List of Papers Browse by Subject Areas Author List
Abstract
Spatial transcriptomics (ST) prediction from histology requires modeling sparse, zero-inflated, and intrinsically non-negative gene expression. However, existing generative models evolve in unconstrained real-valued spaces, allowing trajectories to leave the biological valid support of count-based data and degrading spatial structure modeling. We propose BioFlow, a support-preserving flow-matching framework that embeds non-negativity directly into the generative dynamics via velocity reparameterization. By constraining the generative dynamics rather than correcting invalid outputs post hoc, BioFlow maintains trajectory-level support consistency and aligns generation with the intrinsic geometry of sparse expression data. Across multiple public ST datasets, BioFlow achieves state-of-the-art spatial correlation performance while improving training stability and computational efficiency. These results underscore the importance of support-consistent generative dynamics for reliable modeling of sparse biomedical count data.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1400_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hrterry/BioFLow
Link to the Dataset(s)
PRAD READ datasets: https://huggingface.co/datasets/MahmoodLab/hest
BibTex
@InProceedings{XuHao_BioFlow_MICCAI2026,
author = { Xu, Haoran AND Liu, Yang AND Yuan, Wei AND Han, Xiao},
title = { { BioFlow: A Biologically Valid Support-Preserving Flow for Histology-Conditioned Spatial Transcriptomics Prediction } },
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
This paper proposes the BioFlow framework, which introduces a dynamics-based velocity field reparameterization method. While preserving computational efficiency, it mathematically enforces strict constraints on non-negativity and sparse support domain in the spatial transcriptomics prediction pipeline. It not only resolves the spatial pattern distortion caused by traditional generative models deviating from the biological manifold in their generation trajectories, but also experimentally validates its superior performance in balancing prediction accuracy and training stability for complex sparse data modeling.
- 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.
- Highly insightful mathematical reparameterization (novel formula design): The authors acutely identify the “geometric domain misalignment” problem plaguing existing generative models (e. g. , Flow Matching/Diffusion Models) when modeling non-negative, sparse transcriptomics data. The core innovation lies in embedding the non-negativity constraint directly into the velocity field prediction. By introducing the Δ parameter and Softplus activation, the authors construct a dynamical evolution formula that is non-negative by design under Euler integration. This approach, which corrects model biases from the fundamental mathematical mechanism rather than post-hoc fixes, holds significant academic value for processing constrained biomedical data.
- Flexible architecture design (excellent model scalability): The core support-domain-preserving dynamics module is fully decoupled from the backbone network, enabling seamless replacement with any state-of-the-art feature extraction architecture.
- Achieves state-of-the-art prediction accuracy across all datasets and at the whole-gene level. Its performance robustness for sparsely expressed low-abundance genes is markedly superior to all baselines, while delivering orders-of-magnitude efficiency gains. Compared to mainstream diffusion models, training time is reduced from thousands of GPU hours to minutes, drastically lowering the hardware and time barriers for real-world deployment.
- 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 authors failed to control variables when presenting the generation trajectories in Figure 1.B. The initial random noise and final prediction values are inconsistent across methods, making this comparative demonstration unfair. It is clearly visible that most initial noise samples for the diffusion model lie below zero, while those for vanilla Flow Matching, BioFlow, and the ideal curve all lie above zero. The authors should re-plot the diffusion model results using a consistent initial distribution to eliminate the confounding effect of initial noise distribution on perceived performance. Furthermore, vanilla Flow Matching and BioFlow were not evaluated on the same batch of initial data and prediction targets.
- The visualization results in Figure 1.D do not live up to the claims in the text. For instance, TRIPLEX exhibits qualitatively better performance, which raises concerns about potential inflation of quantitative metrics. The authors need to provide more compelling evidence to substantiate their superiority claims.
- The presented work is merely a non-negativization modification to vanilla Flow Matching, which lacks sufficient conceptual novelty.
- No hyperparameter ablation study was conducted, particularly for the hyperparameter Δ.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This work has three core strengths: it introduces a mathematically insightful non-negativity reparameterization to fundamentally solve geometric domain misalignment in generative models for sparse transcriptomics data; it features a flexible decoupled architecture supporting backbone replacement; and it achieves state-of-the-art whole-gene prediction, superior low-abundance gene robustness, and orders-of-magnitude faster training. However, it has critical limitations: unfair comparisons in Figure 1.B due to inconsistent initial noise and targets; contradictory qualitative visualization in Figure 1.D raising metric concerns; insufficient conceptual novelty as a mere non-negativization modification to vanilla Flow Matching; and no hyperparameter ablation, especially for the key Δ parameter.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The authors propose a method to improve gene expression prediction from histology when training from spatial transcriptomics data by applying an inductive prior that gene counts are sparse and non-negative. Specifically, the authors achieve this by inducing a “positivity-preserving residual flow” within their Conditional Flow Matching (CFM) generative framework.
- 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.
-Simple adjustment that is well-motivated by biological prior. -Robust validation using multiple datasets and compared against multiple competing methods -Training speed is quantified and compared against competing methods.
- 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 authors observe that gene counts are non-negative and sparse, however they do not propose a solution for enforcing sparsity.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall this paper proposes a relatively simple fix to improve gene expression prediction from histology, and the results are presented in a clear and convincing manner. The conclusion that “support-preserving dynamics enhance stability and generalization without architectural changes” seems significant too me, and beneficial to the MICCAI community.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After the author response I am still confident that this paper provides a meaningful improvement to a relevant task for this conference.
Review #3
- Please describe the contribution of the paper
The paper proposes BioFlow, a support-preserving conditional flow-matching framework for predicting spatial gene expression from H&E histology. The core idea is to reparameterize the learned velocity field so that forward-Euler integration stays within the non-negative orthant throughout the trajectory, rather than clamping/projecting only at t=1t=1 t=1 as in prior work.
- 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.
Clean, well-motivated contribution. Non-negativity of sparse zero-inflated count data is a genuine mismatch with unconstrained continuous flows, and the trajectory-level vs. output-level distinction (post-hoc clamping vs. support-preserving dynamics) is a conceptually sharp argument. Fig. 1B is pedagogically effective. Nice mathematical derivation. Eq. (7) gives the sufficient condition vθ(xt,t)≥−xt/hv_\theta(x_t,t)\geq -x_t/h vθ(xt,t)≥−xt/h; Eq. (8) reparameterizes as v^=−xt/Δ+ϕ\hat v = -x_t/\Delta + \phi v^=−xt/Δ+ϕ with ϕ≥0\phi\geq 0 ϕ≥0; Eq. (9) shows that under 0<h<Δ0<h<\Delta 0<h<Δ the Euler update is a positive convex combination. The derivation is simple, correct, and directly actionable.
Strong experimental evidence. Consistent PCC gains across three datasets (HER2ST, PRAD, READ) and three gene strata (H/M/All), with larger margins in the harder “All” stratum (+27–62% relative on HER2ST, +36–50% on PRAD). The pattern that gains grow with difficulty is exactly what one would expect if support preservation helps more for sparse, boundary-proximal genes. Impressive efficiency. 18 min on PRAD vs. 4,000 GPU-hours for STEM on comparable hardware (Table 2). 3–23× faster than MERGE/STFlow/TRIPLEX on HER2ST. This is a meaningful practical contribution. Clean ablations. Table 4 directly compares clamping vs. softplus-output vs. BioFlow on the same STFlow backbone — this is the cleanest way to show that output-level fixes are insufficient. Table 5 shows the constraint improves both time and accuracy. Table 6 isolates architectural capacity from the constraint. Anonymous code release is provided. Honest trade-offs. MSE/MAE are higher than some deterministic baselines, and the authors explain why (deterministic regression fits the conditional mean; generative methods fit a distribution). This is mature reporting.
- 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) The theoretical guarantee is specific to forward Euler and to 0<h<Δ0<h<\Delta 0<h<Δ — strengths and limits under-stated
Eq. (7–9) guarantees non-negativity only under forward Euler integration with step h∈(0,Δ)h\in(0,\Delta) h∈(0,Δ). In implementation h=0.05h=0.05 h=0.05, Δ=0.1\Delta=0.1 Δ=0.1, giving 20 integration steps. A few concerns:
Higher-order integrators (RK2/RK4, adaptive step-size, Dormand–Prince) are standard in flow-matching inference and do not inherit the guarantee. The paper does not discuss this. If h≥Δh\geq \Delta h≥Δ the guarantee breaks; why is Δ\Delta Δ treated as a fixed hyperparameter rather than derived from hh h? The choice Δ=2h\Delta=2h Δ=2h is arbitrary.
The guarantee is for the discretized sampler, not the continuous-time ODE. The paper claims “intermediate states remain within the biologically valid expression manifold” (§2.3) — strictly, the continuous trajectory of v^\hat v v^ is not guaranteed to stay in the positive orthant unless −x/Δ+ϕ(x,t)-x/\Delta + \phi(x,t) −x/Δ+ϕ(x,t) pointed into the orthant at the boundary, which requires ϕ(0,t)≥0\phi(0,t)\geq 0 ϕ(0,t)≥0 (softplus gives this, good) AND the radial decay −x/Δ-x/\Delta −x/Δ does not drag xx x negative — which is true because −x/Δ=0-x/\Delta = 0 −x/Δ=0 when x=0x=0 x=0.So the continuous guarantee actually holds too; the paper should make this argument explicitly rather than hiding behind “discretization-aware” language.
Forward Euler with 20 steps is the most basic integrator. How does BioFlow compare to STFlow/STEM when those use their native integrators (often adaptive or many-step)?
(W2) The “structural distortion” argument is not directly demonstrated The motivation story is that boundary violations along the trajectory distort the learned vector field and thus hurt spatial fidelity. Table 4 shows that clamp/softplus on STFlow hurt PCC, but this is consistent with two stories: (a) trajectory-level violations degrade vector field learning; (b) output-level fixes project predictions to bad local regions. The paper does not measure how often standard flow matching trajectories violate non-negativity, nor show a direct link between violation rate and PCC. A diagnostic figure plotting fraction of negative intermediate values at each timestep (for STFlow vs. BioFlow) would make the mechanistic argument concrete. (W3) Residual vs. drift form: the −x/Δ-x/\Delta −x/Δ term is essentially exponential decay
Eq. (9) can be read as xt+h=(1−h/Δ)xt+hϕx_{t+h} = (1-h/\Delta)x_t + h\phi xt+h=(1−h/Δ)xt+hϕ, i.e., a leaky-integrator update. If ϕ≡0\phi\equiv 0 ϕ≡0, the trajectory collapses to zero geometrically fast. This means the model is “biased toward zero” by construction — which for a sparse, zero-inflated target is helpful, but it is also a strong implicit prior that should be characterized. What does ϕ(x,t)\phi(x,t) ϕ(x,t) actually need to do to recover non-zero target values? The paper doesn’t show that ϕ\phi ϕ is well-scaled; if Δ\Delta Δ were much smaller, the decay would dominate the network. This feels like an unexamined hyperparameter choice.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
BioFlow makes a genuinely novel contribution: it is, to my knowledge, the first paper to reparameterize the flow-matching velocity field to enforce non-negativity throughout the sampled trajectory. The derivation is clean, the empirical gains are consistent, the efficiency claims are impressive, and the ablations are generally well-designed
- 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
Author Feedback
We thank the reviewers for thoughtful feedback and are pleased that our work was recognized for its mathematical insight (R1, R3), robust, strong experimental evidence (R2, R3), and significant, genuinely novel contribution (R2, R3). We address the main concerns below. Visualization fairness(R1,R3). Fig. 1B is a conceptual illustration, not a controlled quantitative comparison. Its purpose is to show that native formulations alone are insufficient: STFlow uses a non-negative ZINB count prior, yet its unconstrained velocity can still produce negative states/outputs; STEM, as a diffusion model, naturally involves Gaussian noising/denoising. BioFlow instead enforces support preservation throughout Euler integration. Quantitative evidence is in Tables 1-2 and 4-6.We thank R1 for noting the inconsistency in Fig. 1D. After rechecking the source, we found a LaTeX filename typo: the TRIPLEX panel was inadvertently duplicated from GT, explaining why it appeared visually stronger than the reported PCC. We apologize for this presentation error. It affects only this visualization; all quantitative results and ablations are unaffected. We also thank R3 for the careful suggestion of timestep-wise diagnostics for negative intermediate values and their relation to PCC. This would be a useful mechanistic diagnosis, but under the rebuttal policy, we cannot provide new statistics/figures here. Note that violation frequency should not be expected to correlate directly with final PCC, since it can depend on initialization, timesteps, and sampling path, whereas PCC is measured on final predictions. Novelty(R1). BioFlow is not merely a non-negativization of vanilla flow matching. As R3 also noted, to our knowledge, BioFlow is the first work to reparameterize the flow matching velocity field to enforce non-negativity throughout the flow trajectory, rather than relying on post-hoc output correction. Table 4 directly tests this distinction: STFlow+clamp/softplus ensure non-negative outputs but reduce PCC(H) from 0.503 to 0.329/0.308, whereas BioFlow reaches 0.571 by constraining the dynamics. Thus, the key novelty is embedding a trajectory-level support-preserving inductive bias for sparse biological data directly into the dynamics. Sparsity(R2). We agree that BioFlow does not explicitly enforce hard zeros. Our guarantee is for non-negative support preservation, not exact sparsity. Sparsity and zero inflation motivate the constraint because many ST entries lie at or near the non-negative boundary. Explicit sparsity regularization is orthogonal to the proposed support-preserving flow. Hyperparameters h,Δ(R1, R3). We agree that Δ should be better documented. Due to space limits, the full ablation was not included in submission. As stated in Sec. 3.2, our results are insensitive to h,Δ choices as long as h<Δ holds. All reported results use the fixed setting h=0.05,Δ=0.1 across datasets, rather than dataset-specific tuning. This was chosen for efficiency, as smaller valid pairs require more integration steps with no benefit. Even for smaller Δ, a correspondingly smaller h keeps performance comparable, suggesting that the leaky decay term −x/Δ does not dominate. ϕ(x,t)≥0 robustly serves as a non-negative residual recovery term. Theory(R3). We thank R3 for the careful reading of our theory. In continuous time, BioFlow defines v_hat(x,t)=-x/Δ+ϕ(x,t), with ϕ≥0.If x_i=0, v_hat_i(x,t)=ϕ_i(x,t)≥0, so the vector field is non-outward on the boundary of the non-negative orthant. Therefore, with x_0≥0, the continuous flow cannot leave the orthant. During inference, BioFlow/STFlow follow standard Euler integration, while STEM follows native diffusion inference. We apply x_{k+1}=(1−h/Δ)x_k+hϕ(x_k,t_k), which preserves non-negativity for 0<h<Δ. Arbitrary RK/adaptive solvers do not automatically inherit this inference-time guarantee as their intermediate stages are not necessarily positivity-preserving convex updates. We hope this clarifies the solver-dependent scope.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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 This paper addresses the problem of predicting spatial gene expression from histology images under the constraint that transcriptomic data are sparse and non-negative. Existing generative models operate in unconstrained spaces, which can lead to biologically invalid intermediate states and degraded spatial modeling. To address this, the paper proposes BioFlow, a support-preserving conditional flow-matching framework that reparameterizes the velocity field to enforce non-negativity throughout the generative trajectory rather than only at the output. The method integrates this constraint into the dynamics and demonstrates improved prediction accuracy, training stability, and computational efficiency across multiple spatial transcriptomics datasets.
Key Strengths The paper is recognized for a well-motivated formulation that incorporates biological priors directly into the generative process through a mathematically grounded reparameterization. The method is conceptually clean and integrates seamlessly with existing architectures, offering flexibility and scalability. Empirical evaluation is extensive, covering multiple datasets and baselines, with consistent improvements in prediction performance, particularly for challenging sparse gene settings. The approach also demonstrates substantial gains in computational efficiency, significantly reducing training time compared to competing generative models. Theoretical derivation and ablation studies are generally viewed as clear and supportive of the proposed design.
Key Weaknesses Concerns are raised regarding the level of conceptual novelty, with the method viewed by some as a relatively simple modification of existing flow matching approaches. Several issues are identified in experimental validation, including potentially unfair comparisons due to inconsistent initial conditions and discrepancies between qualitative visualizations and quantitative results. The theoretical guarantees are considered limited in scope and insufficiently discussed, particularly regarding integration methods and parameter choices. Additionally, key aspects such as hyperparameter sensitivity and sparsity enforcement are not thoroughly explored, and some claims about structural advantages are not directly substantiated through targeted analysis.
Review Summary The reviewers agree that the paper tackles an important problem and that incorporating non-negativity into generative dynamics is a meaningful direction. There is also consensus that the method is well-motivated and shows strong empirical performance with notable efficiency improvements. However, there is significant disagreement regarding the level of novelty and the reliability of the experimental evidence. While some reviewers find the formulation novel and the results convincing, others question whether the contribution goes beyond a minor modification and highlight concerns about fairness and consistency in evaluation. There are also differing views on the completeness of the theoretical analysis and whether the supporting experiments adequately validate the claimed advantages.
Decision Invite to rebuttal. The paper shows promising performance and a clear motivation, but concerns about novelty, experimental fairness, and theoretical clarity need to be addressed before a final decision.
Rebuttal Focus The rebuttal should primarily address concerns by clarifying the novelty of the proposed method relative to existing flow matching approaches, particularly whether the contribution goes beyond a simple modification. The authors should also justify the fairness of experimental comparisons, especially in trajectory visualization and benchmarking settings, and explain any discrepancies between qualitative and quantitative results. Additional clarification on theoretical guarantees, including integration assumptions and parameter choices, would strengthen the paper. Finally, providing discussion on hyperparameter sensitivity and the role of sparsity in the method would help resolve remaining concerns about completeness and rigor.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal satisfactorily addresses the main concerns regarding novelty, experimental fairness, and theoretical scope. The authors clarify the role of the disputed figures, provide a reasonable explanation for the visualization inconsistency, and better articulate the novelty of trajectory-level support-preserving dynamics compared with output-level non-negativity corrections. While some limitations remain, the paper presents a well-motivated methodological contribution with consistent empirical gains, strong efficiency improvements, and support from two reviewers. Overall, I believe the strengths outweigh the remaining weaknesses and recommend acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers appreciate the motivation and extensive experimental results, and one author has concerns regarding the implementation details. The initial concerns have been mostly addressed during the rebuttal. I recommend accepting this submission.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
After rebuttal, R2 strongly supports acceptance, while R1 maintains Reject and R3 did not provide a post-rebuttal update. Given the unresolved concerns from R1 and the lack of a clear positive post-rebuttal majority, the paper remains slightly below the acceptance threshold.
