acceptodds
Under review as a conference paper at ICLR 2027

MANGO: Flow Matching on the Multiplicative Geometry of Gene Expression for Spatial Transcriptomics Prediction from Histology

10 views

Abstract

Predicting spatial molecular profiles from routine H&E images can extend spatial omics to large pathology cohorts, but existing methods commonly rely on Euclidean generative paths, fixed tissue graphs, and uniformly weighted gene objectives. These choices overlook the multiplicative geometry of nonnegative gene expression, prevent tissue neighborhoods from adapting to the evolving molecular state, and bias learning toward abundant genes. We propose MANGO, an expression-conditioned Riemannian graph flow for histology-based spatial-omics prediction. MANGO performs conditional flow matching on the multiplicative geometry of gene expression and predicts the clean expression endpoint, yielding closed-form geodesic corruption and geometry-preserving denoising. A dynamic multi-view graph Transformer jointly models spatial proximity, morphological similarity, and the current multiplicative expression state, and reconstructs the graph layer by layer. An expression-aware Riemannian objective further balances supervision across genes. Across 17 transcriptomics datasets spanning spot-level, Visium HD, and cell-level resolutions, MANGO ranks first on 10 and second on the remaining seven, while achieving the highest average PCC in all four evaluation groups. It also leads in four representative all-gene settings and reaches PCC values of 0.409 and 0.292 on spatial metabolomics and proteomics, surpassing the strongest baselines by 0.176 and 0.088, respectively. These results demonstrate that coupling expression-aware geometry with adaptive tissue structure enables accurate and resolution-robust spatial-omics inference from histology.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial transcriptomics, flow matching, manifold learning, spatial gene expression prediction

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

This paper proposes an expression-conditioned Riemannian graph flow for predicting spatial omics from histology. The method performs conditional flow matching under the multiplicative geometry of gene expression, using a dynamic graph Transformer to predict the clean expression endpoint. Its neighborhoods and edge messages integrate spatial proximity, morphological similarity, and evolving expression estimates. An expression-aware Riemannian objective balances supervision across genes. Experiments across multiple datasets demonstrate the method’s effectiveness.

Critical strengths

The paper identifies three limitations of existing methods: unconstrained generative trajectories, fixed graph neighborhoods, and uniformly weighted training objectives. It presents three corresponding contributions to address these limitations.

Critical weaknesses
  1. The paper does not clearly analyze how its key components, including geometry-aware conditional generation, dynamic multi-view tissue modeling, and the expression-weighted Riemannian objective, differ from existing approaches.

  2. The implementation details of the ablation studies are insufficiently described, making it difficult to assess how each component’s contribution is isolated.

Questions for authors
  1. Since the proposed metric is globally isometric to Euclidean space through ϕ(y)=log⁡(y+s)\phi(y)=\log(y+s), could you clarify the methodological distinction between your geometry-aware conditional generation and standard Euclidean flow matching applied to log-transformed expressions?

  2. What distinguishes the proposed dynamic multi-view tissue modeling from existing dynamic graph construction and edge-conditioned attention methods? Please clarify how candidate neighbors are constructed and whether the candidate sets themselves change during refinement.

  3. How does the expression-weighted Riemannian objective differ from conventional inverse-expression-weighted regression in log coordinates? What additional benefit does gene weighting provide beyond the normalization already introduced by the multiplicative metric?

  4. Could you specify the implementation of each variant in Table 4?

  5. Can you provide a controlled comparison between multiplicative and Euclidean transport while keeping the denoiser, dynamic graph, gene weighting, and training budget unchanged? Table 4 evaluates multiplicative edge information, but does not directly isolate the contribution of the transport geometry.

1 comment
AuthorsOpenReview ↗

Thank you for acknowledging that the ablation implementation and gene-weighting concerns have been addressed.

Thank you for acknowledging that the ablation implementation and gene-weighting concerns have been addressed. We provide further clarification and matched fold/seed results for the two remaining points.

1. Transport formulation. The exact Euclidean-coordinate representation is consistent with our Riemannian formulation: it describes the same model in another coordinate system. Following Chen and Lipman’s Flow Matching on General Geometries (ICLR 2024), we use an analytically tractable geometry to specify multiplicative conditional transport. Our contribution concerns this formulation and its coupling with expression-dependent tissue refinement, rather than a claim of intrinsic curvature.

The controls answer complementary empirical questions. Replacing only multiplicative corruption reduces PCC from 0.4671 to 0.3712 on Ovary and 0.3461 to 0.2571 on COAD, supporting this corruption choice within the pipeline. Standard Euclidean FM evaluates a complete alternative with its corresponding prediction target, objective, and integration scheme. We agree that its entire performance gap cannot be attributed solely to the path. We will distinguish these interpretations explicitly.

2. Dynamic neighbor reselection. To clarify the variability, we examined the matched fold/seed results. The reported standard deviations pool all 12 fold×seed evaluations, so they include between-fold heterogeneity. Averaging the three seeds within each fold gives:

FoldFullG1Full−G1
split00.36050.5104−0.1499
split10.34050.1481+0.1925
split20.34100.1849+0.1561
split30.34200.2279+0.1141
Overall0.34600.2678+0.0782

Full improves on three of four folds, with the same direction across all three seeds within each fold. Averaging over folds separately for each seed gives improvements of +0.0864, +0.0824, and +0.0658. Thus, the positive aggregate effect is reproduced across all three seeds, while G1’s large pooled standard deviation primarily reflects differences between folds.

G1 retains expression-edge updates while fixing neighbor indices, so this intervention specifically tests reselection. These matched results support its empirical benefit in the evaluated setting and identify split0 as a clear exception.

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

Authors propose Mango which predicts ST metabolic and proteomics from H&E slides. It models conditional distribution with flow matching and predicts expression endpoint and derives velocity from it. The backbone of the denoiser is a graph transformer whose neighbor graph is rebuilt at every layer from fused multi modal representations. Validation covers 17 transcriptomic datasets. paper reports the best group-average PCC in all four evaluation groups.

Critical strengths

The method is tested across platforms and resolutions with group-level gains in every benchmark metric.

The expression-conditioned dynamic graph uses sparse top-k neighbor retrieval which provides mechanism efficiency and gains.

Critical weaknesses

No ablation is shown which isolates the geometry. Table 4 ablates a manifold edge feature, not the log space flow. Authors have never shown comparison for log-space and raw-space flow, flow and deterministic regression on the same denoiser, or endpoint and velocity parameterization.

HistoPrism ICLR 2026 shows a plain transformer regressor can match generative models on pathway level prediction. GenAR Medical Image Analysis 2026 argues for count native modeling over log transforms. STPath, TRIPLEX, MoLF and HistoPrism are absent, FLAG is cited but not compared, and the multi omics and all gene tables use only four baselines. Please compare against these methods.

Table 15 (top-10 HVG) contradicts the “first or second everywhere” robustness claim. MANGO shows lower score compared to STFlow, MERGE and HINGE on many datasets, and the appendix does not discuss it.

Questions for authors

Is Y raw normalized counts or log1p-transformed? How was s=1 chosen?

Please report an ablation with the same denoiser for raw space flow, log space flow, velocity vs endpoint, and a deterministic regressor.

Please explain the identical values for HINGE PRAD and COAD, k=5 and k=8, and Visium HD Lung for SpatialEx and Stem.

In Table 2, IDC is listed with 8,270 genes but is labeled Xenium in Table 6.

How many samples are averaged for reported PCC? Please report sample diversity and calibration.

3 comments
AuthorsOpenReview ↗

# Additional all-gene baseline results

Following up on our earlier response, the GenAR all-gene experiments are now complete. We also provide the standard deviations for HistoPrism and update its IDC mean from 0.1273 to 0.1274 using the latest aggregate. Both methods follow the original evaluation protocol.

DatasetPlatformGenesFoldsUNIH0-miniSTFlowMERGEHistoPrismGenARMANGO
READVisium8,82840.114±0.0560.112±0.0730.105±0.1370.151±0.0470.1524±0.10840.1107±0.02570.186±0.045
IDC / BreastVisium HD8,27030.213±0.1060.171±0.0620.216±0.1090.246±0.1900.1274±0.05140.0336±0.02480.271±0.168

Existing baseline and MANGO entries follow Table 2. MANGO retains the highest mean PCC in both tasks. On READ, HistoPrism is the strongest baseline, giving an absolute PCC margin of approximately 0.0336. On IDC/Breast, MERGE remains the strongest baseline reported in Table 2 (0.246±0.190), giving a margin of 0.025. We will incorporate these results into the final manuscript.

Reviewer_CzgsOpenReview ↗

Staying at the score.

Given the mixed performance advantage (W3) and lack of clear ablations (W1), I am inclined to stay at my current score.

AuthorsOpenReview ↗

Thank you for considering our clarifications and additional results.

Thank you for clarifying your remaining concerns. We would like to make the evidence for W1 and the scope of our claim regarding W3 more explicit.

W1 — Transport and parameterization ablations. We agree that the original manifold-edge ablation did not isolate transport. The additional experiments address this separately:

ConfigurationOvary (Visium HD)COAD (Visium)
Multiplicative path + endpoint prediction0.46710.3461
Additive corruption only0.37120.2571
Standard Euclidean FM0.31170.2148
Direct regression0.31950.3257
Chart-velocity prediction0.37160.2395

To clarify exactly what these comparisons isolate:

  • Additive corruption only replaces the interpolation with (Y_t=(1-t)Y_0+tY_1), retaining the baseline endpoint objective and sampler. It tests the corruption choice within our pipeline.
  • Standard Euclidean FM uses additive interpolation, velocity prediction, the corresponding CFM loss, and ambient Euler integration. This is a complete alternative formulation, rather than a single-factor intervention.
  • Direct regression retains the denoiser backbone but removes corruption and time weighting, disables the expression-node input, and predicts in one step.
  • Chart velocity retains the multiplicative path, prior, and sampling steps, while replacing endpoint prediction with chart-displacement prediction and its MSE objective.

These results support the selected formulation on both tasks; we do not interpret them as disproving the metric’s Euclidean-coordinate equivalence or establishing superiority over every possible Euclidean formulation.

W3 — Mixed performance on top-10 HVGs. We agree that the original “first or second everywhere” statement was too broad unless explicitly restricted to the main top-50 setting. In Table 15’s top-10 setting, MANGO ranks first on 11 tasks, second on one, third on four, and fourth on one. It nevertheless has the highest group-average PCC in all four groups. Both observations should be reported together: strong aggregate performance does not imply uniform task-level dominance.

We will explicitly restrict the ranking claim and discuss the weaker top-10 results in the revised manuscript. We hope these clarifications make the ablation evidence and the limits of our performance claim easier to assess.

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

The paper proposes MANGO, a method for predicting spatial gene expression (and other omics) from H&E images. It applies flow matching in log space, with a graph transformer denoiser. In the main evaluation, MANGO is compared against 8 baselines using PCC on the top 50 highly variable genes. It ranks first on 10 tasks and has the best average in all four resolution groups. Further experiments cover all-gene prediction, transfer to metabolomics and proteomics, ablations on COAD with two encoders, a CCL19 case study, and runtime and memory scaling.

Critical strengths

Broad evaluation. The benchmark spans spot, Visium HD and cell resolution, several platforms and organs, all-gene panels, and two non-transcriptomic modalities.

Critical weaknesses

Limited methodological novelty.

  • The “Riemannian” flow is indeed Euclidean flow matching after a log(y+1) transform. With one global chart and the pulled-back Euclidean metric, the manifold is flat and isometric to RG\mathbb{R}^G. The geodesics, the “wrapped” prior and the loss are all Euclidean in log coordinates. In Eq. 10 the metric factor simply cancels. Log-transforming expression is standard practice, and the paper’s own evaluation uses log1p. The domain y > −1 also does not enforce the nonnegativity used as motivation.
  • Lack of innovation compared to prior art. STFlow already uses local k-NN attention conditioned on the noisy expression and on neighbor expression differences, and averages the layer-wise predictions. The edge-aware attention follows standard graph Transformers (Graphormer, TransformerConv).
Questions for authors
  1. How is the candidate set defined in the top-k neighbor selection?
  2. App. E conduct a hyperparameter search for γ\gamma, kk, wpw_p and wmw_m on SKCM and LUNG. Both are Table 1 test tasks with only two slides each. The chosen values reproduce the Table 1 numbers exactly (0.701 and 0.618). Please also suggest hyperparameter selection for the test slide without ground truth (e.g., use MANGO as a discovery tool on a new slide without ST profiling).
  3. What is the patch size in pixels, the resolution (μ\mum/px), and any resizing before the encoder, for each resolution regime? At cell level, how are patches defined? At bin or cell resolution, is there a heavy overlap between neighboring patches, so the morphology distance may mostly repeat the spatial distance?
  4. For HEST-Bench, are these HEST’s official top-50 gene lists? For Visium HD and cell level, were the 50 genes chosen per fold from training slides only, or with the held-out slide included?
  5. Table 6 lists IDC as Xenium, but Table 2 reports 8,270 genes for IDC. That is more than any Xenium panel holds (the largest has about 5,000). Which slides and genes are used?
  6. HINGE has identical PRAD and COAD results (0.351 ± 0.063). SpatialEx and Stem are identical on Visium HD Lung (0.343 ± 0.099). Are they floating-point precision coincidences?
1 comment
AuthorsOpenReview ↗

Thank you for considering our clarifications and additional results.

Thank you for considering our clarifications and additional results. We appreciate that your remaining concern is methodological novelty.

Our intended contribution is a specific coupling of multiplicative conditional transport and expression-dependent tissue refinement. The endpoint estimate serves two connected roles: it parameterizes the conditional flow and updates the expression relations used to select neighbors and construct edge messages at the next graph layer. Spatial locality, morphology, and the evolving expression estimate therefore jointly determine how tissue context is propagated during denoising.

We recognize that Riemannian flow matching, dynamic graphs, and edge-conditioned attention have established foundations. Following Chen and Lipman’s Flow Matching on General Geometries (ICLR 2024), we deliberately choose an analytically tractable geometry with closed-form paths and sampling. Its flatness is compatible with this purpose; introducing curvature would add complexity without an established benefit for this task. We will sharpen the manuscript’s positioning around the task-specific formulation and coupling, rather than suggest that the metric or individual attention operations are themselves new.

The added controls provide evidence that these choices matter empirically: replacing only the multiplicative corruption reduces PCC from 0.4671 to 0.3712 on Ovary and from 0.3461 to 0.2571 on COAD. Fixing the fused adjacency while retaining expression-edge updates reduces COAD PCC from 0.346 to 0.2678. These results support the roles of the transport choice and adaptive topology within the proposed system, while not constituting a claim of mathematical novelty by themselves.

We hope this more precise positioning clarifies the methodological contribution and distinguishes it from the established ingredients on which it builds.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.