acceptodds
Withdrawn from ICLR 2027

BRIDGE: Bridging Histology and Spatial Gene Expression with Latent Co-Expression Programs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predicting spatial gene expression from hematoxylin and eosin (H&E)-stained histology could extend routinely acquired pathology images toward spatial molecular characterization. Existing methods primarily map local morphology to a selected target-gene panel, yet the same visual appearance can accompany different target-gene levels, and similar target expression can occur in morphologically distinct regions. Paired training spatial transcriptomics profiles also measure non-target genes whose coordinated expression provides complementary molecular context for these cases. To use this training signal, we introduce BRIDGE, a histology-based spatial gene expression prediction framework that learns cohort-adaptive co-expression programs from a disjoint support-gene set and converts them into target-specific visual predictors. The framework comprises three complementary modules: the Local Morphology Encoding Module (LME) extracts target-gene-conditioned morphology features; the Co-expression Program Learning Module (CPL) converts support-gene programs into queries for program-specific morphology aggregation; and the Gene-Adaptive Decoding Module (GAD) combines pooled and program-specific morphology with target-gene features through gene-adaptive experts. Across patient-disjoint evaluations, BRIDGE achieves the best MAE and MSE on all three datasets and the highest PCC on ANDERSSON and LUAD. On ANDERSSON, its PCC is 0.734 versus 0.710 for SpaHGC, while its MSE is 0.399 versus 0.497 for MERGE. Controlled comparisons, cross-cohort transfer, and ablations assess the roles of support-derived supervision, program-guided morphology aggregation, and gene-adaptive decoding.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
computational pathology, histology-based gene expression prediction, multimodal 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_bAApOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BRIDGE, a framework for predicting spatial gene expression from H&E histology images by leveraging non-target support genes during training. The central idea is to learn latent co-expression programs from support-gene expression and use these programs to guide morphology representation learning and target-gene prediction.

The framework consists of three modules: (1) a Local Morphology Encoding Module that extracts morphology features and target-gene representations using pretrained encoders; (2) a Co-expression Program Learning Module that applies non-negative matrix factorization (NMF) to support-gene expression and uses the resulting program representations as queries for morphology aggregation; and (3) a Gene-Adaptive Decoding Module that combines program-specific morphology features and target-gene embeddings through adaptively weighted experts. Experiments on three spatial transcriptomics datasets (ANDERSSON, LYMPH IDC, and LUAD) demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The paper introduces an interesting approach to exploiting non-target gene expression as additional supervision for spatial gene expression prediction. Specifically, BRIDGE derives cohort-adaptive latent co-expression programs using NMF and transforms them into program-specific queries for morphology aggregation. These representations are subsequently incorporated into a gene-adaptive prediction framework. Importantly, support-gene expression is used only during training and is not required during inference. This design provides a potentially useful way to leverage additional molecular information without increasing the inference-time data requirements.

  2. The consistent improvements across multiple evaluation datasets demonstrate the effectiveness of the proposed method. For example, BRIDGE achieves the lowest MAE and MSE on all three evaluated datasets and the highest PCC on ANDERSSON and LUAD.

Critical weaknesses
  1. My main concern is that the paper claims that non-target gene co-expression programs provide complementary molecular information that improves morphology-based prediction. However, most of the evidence demonstrates improved prediction performance, rather than directly validating the learned programs. For example, the original NMF programs may not remain biologically meaningful after optimization. The performance improvements attributable specifically to program structure are relatively modest.

  2. My second concern is the generalization. The experimental evaluation relies on relatively small numbers of independent patients, particularly in LYMPH IDC (four patients) and LUAD (two patients). Although the use of patient-disjoint splits is appropriate, the limited cohort sizes make it difficult to establish the robustness of the reported improvements. Several PCC gains are relatively modest, and comprehensive patient-level paired statistical analyses are not provided. More importantly, the cross-cohort transfer experiment shows only a marginal improvement in PCC (0.226 versus 0.224), despite improvements in MAE and MSE, suggesting that the learned representations may not reliably capture spatial gene expression variation under domain shift.

  3. My last concern is that the submission exceeds the 9-page limit, with four lines extending onto Page 10.

Questions for authors

Q1: How can the authors verify that the learned programs preserve meaningful co-expression information?

Q2: Are the reported improvements statistically reliable at the patient level?

Q3: How much improvement comes specifically from NMF co-expression structure rather than generic additional supervision?

Q4: How should the cross-cohort results be interpreted given the low PCC?

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

In this manuscript, the authors propose BRIDGE, a method that predicts spatial gene expression from H&E histology. Standard H&E-to-ST methods supervise only a selected target-gene panel and ignore the other genes that spatial transcriptomics already measures. The proposed BRIDGE framework uses those non-target 'support' genes as training-only information. Within each training split, it fits NMF to the support-gene expression, which yields K latent co-expression programs.

Critical strengths
  • This manuscript is well-motivated. The proposed approach reuses non-target genes that ST experiments already measure, but benchmarks usually discard, treating them as privileged training-only information. This adds no inference-time requirements, and the design avoids leakage: support and target sets are disjoint by construction, and NMF is fit on training slides only within each split. The idea is simple, broadly applicable to other H&E-to-ST pipelines, and directly testable.

  • All comparison experiments use patient-disjoint splits with shared target panels, which is in a reasonable evaluation protocol.

Critical weaknesses
  • The image encoder in this manuscript seems to confound the comparison with baselines. The proposed BRIDGE method uses frozen UNI v1 and scGPT features, while the main text does not state which encoders the baselines used. The manuscript's own encoder control shows that much of the PCC advantage depends on UNI v1. With ResNet50, BRIDGE's PCC falls to 0.706 on ANDERSSON, below SpaHGC's 0.710, and to 0.530 on LYMPH_IDC, below SpaHGC's 0.574. However, the headline PCC claims cannot be attributed to the proposed method without baselines re-run on the same frozen UNI v1 features.

  • Given that the evaluated datasets comprise only 8, 4, and 2 patients, several splits are essentially limited, e.g., training on just two patients per fold in LYMPH_IDC and estimating LUAD variance from two runs. Furthermore, most performance margins across main benchmarks and ablations fall within one fold-level SD, and the sole paired test is non-significant. Given the shared fold structure, paired fold-level differences with confidence intervals are necessary.

Questions for authors
  • Which image encoder and feature extractor did each baseline use? Can you report the strongest baselines, such as SpaHGC, MERGE, TRIPLEX, and STEM, re-run on the same frozen UNI v1 features and splits? If BRIDGE's PCC advantage disappears in an encoder-matched comparison, the headline claims would need to be restated in terms of error metrics only.

  • Please provide per-fold results and paired fold-level differences, with confidence intervals or paired tests, for the main comparisons and for each ablation and control on ANDERSSON and LYMPH_IDC. Do the conclusions hold under these paired analyses?

Reviewer_a5XTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes BRIDGE, a DL framework for predicting spatial gene expression from H&E-stained histopathology images.

Motivation: The authors observe that local H&E morphology alone can be ambiguous, since morphologically similar tissue regions can exhibit different target-gene expression levels, while morphologically distinct regions can share similar target-gene levels under different contexts.

So to address this, BRIDGE partitions the measured transcriptomic profile during training into a predicted target-gene set and a disjoint support-gene set , using the support genes only during training to construct cohort-adaptive latent co-expression programs via non-negative matrix factorization. So my understanding of the term “cohort-adaptive” are gene clusters learned directly from the training dataset, rather than using a fixed, pre-existing list of biological pathways.

While BRIDGE addresses a meaningful problem, presents an interesting architecture, and reports encouraging results, its relatively broad novelty claim is weakened by close relevant prior work that is neither discussed nor compared.

More importantly, with all due respect, the readability of this paper is extremely bad, and there are 400+ hyphens overused in this draft (e.g., “cohort-adaptive”, “patient-disjoint”, “target-gene-adaptive”, “support-gene-weighted”) without any definition, many of those hyphenated technical phrases are used to describe very simple operations or concepts. Details in the Questions section. So I suspect this paper is a direct LLM generation without enough human review or refinement given the hyphen usage tradition [1,2], even if the AI usage claims “The authors did not use generative AI tools in conducting the research or preparing this manuscript”.

Additionally, the main-text exceeds the 9-page template limit.

Reference:

[1] https://www.mcgill.ca/oss/article/critical-thinking-student-contributors-technology/why-did-llms-steal-our-em-dashes

[2] https://www.seangoedecke.com/em-dashes/

Critical strengths

1. The problem is meaningful

Using additional expression measurements that are already available in paired training data is a reasonable way to improve histology-based prediction. This method requires no additional molecular assay at inference time, so I think the direction is well motivated as an auxiliary supervision for learning.

2. It provides a concrete use of molecular programs

Beyond adding an auxiliary expression loss, it also uses support-derived gene loadings to initialize the queries, and uses those queries to aggregate morphology, and combines program-specific features via gene-dependent decoding.

3. The reported improvements are encouraging

Critical weaknesses

1. (Major) originality

Using disjoint non-target genes as auxiliary supervision for histology-based target-gene prediction already appears in earlier work. For instance:

HE2Gene (Bioinformatics 2024) explicitly separates target-gene prediction from auxiliary non-target-gene prediction. Also, paper Auxiliary Gene Learning: Spatial Gene Expression Estimation by Auxiliary Gene Selection (AAAI 2026), is very close to BRIDGE’s motivation. It explicitly studies genes excluded from the evaluation panel, uses their expression prediction as auxiliary tasks, and then learns which auxiliary genes benefit the primary prediction task through selection and bilevel optimization.

Neither paper appears in the paper’s discussion or references.

2. Inconsistent description of LME

The abstract claims that LME extracts "target-gene-conditioned morphology features." However, Eq. 2 shows that image features are computed independently of target genes, which are encoded separately. Target-gene conditioning occurs only in the final decoding module. The abstract should accurately describe the model.

3. Unconstrained program queries & confounded ablation

Eq. 7 adds an unconstrained learnable residual to each NMF-derived query, making it equivalent in representational capacity to a learnable query initialized from NMF. Thus, preservation of the NMF structure is not guaranteed.

More importantly, Table A compares learnable gene-weighted queries without program activity supervision vs. full BRIDGE with this supervision. This confounds the effects of NMF initialization and auxiliary supervision.

Questions for authors
  1. Here is a list of awkward, unnecessary hyphened terms in the draft and their occurrence. I would kindly request the authors to clarify their meaning.
                       Hyphened Terms                                                           Meaning
train-derivedtraining?
cohort-adaptivelearned?
data-adaptivedata-driven?
support-derivedcomputed from auxiliary genes?
target-gene-conditioned
target-gene-adaptive
support-gene-weighted query construction
training-only support-gene paradigm
training-only molecular context
program-guided morphology aggregation
program-relevant morphology
gene-adaptive decoding
program-expert decoder
program-wise estimates
morphology-derived score estimate
patient-disjointpatient-level cross validation?
module-subtraction ablationsAblation study?
small-cohort reciprocal evaluationCross validation?
post-selection parameter sensitivity analysisHyperparameter sensitivity analysis?
split-wise evaluation protocolcross-validation protocal?
outer-test observationsThe paper does not use nested cross-validation so outer-test is misleading.
extracellular-matrix

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

Related papers

Discussion (0)

Sign in to comment.