design choice. In MoSPR, spatial adjacency is not a patient-level feature; it is used once, at the cohort level, to define a shared dictionary of tissue states. Each slide is then described by how much of each state it contains and how each state deviates from the slide average. Encoding each patient's spatial layout would add patient-specific descriptors that must be learned from limited paired data, contrary to the compact, data-efficient design described in our response to W1.

We also expected this choice to suit the prediction target. Bulk expression is a spatial average. Its variation should be explained mainly by which states are present, in what proportion, and with what morphology, rather than by how those states are arranged within the slide. To test this directly, we augmented the representation with each patient's macrostate contact matrix, In=R⊤C(n)RI_n = R^\top C^{(n)}R. Here C(n)C^{(n)} counts adjacent microstate pairs in slide nn, and RR is the shared microstate-to-macrostate assignment. InI_n (normalized, upper triangle; 36 dimensions for K=8K=8) records how often each pair of macrostates touches in that slide, so it distinguishes slides with identical composition but different spatial organization. Using the same dictionary, the same two-stage ridge protocol, and the same rank and regularization search:

Gene SCCBRCAKIRCLUAD
MoSPR [M | S]0.41000.34740.3752
+ [M | S | I]0.41000.34750.3751

The change was within ±0.0004 in all 12 cohort–fold settings. For bulk expression, patient-specific arrangement, at least as captured by macrostate contact frequencies, adds no information beyond state composition and morphology, and the cohort-level design incurs no measurable loss.

Regarding the isolation of spatial information at the dictionary level, we re-learned the dictionary after randomly permuting patch coordinates within each training slide. This preserves state composition but removes the true spatial arrangement. Compared with the original dictionary on the same folds and seeds (4 folds × 5 seeds per cohort), coordinate permutation lowered gene-level SCC in KIRC (0.348 → 0.340; −0.008, p<10−4p<10^{-4}) and LUAD (0.373 → 0.359; −0.014, p<10−3p<10^{-3}), whereas the change in BRCA was not significant (0.412 → 0.411; −0.002, p=0.45p=0.45). Overall, SCC decreased by 0.008 on average and was lower in 49 of 60 paired comparisons. Cohort-level adjacency therefore contributes some information beyond co-abundance, but its effect is modest. A likely reason is that our adjacency is defined by immediate (8-neighbour) contacts, and because tissue morphology is spatially continuous, immediately adjacent patches are usually morphologically similar. The adjacency graph thus largely overlaps with appearance similarity, which is consistent with our finding that a dictionary built from appearance alone performs at least as well (see our response to Q1).

Because MoSPR is linear and block-structured, patient-specific spatial terms such as InI_n can be appended directly when the target is spatially resolved (e.g., spatial transcriptomics), where arrangement is expected to matter. We now discuss this as a limitation of the current bulk-oriented design and a natural extension.

AuthorsOpenReview ↗

Response to Reviewer XgSE (3/4): W3 and Q1

W3. A direct comparison with feature-based clustering without spatial adjacency would help clarify the contribution of spatial modeling.

Please see our response to Q1.

Q1. How does the proposed spatial macrostate discovery compare with directly clustering patch embeddings into the same number of groups without using spatial adjacency?

We thank the reviewer for suggesting this control. We built an appearance-only dictionary by clustering the microstate centroids in the CONCH embedding space into the same number of groups (K = 8, size-weighted k-means). This replaces adjacency-based spectral grouping and keeps everything else identical: the microstate vocabulary, the [M | S] representation, and the same rank and regularization search. Over four folds and five seeds, appearance-only grouping achieved slightly higher gene-level SCC than our adjacency-based grouping:

Gene SCCBRCAKIRCLUAD
MoSPR0.4120.3480.373
Appearance-only0.4210.3570.386

The difference is +0.011 on average, with appearance-only higher in 56 of 60 fold–seed pairs.

We consider this an informative result and a limitation of the current adjacency definition. MoSPR builds the microstate graph from immediate (8-neighbour) patch contacts. Because tissue morphology is spatially continuous, immediately adjacent patches are in most cases morphologically similar. The resulting adjacency graph therefore largely recapitulates appearance similarity, rather than capturing how distinct tissue compartments (e.g., tumor, stroma, immune infiltrates) border or intermix.

This has two implications.

Source of the gain. MoSPR's improvement comes mainly from how each slide is represented: as the overall morphology plus, for each tissue state in a shared dictionary, how abundant that state is and how it differs from the slide average. It does not come mainly from using adjacency to define those states. Supporting this, states defined by appearance alone work at least as well as adjacency-based states, whereas randomly grouped states perform no better than using overall morphology alone (Table 3).

Future directions. Capturing tissue architecture will require a broader notion of neighbourhood than immediate contact. Examples include dilated neighbourhoods that skip the nearest ring (analogous to dilated convolutions), multi-scale neighbourhoods combining several spatial radii, and definitions that emphasise contacts between morphologically distinct states. We consider these as directions for future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.