- 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
- 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.
- 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.