acceptodds
Withdrawn from ICLR 2027

GeneRoute: From Gene Knowledge to Histological Evidence for Spatial Gene Expression Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predicting spatial gene expression from routine H&E histology could make molecular tissue maps widely accessible, but fixed output panels restrict many models to genes observed during training. We introduce GeneRoute, a gene-conditioned framework that uses molecular knowledge to guide multiscale image interpretation. Sequence and functional priors define shared gene representations, while cancer-specific dependencies and molecular programs adapt them to tissue context. These queries retrieve histological features at three scales, and a shared scalar decoder predicts expression without gene-specific output heads. Across five cancers, GeneRoute leads all five macro-averaged metrics on cancer-related genes, improving gene-wise Pearson correlation by 25% over the strongest comparator. For 200 genes excluded from expression supervision and model selection, it also achieves the highest macro-averaged gene-wise Pearson and Spearman correlations. On the six-gene immune-receptor subset, GeneRoute achieves positive mean gene-wise correlations across all five cancers, with predicted IGHG1 enrichment partly overlapping outlined immune infiltrates in a colorectal section.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, Histology, Spatial gene expression prediction, Gene representation, Knowledge-guided learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AerjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces GeneRoute, a framework that predicts spatial gene expression from H&E images conditioned on specific genes. Gene queries are constructed by fusing embeddings from frozen ESM2 (protein sequences), RNA-FM (RNA sequences), and GenePT (functional text), based on availability, and refining them via the STRING graph; these are further adapted to specific cancer types using DepMap dependency graphs and six program banks (Cancer Card, Hallmark, GO, Reactome, JUMP ORF/CRISPR). The adapted queries perform cross-attention with frozen Virchow2 image features across three spatial scales. Following fusion by a router, a shared scalar decoder outputs a value for each gene-location pair, enabling the querying of genes not included in the expression supervision. During training, genes are sampled in blocks, with 20% of the sampling budget allocated to priority genes for each cancer type. Experiments cover five cancer types, utilizing 108, 23, and 33 slides for training, validation, and testing, respectively. The framework is compared against 12 existing methods across four supervision panels (high variance, low variance, HVG, and cancer-related) and evaluated on 200 genes excluded from both expression supervision and model selection (U200) including six immune receptor genes, alongside spatial visualization of EPCAM and component ablation studies.

Critical strengths

The problem formulation is meaningful. By constructing gene queries using molecular priors and employing a shared decoder, the model can predict genes not included in the expression supervision, thereby overcoming the limitations of a fixed output panel. This represents a substantial capability advantage over most baselines and constitutes the technical contribution likely to support the paper's acceptance.

Critical weaknesses
  1. The results reported in the paper cannot be attributed solely to the proposed method.

The cancer-related panel corresponds to the same set of cancer-type priorities used for the 20% prioritized sampling during training; in other words, the panel yielding the most prominent results consists of the very genes that were heavily weighted during the training process. When prioritized sampling was removed (Table 3), the gene PCC for this panel dropped from 0.1020 to 0.0472, a decline of more than 50% and the largest effect among all ablated components. Furthermore, GeneRoute did not outperform STFlow on two general-purpose panels that were independent of these priorities (Table S1). On high-variance panels (selected based on the variance of log-normalized expression), STFlow achieved superior gene PCC, gene Spearman, and RMSE scores (0.1196 vs. 0.1183, 0.1109 vs. 0.1068, and 0.7835 vs. 0.8319, respectively). On HVG panels (selected using Seurat v3), STFlow also performed better in terms of gene PCC (0.0887 vs. 0.0850) and RMSE (0.6051 vs. 0.6450), with GeneRoute leading only in gene Spearman (0.0938 vs. 0.0892). STFlow also outperformed the other methods on spot-wise metrics for both panels.

The backbones are not comparable. GeneRoute employs a frozen Virchow2 model (ViT-H, approximately 630 million parameters), reading over a hundred fields of view across three scales for each spot; in contrast, the baselines "retain their respective architectures and pretrained resources" (Appendix L). For instance, ST-Net continues to use an ImageNet-pretrained DenseNet-121. The paper does not provide a comparison using a "same encoder + simple regression head" setup.

Some baselines appear to have failed during training. The gene PCC values for Hist2ST and THItoGene are merely 0.0004 and 0.0002 (Table 1), yet Appendix L fails to provide detailed training configurations for any of the baselines. Figure S1 corroborates this. Using method-specific color scales, the prediction ranges for Hist2ST and THItoGene on the colorectal tissue section are merely 0.48372 to 0.48405 and 0.87311 to 0.87357, respectively, values that are virtually constant yet visually rendered as patterns with spatial structure; meanwhile, BrSTNet yields PCC values of −0.31 and −0.42 across the two sections. Furthermore, GeneRoute is absent from Figure S1 and can only be cross-referenced with Figure 2; however, as both figures employ method-specific color scales, a direct comparison is impossible.

Consequently, the "+25%" figure cited in the abstract cannot be disentangled from variables such as training allocation, encoder differences, and baseline status.

  1. There is a lack of evidence supporting the two core selling points ("knowledge-guided" design and zero-shot capability), and the descriptions of the methodology and reproducibility information are insufficient.

The impact of the priors is minimal. In the fixed-weight analysis, disabling the "Cancer Card" during inference results in a change in gene PCC of no more than 4×10⁻⁵ (Appendix N); in the retraining ablation study, removing all priors causes a performance drop of only 14.4%, a smaller reduction than replacing sequence embeddings with random vectors (18.8%). The prior ablation in Table 3 involved training for only 20 epochs, and the results are not additive. Removing DepMap alone (−14.2%) yields nearly the same result as removing all priors (−14.4%), making it difficult to assess the individual contribution of each prior.

The zero-shot results are very weak. The macro-average gene PCC for the U200 genes is only 0.027, less than 0.01 higher than the runner-up method; furthermore, the RMSE is inferior to that of TissueFormer. Some of the six immune receptor genes likely have highly homologous paralogs in the training set (e.g., IGHG2/IGHG4, IGLC2/IGLC3), suggesting that predictions may simply be leveraging these genes.

Figure 1 is inconsistent with the methodology section. Sub-figures A to D are neither cited nor explained in the figure caption or the main text; the module labeled "Gene-Conditioned Scale Routing" in sub-figure C is drawn only as a summation node, while the actual gene-conditioned cross-attention, the router weights (Equation 4), and the image-reading query q'_{c,g} described in the text are not depicted; the terms u_shared and u_cancer in the figure are defined only in the appendix; and the "Cancer Card" in the figure includes "Morphology," a feature never mentioned in the main text. Readers cannot understand the method based on Figure 1.

The results are difficult to reproduce. The main results rely on a single "selected run"; the ablation study involved training for only 20 stage-1 epochs and was evaluated on just 20 test sections; the "Cancer Card" (13 programs per cancer type derived from a review of the past decade's literature) lacks a documented construction process and potentially overlaps with the reporting panels. Although the appendix spans approximately 24 pages, it omits critical information needed to evaluate the claims, such as detailed baseline configurations, Cancer Card details, sources of priority weights, and measured computational costs. No code has been provided, nor is there any indication that it will be released; furthermore, the supplementary files mentioned in the text (e.g., u200_genes.csv, supervised_panel_membership.csv, gene_panel_audit.json) have not been uploaded. While ICLR does not mandate code submission, this method relies heavily on a vast array of self-curated external resources (including pre-processed ESM2/RNA-FM/GenePT embeddings, STRING and DepMap graphs, six program banks, and the Cancer Card), alongside the reproduction of 12 baselines. Without the aforementioned information or the code itself, reproducing the main results is impossible.

Consequently, the claims of knowledge-guided approaches and out-of-panel prediction, highlighted in the title and abstract, currently lack reliable substantiation. I cannot draw reliable conclusions from the current version of the paper.

Questions for authors

Table 3 shows that the performance of the cancer-related panel drops by more than half when priority sampling is removed. Please report the results for GeneRoute, trained using the full training budget without priority sampling, across the four panels, or ensure that all baselines use the same loss weights for the set of priority genes. Additionally, please clarify the overlap between the priority weights, the cancer-related panel, and the Cancer Card genes, and provide the full content and construction process for the Cancer Card.

Please report results for simple baselines (e.g., Virchow2 + ridge/MLP) and versions using concatenated multi-scale features, all based on the same Virchow2 features; also, provide the training configurations for each baseline. Since the gene PCC values for Hist2ST and THItoGene are close to zero, please clarify whether training failed. I would also like to know if the authors' modules work with other foundation model (FM) backbones.

Please explain which gene-specific inputs are retained in the "w/o all priors" setting and report the gene PCC for a gene-agnostic baseline (one that outputs the same spatial pattern for all genes) to quantify the actual gains derived from the priors.

Appendix D merely states that STRING data is stored as "normalized top-32 neighbor lists." Please specify which STRING score is used (combined score vs. specific evidence channels), the confidence threshold, and the normalization method. If the combined score is used (where co-expression and text-mining channels might allow unseen genes to adopt the expression patterns of supervised genes), please report U200 results broken down by channel (e.g., using only experimental and database evidence). Also, please report the values of γSTRING, γDepMap, λm, and λctx after training.

Please report the embedding similarity between each gene in the U200 set and its most similar training-set gene, as well as the results for a baseline that simply uses the prediction of the most similar training gene. Do the zero-shot conclusions still hold after excluding genes that have highly similar paralogs in the training set?

If the authors address these questions thoroughly in the rebuttal, I will consider raising my score to 2.

Reviewer_yDg4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeneRoute for spot-level gene expression prediction from H&E images, with a focus on genes excluded from expression supervision. It combines molecular priors and cancer-specific context to guide multiscale image readout. Across five cancer contexts, the model outperforms the compared methods on the cancer-related gene panel and obtains the highest average gene-wise correlations on 200 genes excluded from expression training and model selection.

Critical strengths
  1. The design brings gene-level information and cancer context into the same prediction framework. Using these representations to guide image readout is a reasonable way to connect molecular knowledge with histology.
  2. The held-out gene evaluation is carefully defined. The 200 genes are excluded from both training stages and model selection, and the data splits are disjoint at the patient and specimen levels.
  3. The model improves all five reported metrics on the cancer-related gene panel across the five-cancer macro average. The paper also reports broader gene panels where other methods perform better, making the scope of the gains clear.
Critical weaknesses
  1. Table 3 reports component ablations only on supervised gene panels, although prediction beyond expression supervision is a main motivation. These results do not show which molecular priors or model components help predict the 200 held-out genes. Key ablations on U200 would provide more direct support for this contribution.
  2. The R0-only comparison removes the wider image context as well as scale selection. It therefore does not show whether gene-dependent routing works better than simply averaging or concatenating the same three sets of image features. Gene information also conditions the decoder, so the experiments do not establish how much is gained by using it to guide image readout.
  3. On U200, the mean gene-wise PCC is 0.0273 and Spearman correlation is 0.0515. These scores outperform the compared methods but remain weak in absolute terms. Per-gene results would help show how many unseen genes have useful spatial predictions and whether the gains extend beyond the six immune genes highlighted in the paper.
Questions for authors
  1. How do the main prior and component ablations perform on U200, keeping these genes excluded from training and model selection?
  2. With the same three image ranges, how does learned routing compare with simple averaging or concatenation? How much does gene-guided image readout add when gene conditioning in the decoder is retained?
  3. Could you report the distribution of per-gene correlations on U200, including the proportion near zero or negative, and break down performance by expression level and available prior sources?
Reviewer_NhMvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces GeneRoute, a gene-conditioned framework for predicting spatial gene expression from H&E histology. It combines sequence embeddings, functional-text embeddings, and gene relationships into shared gene representations, then adapts the queries using cancer-specific dependency graphs and molecular program banks. The queries retrieve image features at three spatial scales, which are combined through gene- and location-dependent routing before shared scalar decoding.

Critical strengths

1. A coherent and practically meaningful framework for gene-conditioned prediction.

GeneRoute integrates shared gene identity, cancer-context adaptation, and multiscale image readout into a unified prediction pathway, using a shared decoder to support queries for genes excluded from expression supervision. The handling of different prior sources and missing information is also explicitly described. Although the underlying ideas have precedents, this integration is reasonable and provides a concrete implementation for extending histology-based prediction through external molecular knowledge. This is the principal reason I consider the work to have research value.

2. Consistent predictive gains on cancer-related genes.

GeneRoute leads all five macro-averaged metrics on the cancer-related panel, improving gene-wise PCC from 0.0999 for STFlow to 0.1249. The evaluation covers five cancer contexts; the authors state that the training, validation, and test partitions are disjoint at both patient and specimen levels, and explicitly report the reference-information access of different baselines. These results directly support the effectiveness of the complete framework and constitute the strongest empirical consideration in favor of acceptance, although the conclusion should remain limited to the corresponding panel and evaluation conditions.

Critical weaknesses

1. The independent value of the added mechanisms over prior approaches is not sufficiently established.

Gene inputs, shared prediction, and prediction of unseen genes have precedents in GeneQuery and SGN [1, 2], while DeepSpot-M already uses gene queries to read image features [3]. The contributions most in need of validation are therefore cancer-context adaptation and gene-dependent scale routing. However, the R0-only ablation also removes additional visual information, so it directly supports the usefulness of broader tissue context rather than the routing mechanism specifically. Likewise, the decrease after removing DepMap does not establish that correctly matched cancer context is superior to generic context (Table 3 and Appendix N). Without alternative-design controls that hold information inputs and training conditions fixed, it is difficult to determine whether the added complexity constitutes a sufficiently clear methodological advance. This is my primary concern affecting the recommendation.

2. The strongest performance gains have not been separated from the effect of priority sampling.

In Table 3, removing priority sampling produces the largest performance decrease: mean gene-PCC across the four panels falls from 0.0754 to 0.0472, and the cancer-related-panel value falls from 0.1020 to 0.0472. The reporting panel itself is also selected using cancer-specific priority (Table 3 and Appendix J). This is a legitimate task-focused strategy, but the experiments do not separate the contributions of training allocation and the added biological mechanisms. A comparison crossing priority versus uniform sampling with full versus simplified priors would help determine whether the biological design provides consistent gains under the same sampling conditions. Furthermore, Table S1 shows that the advantage does not extend consistently to broader gene panels; for example, STFlow leads all five metrics on the high-variance panel (Table S1). Consequently, the complete model's advantage does not yet establish stable, generalizable gains from the core proposed design, substantially weakening the empirical support for its methodological contribution.

[1] Xiong, Ying; Liu, Linjing; Cui, Yufei; Wu, Shangyu; Liu, Xue; Chan, Antoni B.; and Xue, Chun Jason. GeneQuery: A General QA-based Framework for Spatial Gene Expression Predictions from Histology Images. arXiv:2411.18391, 2024.

[2] Yang, Yan; Hossain, Md Zakir; Li, Xuesong; Rahman, Shafin; and Stone, Eric. Spatial Transcriptomics Analysis of Zero-shot Gene Expression Prediction. In Medical Image Computing and Computer Assisted Intervention — MICCAI 2024

[3] Nonchev, Kalin; Dawo, Sebastian; Silina, Karina; Koelzer, Viktor H.; and Rätsch, Gunnar. DeepSpot-M: A Multimodal Foundation Model for Transcriptome-wide Virtual Spatial Transcriptomics from Histology. medRxiv preprint, 2026.

Questions for authors
  1. Controls for the key mechanisms: With identical three-scale image inputs, does the proposed router outperform simple fusion? Do correctly matched cancer resources outperform shared or mismatched resources? Can full and simplified priors be compared under the same sampling strategy to separate their contributions?

  2. Held-out-gene prediction: Macro-averaged gene-PCC on U200 is 0.0273, and RMSE is higher than that of TissueFormer (Table 2). Which prior sources directly improve spatial prediction for these genes? Please report results by the predefined prior-availability strata and provide per-gene distributions to show how broadly the gains are distributed. Are the results consistent on an independently selected held-out set?

  3. Result stability: The reported standard deviations describe variation across cancers or sections (Section 4 and Appendix L.1). Are the principal improvements robust when assessed using paired confidence intervals that account for patient clustering and across different training runs? Do the main ablation conclusions hold under the final two-stage training procedure?

  4. Effective evaluation subsets: Different methods may exclude different genes because their correlations are undefined (Appendix L.1). How many genes are excluded for each method, and do the principal rankings remain unchanged on a common valid subset?

  5. Reproducibility: Can the exact Cancer Cards, priority-sampling weights, program-selection lists, and patient/specimen split identifiers be provided to enable reproduction of these settings, which materially affect the results?

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

Related papers

Discussion (0)

Sign in to comment.