{"id":"cc0a6f62-16f2-4ba3-9168-dd46d212a653","arxiv_id":"2506.05361","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"STFlow uses whole-slide flow matching with local spatial attention to jointly predict gene expression across all spots in a histology image, outperforming prior spot-wise and slide-wise baselines on two benchmarks.","lead":"STFlow is a machine learning model that predicts gene expression at thousands of tissue spots from a standard H&E-stained histology image by iteratively refining initial guesses with a flow matching process. If reliable, it could let researchers obtain spatial gene expression data from routine pathology slides without costly specialized sequencing, which may accelerate cancer biomarker discovery.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark curation (STImage organ filter + older COAD) gates the 'substantially outperforms SOTA' claim; rerunning on official protocols is needed before acceptance.","rationale":"I read the paper as making a concrete empirical claim: STFlow beats SOTA on two public collections. The internal comparisons are plausible: same features, same splits, three seeds, disclosed hyperparameters, consistent average gains, and ablations showing the components help. I do not see an internal inconsistency in the flow-matching objective or the E(2)-invariant attention that would invalidate the method. The weakest link is the external validity of the benchmark. The authors are transparent about the STImage organ filter and the COAD version, and both choices can have legitimate reasons, such as avoiding datasets where no method performs above noise; however, filtering on a correlation threshold makes the average comparison depend on which organs are included, and the COAD version mismatch prevents direct comparison with the published HEST-1k leaderboard. These are exactly the conditions under which a reader should accept the result conditionally rather than as an unconditional SOTA result. The reader's verdict already captures this, so I do not move the verdict; I agree with the reader's weakest assumption.","tokens_in":20973,"tokens_out":9592,"duration_ms":103191,"concrete_test":"Run STFlow and the strongest baselines (UNI, TRIPLEX, Gigapath-slide) on the full STImage-1K4M organ list without the >0.1 filter and on the current HEST-1k COAD version, using the same feature extractors, seeds, and hyperparameter search. Report per-organ Pearson correlations for the excluded organs. If STFlow's average margin over TRIPLEX/UNI drops below 18% or STFlow no longer ranks first, the central SOTA claim is benchmark-curation dependent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 and Appendix B define benchmark variants that differ from the published HEST-1k and STImage-1K4M protocols: (1) STImage-Bench keeps only organs with Pearson correlation > 0.1, and (2) the COAD data is an older version than the one in the HEST-1k release (footnote 2). The headline claim of 'substantially outperforming state-of-the-art baselines' with 'over 18% relative improvement' is computed as an average over these curated subsets. Excluding low-correlation organs is a selection on the evaluation metric; if the excluded organs are ones where spatial-context methods gain little, the average advantage is inflated. Because the paper's internal comparisons are consistent and disclosed, this does not invalidate relative results within the chosen subsets, but it does mean the benchmark-level SOTA claim is not yet established on the official datasets. A full-protocol rerun is required before the result can be read as an unconditional benchmark improvement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces STFlow, a flow-matching-based generative model for predicting whole-slide spatial transcriptomics from H&E-stained histology images. The method models the joint distribution of gene expression across spots via an iterative denoising process, using a local spatial attention transformer with E(2)-invariance achieved through frame averaging. A zero-inflated negative binomial prior is used to capture the sparsity and overdispersion of gene expression counts. The authors evaluate STFlow on HEST-1k and STImage-1K4M, reporting large average improvements over spot-based and slide-based baselines, including pathology foundation models, along with better efficiency and lower memory usage. The paper also includes biomarker prediction experiments, ablations, and hyperparameter studies.","tokens_in":21162,"tokens_out":6549,"duration_ms":54703,"significance":"If the claims are validated, STFlow is a significant contribution: it addresses the memory bottleneck of whole-slide processing, explicitly incorporates spatial cell-cell interactions, and demonstrates that a generative formulation with a tailored prior can improve gene-expression prediction over regression-based baselines. The paper's strengths include a public implementation, experiments across 17 datasets and four biomarkers, multiple ablations, and clear disclosure of some benchmark modifications. However, the headline 'substantially outperforms state-of-the-art' claim is currently weakened by benchmark curation choices and by an internal inconsistency in the 'consistently outperforms' statement. These issues are fixable and do not invalidate the core methodology, but they must be addressed before the benchmark-level result can be accepted at face value.","major_comments":[{"comment":"The STImage benchmark used in this paper is not the official STImage-1K4M benchmark: Appendix B states that only organs where all models achieve Pearson correlation greater than 0.1 are included. This is a selection on the evaluation metric, and it can inflate the average improvement reported in Table 1 and the abstract. Since the abstract and Section 4.1 claim superiority on STImage-1K4M, the authors should either rerun the full protocol on the official STImage-1K4M (including all organs) or explicitly qualify the claim as applying only to the curated STImage-Bench subset. This is load-bearing for the central claim.","section":"Section 4.1 and Appendix B (Tables 7-8)"},{"comment":"The COAD dataset used in the experiments is an older version than the one in the HEST-1k release, as disclosed in footnote 2. Because the paper compares against published HEST-1k results and other baselines, this version mismatch affects the comparability of the COAD row in Table 1. The authors should report the exact version used and, ideally, also provide results on the current HEST-1k COAD version so that readers can verify that the relative improvements are not an artifact of the dataset version.","section":"Footnote 2 and Appendix B"},{"comment":"The statement that STFlow 'consistently outperforms all baselines' is contradicted by Table 1. For example, on HEST IDC, TRIPLEX achieves 0.606 while STFlow achieves 0.587; on STImage Breast, TRIPLEX achieves 0.418 vs STFlow 0.404; and on STImage Mouth, BLEEP achieves 0.191 vs STFlow 0.146. The paper should soften the claim to 'outperforms baselines on average' and discuss the datasets where STFlow does not lead, or provide an analysis of when spatial-context modeling helps and when it does not. As written, the overclaim undermines the paper's central narrative.","section":"Section 4.1 and Table 1"},{"comment":"The training objective in Eq. (2) is a plain mean-squared error between the predicted gene expression and the ground truth, not the standard conditional flow-matching loss on the vector field. For the linear interpolation Y_t = tY + (1-t)Y0, the conditional vector field is Y - Y0; minimizing MSE(Y, fθ) is equivalent to the flow-matching objective only up to a time-dependent weighting (specifically, dividing both sides by (1-t)). The text says fθ approximates the marginal vector field, which is imprecise as written. The authors should state this reparameterization explicitly and justify the unweighted denoiser objective, or use the standard flow-matching loss, so that the 'flow matching' claim is technically accurate.","section":"Section 3.2, Eq. (2)"}],"minor_comments":[{"comment":"The term 'decay coefficient' is not formally defined. Algorithm 2 uses Y_t2 = Y_t1 + (Y_hat - Y_t1)/(1-t1) * (t2 - t1), which is an Euler step with velocity (Y_hat - Y_t1)/(1-t1). Please describe it as such and avoid the vague 'decay coefficient' terminology.","section":"Section 3.2, Algorithm 2"},{"comment":"There is a typo: 'UB2EC' should be 'UBE2C' in Figure 4 and the surrounding text.","section":"Figure 4 and text"},{"comment":"The parameter φ is described as 'the number of failures until stopped,' which is non-standard for the negative binomial parametrization used here. It would be clearer to call φ the inverse dispersion parameter.","section":"Eq. (3)"},{"comment":"The evaluation uses Pearson correlation on log1p-transformed counts while training uses MSE on raw counts. This train/evaluation mismatch should be discussed; in particular, a loss on raw counts may be dominated by highly expressed genes. The authors should justify this choice or include an ablation training on log1p-transformed targets.","section":"Section 4.1"},{"comment":"Footnote 2 notes that the COAD dataset was updated after the paper's release, but it does not specify which version was used to obtain the reported numbers. Please add the exact dataset version and, if possible, the download date.","section":"Table 1 footnote"}],"recommendation":"major_revision","confidential_remarks":"The benchmark curation issue (STImage organ filter and COAD version mismatch) is the most important barrier to accepting the headline claim. The 'consistently outperforms' overclaim in Section 4.1 is easily verifiable from Table 1 and should be corrected. The flow-matching objective point is more of a presentation issue than a fatal flaw, since the unweighted denoiser objective is equivalent in optimum to the weighted flow-matching objective, but the text should be precise. Overall, the paper is promising and the requested revisions are well within scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper, worth a careful read. The new thing is a concrete combination: whole-slide gene-expression prediction as an iterative refinement process, with local spatial attention that is E(2)-invariant through frame averaging, and a ZINB prior. On the benchmarks the authors run, STFlow consistently beats strong baselines, and the ablations support the design choices. The efficiency story is credible: local attention plus few sampling steps avoids the OOM failures of global slide encoders. Code is public, and the writing is transparent about most limitations.\n\nThe main soft spot is benchmark curation. Appendix B states that STImage keeps only organs where all methods exceed Pearson 0.1, and the COAD set is an older version than the one in the published HEST-1k. These are disclosed, but they gate the '18% relative improvement' headline. If the excluded organs are ones where spatial context helps little, the average is inflated. Within the chosen subsets the comparisons look fair; it just isn't yet an unconditional win on the official protocols. A rerun on the standard STImage-1K4M and updated HEST-1k COAD would settle it.\n\nThe flow-matching framing is looser than the text implies. Algorithm 1 trains a denoiser by MSE on the clean target, not a vector-field objective. The iterative sampling is a valid denoiser that works, but it's not flow matching in the Lipman et al. sense. Also, training samples random crops, so 'whole-slide joint distribution' is a bit of an overstatement, though inference is whole-slide. The ZINB hyperparameters are grid-searched on the benchmark; the authors flag this as a limitation, so no foul, just not parameter-free.\n\nI don't think this is cherry-picking or hidden. The stress-test concern is accurate but the authors were upfront. The paper is a solid empirical contribution and a useful architectural template. It deserves proper peer review, and I'd cite it for the method after the benchmark cleanup. Bring it to reading group.","headline":"A solid, well-ablated empirical model for ST prediction, but the headline SOTA gains ride on curated benchmarks and a loose 'flow matching' label; referee it, but ask for a protocol-clean rerun.","tokens_in":21682,"tokens_out":4387,"would_cite":true,"duration_ms":38279,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"STFlow predicts a tissue slide's gene expression from its H&E image by modeling all spots jointly with flow matching, and reports over 18% relative improvement over pathology foundation models on two benchmarks.","keywords":["spatial transcriptomics","flow matching","gene expression prediction","whole-slide histology","cell-cell interaction","E(2)-invariant spatial attention","frame averaging","zero-inflated negative binomial prior"],"falsifier":"Reproduce the comparison on the unmodified benchmarks — the published version of COAD in HEST-1k and all STImage-1K4M organs, including those whose baseline correlation sits below 0.1 — and measure STFlow's average relative improvement over the pathology foundation models; a margin materially below the reported 18% would overturn the headline claim. The cell-interaction claim can be tested directly: in held-out slides, check whether spot pairs receiving high spatial-attention weights coincide with known co-expressed or ligand-receptor gene pairs, and whether the one-step (no flow matching) variant does equally well on that pairing test; if it does, the iterative joint modeling is not the source of the biological signal.","tokens_in":20775,"feed_emoji":"🧬","tokens_out":15546,"duration_ms":103706,"temperature":0.7,"pith_summary":"STFlow aims to make spatial transcriptomics scalable by predicting a tissue slide's gene expression directly from its H&E-stained histology image, replacing a slow, specialized laboratory assay with computation. The paper argues that prior prediction methods have two weaknesses: they predict each spot's genes independently, ignoring that neighboring cells regulate each other's expression, and their whole-slide encoders cannot handle the tens of thousands of spots on a typical slide. STFlow instead models the entire slide's gene expression as one joint distribution learned by flow matching, in which a denoiser iteratively refines an initial expression guess for every spot under the guidance of the image and of local spatial context encoded by an $E(2)$-invariant attention mechanism. On seventeen benchmark datasets spanning two collections, the paper reports that this design outperforms all eight baselines, including an over-18% relative improvement in average Pearson correlation over pathology foundation models, and it gives the best predictions for the four clinical biomarker genes tested.","feed_headline":"Flow matching reads H&E slides into gene maps, beating prior models by 18%","feed_subtitle":"Capturing cell-cell signaling lifts average accuracy 18% over prior pathology models while cutting memory use.","key_machinery":"The argument runs on three coupled components. (1) Flow matching with a gene-specific prior: the network $f_\\theta$ minimizes MSE between ground truth $Y$ and its reconstruction from the noised input $Y_t$, and at inference an initial sample from the zero-inflated negative binomial distribution $Z(\\mu, \\phi, \\pi)$ is refined over $S$ steps, so prediction is a denoising trajectory rather than a single forward pass. (2) E(2)-invariant frame-averaging spatial attention: for each spot, direction vectors to its $k$ nearest neighbors are projected through four PCA-derived frames (two principal components with sign flips), embedded, and averaged, making the representation invariant to rotation, translation, and reflection; the attention weight then uses an MLP over query, key, spatial relation, and pairwise expression difference $(Y_{t,i}-Y_{t,j})$, which couples the gene predictions of neighboring spots. (3) Local neighborhood sparsity: attention is computed only among $k$ nearest neighbors, giving $O(Nkd + Nkd^2)$ complexity instead of $O(N^2)$ global attention, which is what makes the whole-slide formulation practical for up to tens of thousands of spots. The ablation studies identify both flow matching and frame averaging as individually necessary for the reported accuracy.","core_discovery":"The paper's central claim is that whole-slide gene expression should be generated, not regressed: rather than predicting each spot's expression $Y_i$ from its own image patch, STFlow models the joint distribution $p(Y_0, \\dots, Y_N \\mid I_0, \\dots, I_N)$ and samples from it through iterative denoising. The denoiser is trained with flow matching on linear interpolation paths $Y_t = tY + (1-t)Y_0$ between ground truth and a sample from a prior distribution, and the zero-inflated negative binomial distribution $Z(\\mu, \\phi, \\pi)$ is chosen as the prior because real expression counts are sparse and overdispersed. Spatial context enters through frame-averaging attention restricted to each spot's $k$ nearest neighbors: the attention weight between two spots is a learned function of their image features, their relative direction vector encoded in an $E(2)$-invariant way, and the difference in their current expression, which is the mechanism the paper identifies with cell-cell interaction. The key reported finding is that this compound design works: average Pearson correlations of 0.415 on HEST-1k and 0.293 on STImage-1K4M place it ahead of every baseline, and ablations show that removing flow matching, removing frame averaging, or replacing the backbone with EGNN or E2CNN each costs accuracy.","pith_inferences":["Editorial inference: the pairwise expression difference $(Y_{t,i}-Y_{t,j})$ inside the attention weight makes the model's attention maps a candidate readout of cell-cell signaling; one direct test would check whether high-attention spot pairs in held-out slides coincide with known ligand-receptor co-expressed gene pairs, which would turn the biological claim from an aggregate-accuracy argument int","Editorial inference: the $E(2)$-invariance property predicts that outputs are unchanged when the whole slide is rotated, translated, or reflected, and that the model transfers across coordinate conventions without retuning; this is testable with straightforward geometric perturbations and, if confirmed, would ease cross-institution and cross-array deployment.","Editorial inference: the largest relative gains in the reported tables occur on the hardest organs (HCC, READ, CCRCC), where baseline correlations are below 0.3; this suggests joint generative modeling may matter most when spot-level morphology alone carries the least signal, a pattern worth validating prospectively on new organs.","Editorial inference: the paper's stated limitation that the ZINB prior's hyperparameters are grid-searched rather than estimated suggests a concrete next step it does not take — estimating them from the training distribution, for instance with a small amortized model — which could both remove a free parameter and tighten the prior's match to each slide."],"forward_implications":["If the reported gains hold, spatial transcriptomics becomes a computation on a histology scan: full slides with thousands of spots are processed in one pass, with the paper reporting orders-of-magnitude faster inference and lower memory usage than slide-level attention baselines.","The four biomarker genes tested (GATA3, ERBB2, UBE2C, VWF) are all predicted more accurately than by the best baseline or by the foundation model alone, which strengthens the case that histology-only prediction could support clinical biomarker screening.","The improvement is consistent across three pathology foundation models (Ciga, UNI, Gigapath), meaning the method stacks on top of image encoders and should inherit future encoder advances.","The reported step-count analysis shows that two refinement steps already beat one-step prediction and that five steps suffice, so the generative formulation buys accuracy at a small, predictable inference cost."],"supporting_citations":[{"why":"Supplies the flow matching framework whose marginal-vector-field objective STFlow trains with.","marker":"Lipman et al., 2022"},{"why":"Companion stochastic-interpolant formulation of flow matching that motivates the linear interpolation path.","marker":"Albergo & Vanden-Eijnden, 2022"},{"why":"Provides frame averaging, the E(2)-invariance machinery used in the spatial attention encoder.","marker":"Puny et al., 2021"},{"why":"Contributes the HEST-1k benchmark, its patient-stratified split, and the pathology-foundation-model regression setup STFlow is evaluated against.","marker":"Jaume et al., 2024"},{"why":"Contributes the STImage-1K4M collection from which the seven-organ test benchmark is built.","marker":"Chen et al., 2024a"},{"why":"Provides the UNI pathology foundation model that extracts the visual features both STFlow and its strongest baselines consume.","marker":"Chen et al., 2024b"},{"why":"The TRIPLEX slide-based baseline that STFlow outperforms; also evidence for multi-resolution context modeling.","marker":"Chung et al., 2024"},{"why":"The BLEEP contrastive baseline, which like STFlow uses the UNI encoder and is a key comparison point.","marker":"Xie et al., 2023"},{"why":"Precedent for flow matching with five refinement steps, the default S used in STFlow's inference.","marker":"Jing et al., 2024"},{"why":"Supplies the biological rationale that expression of neighboring cells is coupled, motivating joint whole-slide modeling.","marker":"Biancalani et al., 2021"}],"fun_headline_variants":["STFlow: whole-slide gene prediction that models cell-cell signaling","H&E to gene maps, 18% better: STFlow models cell interactions","Memory-efficient flow matching maps H&E to full-slide gene expression","Generative ST: flow matching captures cell-cell context from H&E"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The performance claim rests on the benchmark comparison being fair and representative: the STImage benchmark used here keeps only organs where every method exceeds a Pearson correlation of 0.1, and the COAD dataset in the HEST benchmark differs from the version in the published collection, so gains measured on this curated comparison may not transfer to the unmodified benchmarks.","fun_headline_variants_meta":{"raw":{"variants":["STFlow: whole-slide gene prediction that models cell-cell signaling","H&E to gene maps, 18% better: STFlow models cell interactions","Memory-efficient flow matching maps H&E to full-slide gene expression","Generative ST: flow matching captures cell-cell context from H&E"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001691,"raw_usage":{"total_tokens":6748,"prompt_tokens":1040,"completion_tokens":5708,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":5639}},"tokens_in":656,"tokens_out":5708,"duration_ms":38296,"temperature":1.0,"reasoning_tokens":5639,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:22:36.729677+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the comparison on the unmodified benchmarks — the published version of COAD in HEST-1k and all STImage-1K4M organs, including those whose baseline correlation sits below 0.1 — and measure STFlow's average relative improvement over the pathology foundation models; a margin materially below the reported 18% would overturn the headline claim. The cell-interaction claim can be tested directly: in held-out slides, check whether spot pairs receiving high spatial-attention weights coincide with known co-expressed or ligand-receptor gene pairs, and whether the one-step (no flow matching) variant does equally well on that pairing test; if it does, the iterative joint modeling is not the source of the biological signal.","supporting_citations":[],"review_version":1}