acceptodds
Under review as a conference paper at ICLR 2027

RESOLVE: Residual Decomposition with Single-Cell Prototype Guidance for Spatial Gene Expression Prediction

Abstract

Predicting spatial gene expression from H&E-stained histology offers a promising low-cost alternative for reconstructing spatial transcriptomic maps. However, existing methods typically model all expression variation jointly, conflating stable morphology-associated patterns with slide-specific cell-state deviations and thereby limiting the recovery of signals that are not directly explained by histology. To this end, we propose RESOLVE, which reformulates spatial gene expression prediction as two complementary processes, a morphology-driven base prediction and a cell-state-driven residual correction. A lightweight base predictor learns the morphology-expression mapping shared across slides, capturing stable tissue-level expression patterns. A prototype-guided residual pathway builds cell-state prototypes with explicit biological meaning from a single-cell reference and uses prototype routing to correct the slide-specific deviations that the base prediction cannot adequately explain. A per-spot gate modulates the residual magnitude according to the local tissue state, and an orthogonality constraint encourages the two pathways to learn complementary and decoupled representations, explicitly separating stable morphological signals from slide-specific cell-state changes. Experiments on three public ST datasets show that RESOLVE attains the best results on Her2ST and the best PCC on ST-Net and SCC. The gains of the residual pathway are interpretable. Genes that benefit most improve mainly by correcting slide-level differences and concentrate in interferon- and estrogen response programs. This pattern indicates that the residual corrects cell-state deviations rather than re-fitting what morphology already states.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, Gene expression prediction, Interpretable deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_scEVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RESOLVE for predicting spot-level gene expression from H&E images. The model combines a base predictor with a residual branch guided by cell-type expression profiles from a single-cell reference. Both branches use image features from a frozen UNI2-h encoder followed by local and global attention. The residual branch learns to correct the base prediction, with a gate controlling the size of the correction at each spot. The authors evaluate 250 genes on three cancer spatial transcriptomics datasets and report higher PCC than the compared methods. Ablation studies examine the model components, and further analyses show that the residual branch mainly improves expression differences between slides. The authors interpret these corrections as cell-state-related variation.

Critical strengths
  1. The base-plus-residual design is a good idea. Keeping a separate base prediction allows the single-cell reference to guide the correction without determining the entire output. Local and global attention also give the model access to both nearby tissue context and broader slide structure.
  2. The authors compare against several recent methods, including CPNN, DKAN, and HyperST. They also use the same single-cell reference for CPNN, which makes that comparison more informative.
Critical weaknesses
  1. It is unclear whether the residual captures cell-state variation or differences in cell-type composition. Each spot contains multiple cells, and the prototype control, pathway enrichment, and between-slide analyses do not distinguish these explanations. An analysis controlling for cell-type composition would help. As a possible follow-up, paired H&E and Xenium data could be used to examine residual gains within the same cell type. I would not require this additional experiment, since adapting the model to single-cell resolution would take further work, and the larger number of cells could make global attention expensive.
  2. The folds are split by slide, although the datasets contain multiple slides from each patient. Can slides from the same patient appear in both training and validation sets? Patient-held-out results would be important for assessing generalization, particularly given the emphasis on correcting between-slide differences.
  3. RESOLVE uses UNI2-h features, while MERGE uses ResNet18. The comparison therefore does not isolate the contribution of the proposed method from that of the image encoder. It would be useful to compare the models using the same UNI2-h features and comparable training losses.
Questions for authors
  1. Can the authors distinguish residual corrections due to cell-state changes from those due to differences in cell-type composition? Does the residual still help after controlling for composition?
  2. Are all slides from a patient kept in the same fold? If not, how does the model perform when patients, rather than slides, are held out?
  3. Were any baselines evaluated using the same UNI2-h features as RESOLVE? How much of the improvement remains when the image encoder and training losses are matched?
9 comments
AuthorsOpenReview ↗

Response to Reviewer scEV

We thank Reviewer scEV for recognizing the value of RESOLVE's base-plus-residual decomposition. We address all three questions below.

Reporting conventions. The original RESOLVE paper follows MERGE (CVPR 2025) exactly for data processing and splitting. Upon verification, we find that this splitting protocol (slide-level splitting) induces patient-level overlap, i.e., slides from the same patient may appear in both training and test sets. In addition, MERGE computes PCC per gene over spots pooled across all validation slides of a fold, which mixes inter-sample variation and may artificially affect correlation estimates, failing to faithfully reflect the model's ability to capture within-slide spatial expression patterns. Accordingly, all experiments in this response adopt a revised splitting protocol and PCC computation following DKAN (AAAI 2026). Splits are made at the patient level, and PCC is computed per gene across spots within each slide and then averaged over folds. For the prediction target, we retain MERGE's counts_spcs_to_8n preprocessing, as MERGE explicitly argues that gene-aware smoothing is necessary to mitigate technical artifacts in ST data and is more biologically justified than earlier spatial smoothing. This preprocessing acts only on the prediction target and does not involve sample splitting. It is therefore orthogonal to the evaluation-protocol corrections above, so that performance differences can be attributed to the splitting and PCC corrections rather than to changes in input distribution.

AuthorsOpenReview ↗

W1/Q1: Residual correction analysis

Response. W1 and Q1 address the same question: how much of the residual correction originates from cell-type composition versus state differences within the same cell type, and whether the residual remains effective after controlling for composition. The prototype control, Hallmark enrichment, and between-/within-slide decomposition originally cannot answer this question alone. We therefore add a composition-control analysis.

We use cell2location to deconvolve each spot over the same 29 cell types and 250-gene signature used for the prototypes, yielding composition vector p_k. The per-slide control applies within-slide centering. Within each training slide, expression and composition are each centered by subtracting the slide mean, and expression is then regressed per gene on the composition vector, removing only the within-slide component that covaries with composition. This avoids a cross-slide global regression mistaking between-slide mean differences for composition effects. That global regression yields negative R² on held-out slides (median −0.47) and is therefore not used. Before and after residualization, target mean and standard deviation remain at 2.508/1.002, so changes in ΔPCC are attributable to removal of the composition component rather than to a change in target scale. The between-slide control targets slide means. Each slide has only 29 composition dimensions and each training fold contains about 30 slides, so direct 29-dimensional regression would overfit. We therefore first compress composition to 3 principal components via PCA, retaining the dominant between-slide composition variation, and then regress slide-mean expression per gene and apply the result to held-out slides. Composition entropy H = −Σ p_k log p_k / log 29 is stratified into tertiles within each slide: high indicates relatively mixed composition across cell types, low indicates dominance by a single cell type, and mid lies between.

Table scEV-1a: Residual gains after composition control

Settingraw ΔPCCafter composition controlretained fraction
per-slide (primary metric)0.004460.0036281%
between-slide (PCA-3)0.03400.017451%

Table scEV-1b: Per-slide gains stratified by composition entropy

Composition-entropy stratum (per-slide)raw ΔPCCafter composition controlretained fraction
high (mixed composition)0.00800.006986%
mid0.00600.005490%
low (single-type dominated)0.00100.000330%

After composition control on the within-slide component, ΔPCC drops from 0.00446 to 0.00362, retaining 81%. Composition thus explains only about 19% of this gain in ΔPCC terms. The stratified results are directionally consistent. The high and mid entropy groups retain 86% and 90%, respectively, whereas the gain in the low entropy group drops from 0.0010 to 0.0003 and essentially vanishes. This indicates that the state correction occurs mainly on spots with mixed composition, whereas on spots dominated by a single cell type the residual gain is largely explained by composition. The raw ΔPCC in the low entropy group is itself small, so the ratio is more sensitive to noise, but its direction is consistent with the high and mid groups. The between-slide control retains 51% (from 0.0340 to 0.0174). Composition accounts for a substantial share of slide-level differences, but a non-zero state component remains. After composition removal, the between-slide residual (0.0174) remains clearly larger than the within-slide residual (0.0036). The original conclusion that the residual is positioned as a between-slide signal therefore holds, although the description should be revised to a mixed correction of state and composition.

Table scEV-2: Hallmark enrichment of between-slide surviving genes after composition adjustment (top-15)

TermAdjusted pOdds ratio
Interferon Gamma Response3.2×10⁻⁶50.7
Allograft Rejection6.7×10⁻⁵36.7
Interferon Alpha Response1.8×10⁻⁴52.9
Apoptosis0.01719.2
Inflammatory Response0.02115.4

Among genes that retain a between-slide gain after composition removal, IFN-γ, allograft rejection, IFN-α, apoptosis, and inflammatory response rank in the top 5. This enrichment pattern is consistent with a between-slide cell-state correction interpretation and is observed after the composition effect has been removed.

Overall, the residual contains both a cell-state and a cell-composition component. The within-slide main effect is dominated by state, with composition accounting for about 19%. In between-slide differences, composition accounts for a relatively high share (about half), but a non-zero state component remains, and this component is still enriched for IFN programs. The revision will change "pure cell-state correction" in the main text to a mixed correction of state and composition.

AuthorsOpenReview ↗

W2/Q2: Data-splitting protocol

Response. As stated under Reporting conventions, the original data-splitting scheme allowed slides from the same patient to appear in both training and validation sets. Following DKAN (AAAI 2026), we re-adopt a patient-based splitting scheme to evaluate the model's true generalization ability. Table scEV-3 reports the comparison of representative algorithms under the updated evaluation protocol. RESOLVE is best on all 9 metrics (PCC/MAE/MSE across three datasets). For example, on Her2ST, RESOLVE achieves a PCC of 0.3245, higher than DKAN's 0.3138 and MERGE's 0.1938.

Table scEV-3: Comparison of representative algorithms under patient-level splitting and the DKAN evaluation protocol

MethodST-Net PCCST-Net MAEST-Net MSEHer2ST PCCHer2ST MAEHer2ST MSESCC PCCSCC MAESCC MSE
TRIPLEX (CVPR 2024)0.16730.46730.36180.30160.69420.80020.40330.68850.7428
MERGE (CVPR 2025)0.07800.46830.36440.19380.75530.89520.29400.71990.7600
DKAN (AAAI 2026)0.17020.45350.33910.31380.72690.85420.43630.67860.7007
HyperST (CVPR 2026)0.14790.44990.33920.27200.70340.81850.39310.69850.7604
RESOLVE0.18130.42960.30270.32450.69060.76940.45170.66010.6819
Reviewer_u7GoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper predicts spot-level gene expression from H&E images. Specifically, the framework (i) combines local and global attention mechanisms that are used sequentially to extract a joint representation. Then, (ii) the joint representation is split into two linear projections, with a loss used to encourage their orthogonality. (iii) A small MLP is used to predict the gene expression profile from one of these projections. (iv) A residual MLP corrects this using the other projection together with a "biological condition vector". That vector is from the top-K routing over learnable prototypes, initialised from a scRNA-seq atlas.

Experiments use the MERGE protocol with three datasets, the 250 most highly expressed genes, SPCS smoothing and an 8-fold cross-validation split by slide. The proposed method achieves competitive performance compared with a wide variety of approaches. No downstream task is reported.

Critical strengths

None

Critical weaknesses
  1. The experimental protocol leaks patient information (slice-level split). Since the method leverages a learnable residual target, the main claim that the residual corrects slide-specific cell-state deviation could be in fact due to patient leakage. This drives my reject recommendation.
  • Folds are split by slide instead of at the patient level. This means that slices from the same patients can be found in the training and testing sets. For example, 4 subjects in cSCC with an 8-fold cross-validation.

  • In App. D.4, the top-15 genes gain +0.0321 PCC between slides but only +0.0021 within slides. Within-slide spatial prediction thus barely improves. Under this protocol, a "slide-level cell-state correction" also can't be distinguished from recognising a patient and reproducing that patient's mean expression, learned from the patient's other sections in the training set.

  • The protocol inflates the performance of baselines too. Two baselines were originally evaluated on these datasets with patient-disjoint folds and per-section PCC, and both score higher here than in their own papers (Her2ST / ST-Net / SCC). TRIPLEX reaches 0.393 / 0.232 / 0.542 here against 0.314 / 0.206 / 0.374 originally (CVPR 2024, Table 1). DKAN reaches 0.413 / 0.241 / 0.558 against 0.330 / 0.219 / 0.407 (AAAI 2026, Table 1).

  1. Seven of the ten baseline performances in Table 1 (ResNet+FCN, BLEEP, HisToGene, Hist2ST, THItoGene, TRIPLEX, MERGE) are identical to MERGE's Table 1. It is also unclear whether RESOLVE uses MERGE's exact fold assignment.
Questions for authors
  1. How is PCC in Table 1 computed? Is it per gene over spots pooled across all validation slides of a fold, as "the pooled one used elsewhere" suggests, or per section and then averaged?

  2. How were the 8 folds formed, and are they MERGE's exact fold assignment?

  3. How were the dataset-specific hyperparameters chosen?

  4. Which normalisation of the ST data was used?

9 comments
AuthorsOpenReview ↗

Response to Reviewer u7Go

We thank Reviewer u7Go for the review. The patient leakage issue raised by the reviewer does hold under the original slide-level protocol. We accept this criticism and have re-examined all experiments accordingly. We address W1 and W2 separately below and clarify overlaps with Q1–Q4 where relevant.

Reporting conventions. The original RESOLVE paper follows MERGE (CVPR 2025) exactly for data processing and splitting. Upon verification, we find that this splitting protocol (slide-level splitting) causes patient-level overlap, i.e., slides from the same patient may appear in both training and test sets. In addition, MERGE computes PCC per gene over spots pooled across all validation slides of a fold, which mixes inter-sample variation and may artificially affect correlation estimates, and cannot faithfully reflect the model's ability to capture within-slide spatial expression patterns. Therefore, all experiments in this response adopt a revised splitting protocol and PCC computation following DKAN (AAAI 2026): splitting by patient, with PCC computed per gene across spots within each slide and then averaged across folds. For the prediction target, we retain MERGE's counts_spcs_to_8n preprocessing, because MERGE explicitly argues that gene-aware smoothing is a necessary step to mitigate technical artifacts in ST data and is more biologically justified than earlier spatial smoothing. This preprocessing acts only on the prediction target and does not involve sample splitting, and is therefore orthogonal to the evaluation-protocol corrections above. Keeping it unchanged allows performance differences to be attributed cleanly to the splitting and PCC protocol corrections rather than to changes in input distribution.

AuthorsOpenReview ↗

W1 + Q1 + Q2: Patient leakage, slide-level splitting, PCC protocol, and fold assignment

Response. The reviewer's assessment is accurate. The original fold assignment is identical to MERGE's, and this slide-level splitting allows slides from the same patient to appear in both training and test sets. The residual branch is a learnable, independent branch. Under the original protocol, it could indeed reproduce a patient's mean expression via patient identity without learning cell state. Thus, the original protocol cannot distinguish "residual correction of slide-specific cell-state deviations" from "recognizing the patient and reproducing their mean expression." This is a genuine flaw of the original protocol and the direct reason we revised the splitting.

Under patient-level splitting, all slides from a patient are constrained to the same fold, so training and test sets are disjoint at the patient level and the model cannot reproduce a patient's mean expression via patient identity. Therefore, residual gains under the revised protocol cannot be a product of patient leakage, and the alternative explanation raised in W1 no longer holds under the revised protocol.

Regarding the Appendix D.4 evidence noted by the reviewer, we recomputed within-slide and between-slide residual gains (Table u7Go-1). Under the original protocol, the top-15 genes show a between-slide gain of +0.0321 and a within-slide gain of only +0.0021, a ratio of about 15×, which is exactly the source of the reviewer's concern. This decomposition is computed on the same set of predictions; the dominance of the between-slide gain is not an artifact of the PCC protocol and holds under both the original and revised protocols. It is consistent with the design goal of the residual branch, namely correcting slide-level state deviations, but consistency alone does not constitute evidence of validity. Supporting evidence comes from two other points. First, patient-level splitting rules out patient identity as an alternative explanation. Second, after removing cell composition, the between-slide gain retains about half (from 0.0340 to 0.0174), and the surviving genes remain enriched for IFN programs.

Note that absolute values across rows in Table u7Go-1 are not directly comparable. The first three rows are supplementary analyses under the patient-level protocol, whereas the Manuscript D.4 row reports the same-configuration residual ablation under the original protocol; the evaluation protocol and gene subset differ. The top-15 rows are selected by the corresponding gain ranking and are subject to selection effects; they are descriptive only. A more reliable estimate of the within-slide gain comes from the all-250-genes row, +0.0045. The residual branch therefore also contributes within slides, although its main effect is concentrated between slides.

On the PCC protocol in Q1: the original Table 1 uses MERGE's pooled protocol, i.e., spots from all validation slides of a fold are pooled and PCC is computed per gene. This protocol mixes between-slide variation into the correlation estimate and makes between-slide gains inseparable from within-slide spatial prediction gains. The revised experiments follow the DKAN (AAAI 2026) protocol: PCC is computed per gene across spots within each slide and then averaged across folds.

On the fold assignment in Q2: the revised fold assignment differs from MERGE. We keep the 8-fold framework but impose a patient-level constraint: all slides from a patient must fall in the same fold, so fold boundaries differ from MERGE's original assignment. This difference is necessary to eliminate leakage. The SCC dataset contains only 4 patients; under the patient-level constraint this becomes 4 folds, i.e., leave-one-patient-out cross-validation, consistent with DKAN's practice on SCC. Her2ST and ST-Net still yield 8 folds after grouping by patient.

Table u7Go-1: Within-slide and between-slide residual gains

Protocolwithin-slide ΔPCCbetween-slide ΔPCCbetween / within
All 250 genes+0.0045+0.0340~7.6×
top-15 (ranked by between)+0.0164+0.2818~17×
top-15 (ranked by within)+0.0159+0.3232~20×
Manuscript D.4 (top-15, same-configuration residual ablation, MERGE pooled protocol)+0.0021+0.0321~15×
AuthorsOpenReview ↗

W2: Baseline performance inflation and the source of numbers in Table 1

Response. The reviewer notes that seven of the ten baselines in Table 1 (ResNet+FCN, BLEEP, HisToGene, Hist2ST, THItoGene, TRIPLEX, MERGE) are identical to those in MERGE's Table 1. We confirm this. The original Table 1 uses exactly the same fold assignment and PCC computation as MERGE, and these seven rows are directly cited from MERGE's published numbers to align with the original paper under the same protocol. The original wording did not distinguish "cited" from "retrained by us," which may lead readers to assume that all ten baselines were retrained by us in the same environment. The original description of baseline sources was insufficiently clear. We accept this criticism and will annotate the source of each entry in the revision.

The reviewer further notes that some baselines score higher under the slide-level protocol than under the patient-level results in their original papers (e.g., TRIPLEX, DKAN). This reflects protocol inflation rather than a source of RESOLVE's advantage. We agree. The slide-level protocol systematically inflates metrics, and these numbers cannot support RESOLVE's conclusions under the patient-level protocol.

To this end, we re-tested RESOLVE and representative algorithms following the DKAN (AAAI 2026) protocol; results are in Table u7Go-2. In Table u7Go-2, RESOLVE is best on all 9 metrics (PCC/MAE/MSE across the three datasets). It is not unexpected that the re-tested TRIPLEX and DKAN results differ from their original papers. We retain MERGE's data-processing pipeline, whereas each method's original paper uses its own data-processing and tuning pipeline, and the reproduction environment also differs from the original. Therefore, Table u7Go-2 is a controlled comparison under the same preprocessing and the same splitting. Its numbers should not be compared entry-by-entry with those in the original papers, nor should it be interpreted as each method's best performance.

Table u7Go-2: Re-tested RESOLVE and representative algorithms under patient-level splitting and the DKAN evaluation protocol

MethodST-Net PCCST-Net MAEST-Net MSEHer2ST PCCHer2ST MAEHer2ST MSESCC PCCSCC MAESCC MSE
TRIPLEX (CVPR 2024)0.16730.46730.36180.30160.69420.80020.40330.68850.7428
MERGE (CVPR 2025)0.07800.46830.36440.19380.75530.89520.29400.71990.7600
DKAN (AAAI 2026)0.17020.45350.33910.31380.72690.85420.43630.67860.7007
HyperST (CVPR 2026)0.14790.44990.33920.27200.70340.81850.39310.69850.7604
RESOLVE0.18130.42960.30270.32450.69060.76940.45170.66010.6819
Reviewer_d1CROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Authors propose RESOLVE, which reformulates spatial gene expression prediction as two complementary processes, a morphology-driven base prediction and a cell-state-driven residual correction. A lightweight base predictor learns the morphology-expression mapping shared across slides, capturing stable tissue-level expression patterns. A prototype-guided residual pathway builds cellstate prototypes with explicit biological meaning from a single-cell reference and uses prototype routing to correct the slide-specific deviations that the base prediction cannot adequately explain. A per-spot gate modulates the residual magnitude according to the local tissue state, and an orthogonality constraint encourages the two pathways to learn complementary and decoupled representations, explicitly separating stable morphological signals from slide-specific cell-state changes. Experiments on three public spatial transcriptomics datasets show that RESOLVE attains the best results on Her2ST and the best PCC on ST-Net and SCC. The gains of the residual pathway are interpretable.

Critical strengths
  1. The paper distinguishes morphology-associated gene expression from slide-specific cell-state variation and motivates modeling these components separately in histology-based spatial gene expression prediction.

  2. RESOLVE combines a dual-bias attention (DBA) encoder for local and global tissue context with a morphology-driven base predictor and a single-cell-prototype-guided residual pathway. Sparse prototype routing, per-spot gating, orthogonality regularization, and gradient isolation support the two-pathway formulation.

  3. The method achieves competitive predictive performance across three public spatial transcriptomics datasets against a broad set of baselines.

  4. The paper examines residual-gain genes through pathway enrichment, visualizes base and residual predictions spatially, and analyzes between-slide and within-slide contributions. These analyses help characterize where the residual correction contributes.

Critical weaknesses
  1. Limited controls for backbone and encoder choices. RESOLVE relies on frozen UNI2-h features and the proposed DBA encoder (Sections 2.2–2.3). The reported ablations remove L-DBA and G-DBA separately (Table 3), but do not compare different pathology backbones or a parameter-matched alternative to DBA. Since the residual branch models errors left by the base predictor, the shared representation may influence which errors remain to be corrected. Experiments using an alternative pathology backbone and a parameter-matched encoder in place of DBA would help establish the robustness of the proposed formulation and the contribution of the encoder design.

  2. Limited statistical evidence for the residual improvement. On Her2ST, the full model improves PCC from 0.5583 without the residual branch to 0.5649 (Table 3), an absolute gain of 0.0066. The paper should report the distribution of residual-branch gains across folds or repeated random seeds, together with confidence intervals, to establish the stability of this improvement.

  3. Insufficiently specified evaluation protocol. The paper reports slide-level eight-fold cross-validation on datasets containing multiple slides per patient (Section 3; Appendix C.1), but does not clarify whether slides from the same patient can appear in both training and validation folds. Appendix C.2 also describes affine alignment of single-cell signatures to the target ST expression scale without specifying which ST data are used for that alignment. The authors should clarify these procedures in the Method or Experimental Setup, including whether the alignment is fitted using training-fold data only.

Questions for authors

Q1. Cross-dataset residual validation**

  • Could the authors extend the residual-branch comparison to ST-Net and SCC, ideally using paired analyses at the slide or patient level?

Q2. Region-stratified biological interpretation**

  • Could the authors evaluate base and residual predictions across histologically distinct regions using established morphology-associated genes and available non-malignant tissue annotations?

Q3. Single-cell reference sensitivity and random-prior control**

  • How were the single-cell references selected with respect to molecular subtype compatibility and batch differences, and how sensitive are the results to reference choice?
  • In the random-prior control (Appendix B.5), was the anchoring target in Eq. (13) randomized along with the prototype initialization?

Q4. Cross-tissue generalization**

  • Additional validation using the public colorectal cancer spatial transcriptomics dataset from Profiling the heterogeneity of colorectal cancer consensus molecular subtypes using spatial transcriptomics (2024) would help assess generalization to another tissue context. Could the authors consider this experiment or discuss practical limitations?
7 comments
AuthorsOpenReview ↗

Response to Reviewer d1CR

We thank Reviewer d1CR for acknowledging the value and predictive performance of RESOLVE in separately modeling morphology-related gene expression and slice-specific cell-state changes. We address the three critical weaknesses (W1–W3) and all four questions (Q1–Q4) below.

Reporting conventions. The original RESOLVE paper follows MERGE (CVPR 2025) exactly for data processing and data splitting. Upon verification, we find that this splitting protocol (slide-level splitting) causes patient-level overlap, i.e., slides from the same patient may appear in both training and test sets. In addition, MERGE computes PCC per gene over spots pooled across all validation slides of a fold, which mixes inter-sample variation and may artificially affect correlation estimates, and cannot faithfully reflect the model's ability to capture within-slide spatial expression patterns. Therefore, for all experiments in this response, we revise both the data-splitting protocol and the PCC computation following DKAN (AAAI 2026): splitting by patient, with PCC computed per gene across spots within each slide and then averaged across folds. For the prediction target, we retain MERGE's counts_spcs_to_8n preprocessing, because MERGE explicitly argues that gene-aware smoothing is a necessary step to mitigate technical artifacts in ST data and is more biologically justified than earlier spatial smoothing. This preprocessing acts only on the prediction target and does not involve sample splitting, and is therefore orthogonal to the evaluation-protocol corrections above. Keeping it unchanged allows performance differences to be attributed cleanly to the splitting and PCC protocol corrections rather than to changes in input distribution.

AuthorsOpenReview ↗

W1:Backbone and encoder choices

Response. We add three sets of experiments. Table d1CR-1 compares replacing the feature-extraction backbone in RESOLVE. CONCH is a CoCa vision-language model (ViT-B/16, 512-dim), CigaR is a SimCLR contrastive self-supervised method (ResNet-18, 512-dim), and ResNet50 is an ImageNet-pretrained general-purpose visual baseline (2048-dim). RESOLVE uses UNI2-h by default (ViT-H/14, 1536-dim). Table d1CR-2 reports the comparison of MERGE and its UNI2-h variant, testing the effect of backbone substitution on the baseline model. Table d1CR-3 replaces DBA with a parameter-matched encoder. All experiments use the revised patient-level splitting protocol and the DKAN PCC protocol, i.e., PCC computed per gene across spots within each slide and then averaged across folds.

Table d1CR-1: Replacing RESOLVE's feature-extraction backbone on Her2ST

BackbonePCCMAEMSEfeature dimension
UNI2-h0.32450.69060.76941536
CONCH0.31160.71570.8069512
ResNet500.21940.75960.85472048
CigaR0.29910.73940.8422512

Table d1CR-2: MERGE and its UNI2-h variant on Her2ST

MethodPCCMAEMSEfeature dimension
MERGE(ResNet18)0.19380.75530.8952256
MERGE+UNI2-h0.26890.71510.80871536

Table d1CR-3: Replacing DBA with a parameter-matched encoder

MethodPCCMAEMSE
RESOLVE0.32450.69060.7694
RESOLVE+DBA replacement0.27740.72020.8212

In Table d1CR-1, UNI2-h achieves the best PCC. Replacing it with CONCH, CigaR, or ResNet50 reduces PCC by 0.0129, 0.0254, and 0.1051, respectively. In Table d1CR-2, MERGE improves by 0.0751 when using UNI2-h relative to its ResNet18 version, but remains below RESOLVE+UNI2-h (0.2689 vs. 0.3245). In Table d1CR-3, replacing DBA with a parameter-matched encoder reduces PCC by 0.0471. These results respectively test RESOLVE's robustness across backbones and the contribution of the DBA encoder under parameter matching.

AuthorsOpenReview ↗

W2/Q1: Cross-dataset residual validation

Table d1CR-4 reports the residual-branch ablation on Her2ST, ST-Net, and SCC under the revised splitting. On the three datasets, RESOLVE improves PCC over RESOLVE_nores by 0.0241, 0.0069, and 0.0060, respectively, with consistent direction. This indicates that the residual-branch gain is not limited to a single dataset.

Table d1CR-4: Residual-branch ablation on three datasets

MethodST-Net PCCST-Net MAEST-Net MSEHer2ST PCCHer2ST MAEHer2ST MSESCC PCCSCC MAESCC MSE
RESOLVE0.18130.42960.30270.32450.69060.76940.45170.66010.6819
RESOLVE_nores0.15720.44190.32060.31760.69100.77540.44570.67550.7179

Comments on the paper

3 comments
AuthorsOpenReview ↗

RESOLVE General Response (Part 1 of 2)

Dear Reviewers,

Thank you for your careful reading and constructive feedback. RESOLVE decomposes spatial gene expression prediction into a morphology-anchored base prediction and a residual correction driven by slide-specific cell states, and models the two components explicitly rather than fitting them jointly. We study this decomposition on Her2ST, ST-Net, and SCC, asking which expression components are stably predictable from morphology and which reflect slide-specific state. We thank the reviewers for recognizing the decomposition, the base-plus-residual design, and the interpretability of the residual.

Recognized strengthRole in RESOLVEReviewer comment
Separating morphology-associated expression from slide-specific cell statesBase learns the cross-slide mapping, residual corrects slide-specific states.“distinguishes morphology-associated gene expression from slide-specific cell-state variation” (d1CR)
Two-pathway design with single-cell prototype guidanceDBA encoder, prototype routing, and per-spot gating.“The base-plus-residual design is a good idea” (scEV)
Competitive performance against diverse baselinesBroad baseline comparison on three public ST datasets.“competitive predictive performance across three public spatial transcriptomics datasets” (d1CR)
Interpretable analysis of residual gainsPathway enrichment, spatial visualization, between/within-slide decomposition.“These analyses help characterize where the residual correction contributes” (d1CR)

Major Questions Addressed in the Rebuttal

Our point-by-point replies appear in the corresponding sections. The reviews overlap, so we consolidate the main concerns into four core questions. Part 2 covers the additional experiments and key takeaways.

The original paper follows MERGE's slide-level splitting and pooled PCC protocol. This protocol induces patient-level overlap and mixes between-slide variation into the correlation estimate. All new experiments use patient-level splitting and the DKAN (AAAI 2026) protocol: all slides from a patient fall in the same fold, and PCC is computed per gene across spots within each slide and then averaged across folds. We retain MERGE's SPCS smoothing for the prediction target, which does not involve sample splitting and is orthogonal to these corrections.

Core questionReviewer feedbackOur response and evidenceDetailed replies
Evaluation protocol: patient-level splitting, PCC, and foldsu7Go W1, scEV W2, and d1CR W3 raise patient overlap. u7Go Q1 asks how PCC is computed.All new experiments use patient-disjoint splits. SCC has four patients, so we use leave-one-patient-out (4 folds). Her2ST and ST-Net remain at 8 folds, with fold boundaries changed to eliminate leakage. PCC is computed per gene across spots within each slide and then averaged across folds. RESOLVE is best on all nine metrics, with Her2ST PCC 0.3245 vs DKAN 0.3138 and MERGE 0.1938. Patient identity cannot explain the residual gains.Reporting conventions, u7Go W1/Q1/Q2, scEV W2/Q2, d1CR W3
Baseline numbers and protocol inflationu7Go W2: seven baselines match MERGE's Table 1, and some exceed their original patient-level results.These seven rows are directly cited from MERGE. The original wording did not distinguish cited from retrained numbers. We re-tested RESOLVE and representative algorithms under patient-level splitting and DKAN PCC, and RESOLVE is best on all nine metrics. The revision will annotate every source.u7Go W2
Backbone and DBA encoderd1CR W1 asks for backbone and DBA replacements. scEV W3/Q3 asks for same UNI2-h features and comparable training losses.We replace the backbone with CONCH, CigaR, and ResNet50, and DBA with a parameter-matched encoder. Under the same UNI2-h backbone, RESOLVE reaches 0.3245 vs MERGE's 0.2689. MERGE gains 0.0751 when switched to UNI2-h, CONCH is only 0.0129 lower, and replacing DBA reduces PCC by 0.0471.d1CR W1, scEV W3/Q3
Source of residual gainsu7Go W1 asks about between vs within. scEV W1/Q1 asks about composition. d1CR Q1/Q3 asks for cross-dataset validation and reference sensitivity.Under patient-level splitting, patient identity cannot explain the between-slide gain. The all-250-gene within/between split is +0.0045/+0.0340. After composition control, the within-slide gain retains about 81% and the between-slide gain about 51%, so composition explains about 19% of the within-slide gain. The between-slide residual (0.0174) remains larger than the within-slide residual (0.0036), so the slide-level positioning holds, now described as a mixed correction. Surviving genes remain enriched for IFN programs. Cross-dataset ablation and reference sensitivity appear in Part 2.u7Go W1, scEV W1/Q1, d1CR Q1, Q3
AuthorsOpenReview ↗

RESOLVE General Response (Part 2 of 2)

Additional Experiments and Clarifications

Beyond the four core questions, the reviewers offered more specific suggestions. We added cross-dataset residual ablation, region-stratified analysis, single-cell reference sensitivity, cross-tissue generalization, and clarification of hyperparameters and normalization.

Additional questionReviewer feedbackOur response and evidenceDetailed replies
Cross-dataset residual ablationd1CR asks for residual validation on multiple datasets.The residual gain over the no-residual model is consistent: +0.0241 on ST-Net, +0.0069 on Her2ST, and +0.0060 on SCC.d1CR Q1
Region-stratified interpretationd1CR asks how the residual affects tumor and non-malignant regions.With eight annotated slides and 3,481 spots, the residual improves tumor and non-malignant regions (+0.0117/+0.0068), strongest in adipose (+0.0138) and breast glands (+0.0145). Paired-test p-values are 0.101 and 0.345. Region labels come from pathologists and do not use the 250-gene panel.d1CR Q2
Single-cell reference sensitivityd1CR asks about sensitivity to reference choice.A random prior reduces PCC from 0.3245 to 0.3070 and eliminates Hallmark enrichment (p = 1.0). Log-normal noise (σ=0.3) gives 0.3057, and removing 25% of cell types gives 0.3141. A multi-atlas comparison is future work.d1CR Q3
Cross-tissue generalizationd1CR suggests validation on COAD and READ.RESOLVE reaches PCC 0.2457 on COAD and 0.4079 on READ, with the lowest MAE and MSE. Data are HEST-1K slices from Valdeolivas et al. DKAN was not reproduced.d1CR Q4
Hyperparameters and alignmentu7Go asks about hyperparameter selection. d1CR asks which data fit the alignment.Hyperparameters use nested validation within training folds. Affine alignment is fitted only on training folds. Details will appear in the appendix.u7Go Q3, d1CR W3
ST data normalizationu7Go asks which normalization is used.We retain counts_spcs_to_8n: SPCS-smoothed counts, min-max scaled to 8-neighbor counts, used as the prediction target. It does not involve sample splitting.u7Go Q4

Key Takeaways from the Rebuttal

  • Protocol correction. Under patient-level splitting and the DKAN PCC protocol, RESOLVE is best on all nine metrics. Seven baseline rows were cited from MERGE, and the revision will annotate every source.
  • Residual correction. Patient identity cannot explain the gain, the residual is consistent across datasets, and the state component survives composition control with IFN-enriched survivors.
  • Backbone and generalization. Under the same UNI2-h, RESOLVE outperforms MERGE, DBA contributes beyond parameter count, and gains extend to COAD/READ and non-malignant regions.

Together, these results return to the motivation for RESOLVE: stable morphology-driven signals and slide-specific cell-state signals in spatial gene expression should be modeled separately, and this decomposition can only be tested reliably under strict patient-level evaluation. The revised protocol does not weaken the conclusions. It makes the residual contribution clearer and more attributable. We believe that reliably separating stable morphology-driven signals from slide-specific state signals under limited evidence is an increasingly important problem in spatial gene expression prediction, and we would welcome the opportunity to bring RESOLVE to ICLR 2027 and discuss it with the community.

We would be glad to answer any further questions.

Best regards,

Authors

AuthorsOpenReview ↗

Rebuttal Update Notice

We have updated the relevant content of this round’s rebuttal in the appendix for your reference. The main text will be updated accordingly as the actual revisions are finalized.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.