Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Predicting spatially resolved gene expression from histology images requires capturing how tissue morphology relates to each gene, but the strength of this coupling differs across genes. Current methods assign every gene the same morphological representation, overlooking this heterogeneity. We introduce gene binding modules that decompose morphological features into functional sub-spaces, letting each gene attend selectively to relevant programs rather than a shared global representation. To capture inter-gene dependencies, compact regulator tokens compress and redistribute information across gene groups, modeling coordinated transcriptional programs. The framework further learns per-gene heteroscedastic uncertainty that focuses model capacity on morphologically predictable genes during training, and modulates stochastic sampling at inference: tightly coupled genes receive near-deterministic generation while loosely coupled genes are sampled more broadly. Experiments on four spatial transcriptomics datasets show consistent improvements over both deterministic and generative baselines. The learned modules recover biologically coherent gene groupings without supervision, validated through pathway enrichment analysis, demonstrating that the framework provides interpretable biological insights beyond prediction accuracy.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1384_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

HESK-1K: https://huggingface.co/datasets/MahmoodLab/hest

BibTex

@InProceedings{NamHyu_GeneSelective_MICCAI2026,
        author = { Namgung, Hyun AND Kim, Soopil AND Park, Sang Hyun},
        title = { { Gene-Selective Morphological Conditioning with Heteroscedastic Flow Matching for Spatial Transcriptomics } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a gene-selective generative framework for predicting spatial gene expression from histology images. The key idea is to model heterogeneous coupling between morphology and gene expression, where some genes are assumed to be tightly coupled (predictable from morphology) and others loosely coupled (less predictable). The method incorporates gene-wise attention, regulator tokens to capture gene–gene dependencies, and heteroscedastic uncertainty to adapt prediction confidence per gene. The main contribution is moving beyond direct black-box regression and attempt to model gene-specific relationships to morphology, and inter-gene dependencies via gene-selective conditioning and uncertainty-aware flow matching. Second, the idea of strongly and weakly morphology-coupled genes in modelling.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Moves beyond standard regression which is a meaningful step forward.
    • Structured modelling of gene–gene interactions using regulator tokens
    • Gene-specific morphological subgrouping instead of a single shared feature, allowing each gene to attend to different morphological programs, which aligns better with biological intuition.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • Limited generalizability / risk of dataset bias: The notion of “tight” vs “loose” coupling is learned implicitly from the dataset and may reflect dataset-specific correlations rather than true biology. For example, ERBB2 appears tightly coupled in Fig. 1, but this could be driven by the dataset being enriched for HER2+ cases. It is unclear how well this would transfer to a different cohort.
    • The paper does not perform cross cohort validation, train on one and test on another. Which is critical given that we know these models are capturing dataset-specific correlation patterns and the notion of tight and loose coupling is likely to make the domain shift even worse [see Dawood et al 2026 nature biomedical engineering]
    • From Table 2, performance appears sensitive to the number of regulator tokens. This raise concerns that the model may be learning dataset-specific configurations, which could limit generalization.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This is an interesting and well-motivated step toward more structured modelling of spatial transcriptomics. However, the current formulation appears to rely heavily on dataset-specific correlations, and lacks evidence of robustness under distribution shift. Addressing generalization and improving biological interpretability would significantly strengthen the work.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper proposes a mean flow matching based framework for spatial transcriptomics prediction and achieves the best overall performance among the compared methods on multiple datasets. On top of the conditional generative formulation, it introduces Regulator Attention to model gene dependencies, Gene Binding Module Attention to capture the associations between tissue morphology and different genes, and Uncertainty-Aware Flow Matching to account for the intrinsic variability in morphology to gene expression mapping. These designs are reasonable and work together well.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The overall idea is meaningful and reasonably novel. The authors cast spatial gene expression prediction as a conditional generative problem with flow matching, which is a natural way to model the one-to-many relationship between tissue morphology and gene expression. 2.The proposed modules are well motivated. 3.The empirical results are solid. The method achieves the best overall correlation performance on multiple datasets, especially on harder settings with more genes, which supports the value of the proposed design.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The paper also has some weaknesses. First, the assumption that the image feature can be decomposed into 6 morphological programs is not fully convincing. This design is reasonable, but the paper does not give strong enough evidence that such a fixed decomposition is biologically meaningful or consistently valid across datasets. The ablation only shows that K=6 works better empirically, but it does not fully justify the modeling assumption itself. Second, the method does not achieve the best MSE or MAE on the four datasets, and this point needs a clearer explanation. The paper argues that deterministic methods are naturally better on pointwise error because they predict the conditional mean, while the proposed method focuses more on modeling uncertainty and diversity. This is reasonable, but it still leaves some concern about absolute prediction accuracy.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I gave this paper a weak accept because the overall idea is meaningful, the proposed modules are well motivated, and the method shows solid experimental performance across multiple datasets. In particular, the paper addresses an important problem and achieves strong correlation results with a reasonable generative framework. At the same time, some parts of the method are not fully justified, and the gains are less clear on MSE and MAE, so I do not see it as a strong accept. Overall, I think the paper has enough novelty and practical value to merit acceptance, but it still has some limitations.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper proposes a flow-matching framework for histology-to-spatial-transcriptomics prediction that addresses three biological observations: (1) genes differ in how tightly their expression couples to tissue morphology, (2) genes are organized into coordinated transcriptional programs, and (3) a well-calibrated generative model should modulate per-gene stochasticity.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    Clear, biologically-grounded motivation. Fig. 1 is excellent: panel (a) shows that among morphologically similar spots, gene expression heterogeneity varies widely (ERBB2 tightly coupled, COL1A1 broadly variable — quantified via coefficient of variation over 10 regions); panel (b) shows the block-diagonal co-expression structure; panel (c) illustrates the adaptive-stochasticity intuition. This is the kind of motivation figure that turns three components into a coherent story. Each component is tied to a specific biological observation. Gene binding modules ↔ heterogeneous morphology–expression coupling; regulator tokens ↔ coordinated transcriptional programs; heteroscedastic flow matching ↔ gene-specific stochasticity. This is a well-constructed narrative. Strong empirical results. Wins or ties best on ρ10/ρ50/ρ200\rho_{10}/\rho_{50}/\rho_{200} ρ10​/ρ50​/ρ200​ across all four datasets; the pattern of growing margins with broader gene panels (42% relative gain on ρ200 BC-I vs. smaller gains on ρ10) is exactly what the gene-selective story predicts.

    Comprehensive baseline comparison. 10 methods spanning deterministic regression (ST-Net, HisToGene, TRIPLEX), graph-based (Hist2ST), retrieval (BLEEP, EGN), multi-scale (M2OST, THItoGene), diffusion (Stem), and flow matching (STFlow). This is the most thorough baseline coverage in this field. Clean ablations. Table 3 isolates each component (B → B+M → B+P → Ours); Table 2 does hyperparameter analysis for MM M and KK K; the ρ200 gains from M and P individually are comparable and combine to give the full method — convincing that both components contribute.

    Interpretability analysis is exceptional. Fig. 4 shows gene binding modules recovering block-diagonal structure (Mod1 = secretory/adhesion, Mod3 = cytoskeletal/metabolic, Mod4 = ECM/signaling, Mod5 = stromal/immune, Mod0 = immunoglobulins), validated against Enrichr-detected pathways including EMT, antigen processing, and neutrophil degranulation. Regulator tokens show the expected complementary pattern (broad, cross-module attention). This is unsupervised emergence of biologically sensible structure and is a significant result beyond prediction accuracy.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    (W1) Heteroscedastic flow matching vs. UA-Flow / Kendall–Gal positioning The heteroscedastic-NLL loss (Eq. 5) is exactly Kendall & Gal’s classical formulation, applied to the velocity regression of flow matching. The paper cites Kendall & Gal [8] but frames “heteroscedastic flow matching” as if it were a new generative-modeling paradigm (“introduces heteroscedastic flow matching” is implicit in the framing). In reality:

    The loss is Kendall–Gal. Using σθ\sigma_\theta σθ​ as the SDE diffusion coefficient (Eq. 6) for per-gene adaptive sampling is the actual novel technical move in this paper.

    UA-Flow (arXiv 2602.10326, Feb 2026) concurrently proposes heteroscedastic velocity-field modeling for image generation; the paper may or may not have been aware of this depending on its submission date.

    The paper’s actual contribution — Kendall–Gal-style heteroscedastic output combined with per-gene SDE coefficient for uncertainty-aware generation — is novel in the ST context but should be positioned more precisely. Please distinguish the output-space per-gene formulation from UA-Flow’s velocity-space formulation in the revision. (W2) Uncertainty calibration is not directly evaluated The paper motivates heteroscedastic modeling with “well-calibrated” per-gene uncertainty (Fig. 1c). But no calibration metric is reported: no per-gene uncertainty-vs-error scatter, no reliability diagram, no prediction interval coverage (PICP / mean interval width), no negative log-likelihood on held-out data. Without these, the claim that σθ\sigma_\theta σθ​ provides “calibrated uncertainty” (end of §2.4 Training) rests on convergence at optimality — but this is a training-time optimum argument, not an empirical verification. Please add:

    Scatter plot of σθ\sigma_\theta σθ​ vs. ∣v−u∗∣|v - u^*| ∣v−u∗∣ on held-out samples.

    90%/95% prediction interval coverage. Comparison with TISSUE (Sun et al., Nat Methods 2024) or a simple MC-dropout baseline.

    This is the single largest gap between the paper’s claimed contribution and its evidence. (W3) Sampling procedure details missing Eq. (6) dx=vθ(xt,t,h)dt+σθ(xt,t,h)dwdx=v_\theta(x_t,t,h)dt + \sigma_\theta(x_t,t,h)dw dx=vθ​(xt​,t,h)dt+σθ​(xt​,t,h)dw is an SDE integrator. Key details are missing:

    What SDE solver? Euler–Maruyama? Higher-order? Number of steps? How many samples are drawn at inference? Is the reported PCC from a mean of multiple samples? Does stochastic sampling help or hurt PCC relative to deterministic flow matching at inference (i.e., σθ=0\sigma_\theta=0 σθ​=0)?

    Without an ablation on stochastic vs. deterministic inference, the claim that per-gene adaptive stochasticity helps is not directly supported. Please add a table comparing (i) deterministic inference, (ii) uniform-σ\sigma σ SDE inference, and (iii) heteroscedastic σθ\sigma_\theta σθ​ SDE inference, all with the same trained model.

    (W4) Pre-trained encoder ambiguity §3 states: “For each spot, we extract patch-level features from two pre-trained pathology foundation models, UNI and CONCH, and concatenate them as the morphological input h∈RDfmh\in\mathbb{R}^{D_{fm}} h∈RDfm​.” This is a non-trivial advantage — BLEEP and TRIPLEX typically use UNI alone, STFlow uses UNI, Stem uses UNI+CONCH. Table 1 does not clarify whether baselines are run with their native encoders or a unified UNI+CONCH input. If BLEEP and TRIPLEX are run with UNI only, the comparison is not strictly fair. Please clarify the encoder setup per baseline, and ideally run the strongest baselines with UNI+CONCH to isolate the encoder contribution from the architectural contribution.

    (W5) “ρ_k averages over top-k best-predicted genes” is a subtle definition The paper reports ρ_10, ρ_50, ρ_200 where each is the mean PCC over the top-k genes ranked by PCC (best-predicted for each method individually). This is the convention established by ST-Net but has a subtle failure mode: two methods’ top-10 gene sets may differ entirely, so ρ_10 compares different gene subsets. The gains on ρ_200 are consistent with the paper’s story, but a fixed gene-set evaluation (e.g., all marker genes, or HVGs fixed per dataset) would be more rigorous. Please add at least one evaluation with a fixed gene set.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    If the authors provide anonymous code link, I can change my rating to accept. (“Talk is cheap, show me the code”) The three-component design (gene binding modules, regulator tokens, heteroscedastic flow matching) is well-motivated by three distinct biological observations, each component is ablated, and each shows complementary gains. The interpretability analysis (Fig. 4) recovering biologically coherent gene modules without supervision is an exceptional result that goes beyond pure prediction accuracy and suggests the learned representations capture meaningful biology. The baseline comparison is the most thorough I have seen in this field.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank the reviewers. Many concerns reflect detail elided under page limits; anonymized code will be released after acceptance.

A. Clarifications (R3-W3,4,5) Sampling (W3): Heun 2nd-order SDE solver, 50 steps, Tweedie final-step correction; 20 samples/spot, PCC is the per-spot mean. σ_θ ∈ [10⁻³,1] is the per-axis SDE coefficient (Eq. 6). To §3.Encoders (W4): All baselines use author-provided native setups. Stem [23] uses the same UNI+CONCH as ours, yet our ρ_200 is much higher on all four datasets (BC-I .169→.352, BC-II .419→.545, SCC .690→.794, Mouse .374→.500); gains cannot reduce to encoder choice. Encoders to Table 1.Protocol (W5): Top 200 HVGs are selected per dataset in preprocessing, and all methods are trained/evaluated on this identical panel. ρ_200 is therefore already a fixed-gene-set evaluation; ρ_10/ρ_50 pick each method’s top-k from the same fixed pool. Largest gains (+42%/+10%/+15%/+27%) appear at ρ_200.To §3. B. UA-Flow (R3-W1) Eq. 5 follows Kendall-Gal [8]; novelty is (i) coupling it to flow-matching velocity regression and (ii) using learned per-gene σ_θ as the per-axis SDE coefficient (Eq. 6). UA-Flow (Han et al., 2026; concurrent) defines per-sample velocity uncertainty and propagates it through Jacobian dynamics for post-hoc reliability under deterministic ODE; we use per-gene σ_θ directly as SDE coefficient for intrinsically stochastic gene-modulated inference. §2.4 will clarify.

C. Calibration (R3-W2) We agree the absence of direct calibration metrics is the most important gap. Indirect support: at NLL optimum σ_θ converges to true per-gene squared error; Table 3 B+P→Ours isolates the heteroscedastic-loss contribution with positive ρ_200 deltas (BC-II +0.036); σ_θ functions as SDE coefficient without degrading top-k metrics. Direct calibration (σ-vs-error scatter, PICP, MC-dropout) is the right verification and will be pursued in future work.

D. Generalization (R1-W1, R2-W1) Cross-cohort (R1-W1): BC-I and BC-II are already independent cohorts (§3: BC-II “an independent cohort with different tumor subtypes”); consistent gains on both (ρ_200 .352/.545) demonstrate cross-subtype transfer. Leave-one-out CV across sections tests within-cohort generalization. σ_θ aggregates over the full training distribution, not a HER2+ prior. Strict train-A-test-B cross-cohort is new experimental work and an important next step we pursue in future work (Dawood et al., 2026). Pathway across tissues (R2-W1): Fig. 4 covers BC; pathway-enrichment on BC-II/SCC (shared cancer framework) and module-to-Allen-Brain-Atlas correspondence for Mouse Brain (hallmark databases being cancer-centric) are planned as future work.

E. Sensitivity (R2-W1, R1-W2) K modules (R2-W1): Fig. 4 justifies K=6: modules learned without supervision recover gene groupings (secretory/adhesion, cytoskeletal/metabolic, ECM/signaling, stromal/immune, immunoglobulins) mapping to hallmark pathways (EMT, antigen processing, neutrophil degranulation). Exact K matters less than the decomposition: K=2,4,6,8 all improve substantially over the no-decomposition baseline (BC-I ρ_200 .169→.347 at K=2). K is empirically chosen; we welcome principled K-selection as an open direction for future work. M tokens (R1-W2): The ρ_200 variation across M ∈ {4,8,16,32} in Table 2 is small (BC-I 4%, BC-II 6%, Mouse 2%). Crucially, M=16 is optimal on all three datasets simultaneously (breast cancer and mouse brain, two distinct domains), indicating a structural property of inter-gene coordination rather than dataset-specific tuning.

F. MSE/MAE trade-off (R2-W2) The trade-off is clearest in STFlow† (deterministic flow matching): on Mouse, STFlow† has the lowest MSE (.415)/MAE (.491) yet a low ρ_200 (.289 vs our .500); same on SCC. Minimizing pointwise error smooths expression and collapses the localized patterns ρ_k captures (Fig. 3). For downstream uses (niche identification, marker discovery), spatial fidelity and correlation are more biologically actionable. To §3.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Paper Summary The paper proposes a generative framework for predicting spatial gene expression from histology images by modeling gene-specific relationships to morphology and inter-gene dependencies. It introduces gene binding modules for gene-selective morphological conditioning, regulator tokens for capturing gene–gene interactions, and heteroscedastic flow matching to model per-gene uncertainty and adaptive stochasticity during generation.

    Key Strengths The method is well-motivated and moves beyond standard regression by modeling the conditional distribution of gene expression. The design is coherent, with each component aligned to biological observations, and the empirical evaluation is strong with broad baselines and consistent improvements in correlation metrics. The interpretability analysis is particularly compelling, showing biologically meaningful gene groupings emerging from the model.

    Key Weaknesses The method raises concerns about generalization due to lack of cross-cohort validation and potential dataset-specific bias. Some modeling assumptions, such as fixed morphological programs and sensitivity to hyperparameters, are not fully justified. There are also missing evaluations for uncertainty calibration, incomplete details on sampling procedures, and fairness concerns in experimental setup. Performance on pointwise error metrics is less competitive.

    Review Summary All reviewers agree the paper presents a meaningful and well-structured contribution with strong empirical results and clear biological motivation. They also consistently note limitations in generalization, justification of design choices, and missing evaluations. While these concerns affect completeness, they do not outweigh the overall contribution and impact.

    Decision Provisional Accept. Please clarify and address the reviewers’ concerns, particularly regarding key design choices such as morphological programs and regulator tokens. Improve the positioning of the heteroscedastic formulation relative to prior work. Provide clearer implementation and sampling details, and ensure a transparent description of the experimental setup for fair comparison in the camera-ready version.



back to top