acceptodds
Withdrawn from ICLR 2027

GENEBRIDGE: HOW TO BRIDGE GENE SPACES FOR TRANSDUCTIVE ZERO-SHOT GRN INFERENCE

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Gene regulatory network (GRN) inference faces a fundamental precision–generality trade-off: supervised methods achieve high accuracy on labeled cell types but cannot generalize to unseen ones, whereas unsupervised and foundation model-based approaches apply broadly at the cost of predictive accuracy. We propose GeneBridge, a zero-shot framework that transfers regulatory knowledge across cell types through a unified semantic gene graph. The central challenge is gene space misalignment: different cell types express different gene subsets, making expression-based representations non-transferable. GeneBridge addresses this by constructing a transferable semantic graph from foundation model embeddings, where hybrid node representations—contextual for expressed genes, frozen pretrained for absent ones—allow all genes to coexist in a shared semantic space. Experiments on seven BEELINE datasets under strict leave-one-out evaluation show that GeneBridge consistently outperforms existing methods by up to 3.1×. Ablation studies further demonstrate that semantic gene selection is foundational: replacing it with standard HVG filtering reduces Top-100 precision to near zero. Together, these results demonstrate that semantic graph learning provides a principled mechanism for transferring regulatory knowledge across heterogeneous biological contexts.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Gene regulatory network, Zero-shot learning, Semantic graph, Gene embedding, Knowledge transfer, Foundation model

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bSgmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies supervised gene regulatory network inference framework, specifically it proposes an approach called GeneBridge that conducts zero-shot unseen cell line GRN prediction. GeneBridge constructs a unified semantic gene graph with contextual gene representations, and trains a graph transformer on soft GRN labels across different cell types. Moreover, this paper proposes a new gene selection process relying on co-expression and attention from the fine-tuned GeneCompass. GeneBridge is evaluated on well-studied BEELINE datasets with leave-one-cell-type-out evaluation.

Critical strengths

This paper studies an important research question, GRN inference, in biology with potential applications in drug discovery. This paper is well written and the proposed method is tested on well-estabilished datasets for GRN inference BEELINE.

Critical weaknesses

My biggest concern is the central assumption of this work that supervised GRN inference method can be transferable to unseen cell types. In a recent study on BEELINE datasets [1], it was shown that supervised GRN inference methods mainly learn TF/TG's degree imbalance, and a simple logistic regression with one-hot gene embedding (without expression measurement) can match the performance of sophisticated supervised GRN methods.

Moreover, I'm not convinced by using co-expression and attention as part of gene selection. In practice, gene selection is driven by drug-discovery applications and HVGs are well motivated because they represent more "important" genes in the cell type. It makes more sense to make co-expression and attention as post-hoc filters after GRN inference rather than gene selection.

[1] InfoSEM: A Deep Generative Model with Informative Priors for Gene Regulatory Network Inference, ICML 2025

Questions for authors
  1. Could the author show the performance of simply predicting the multi-CT joint soft-label without any training?

  2. Could the author show the performance on unseen genes (see [1]) of unseen cell types, rather than seen genes of unseen cell types, which are known to be easily predicted by gene degree? Moreover, other supervised GRN inference methods should be included in the baseline (e.g., directly apply a GENELink trained on hESC on hHep).

[1] InfoSEM: A Deep Generative Model with Informative Priors for Gene Regulatory Network Inference, ICML 2025

Reviewer_UB1qOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

GeneBridge proposes a transductive zero-shot framework for gene regulatory network inference across cell types. It combines correlation- and attention-based gene selection with a shared semantic graph built from GeneCompass representations. An edge-aware Graph Transformer learns regulatory relationships from annotated source cell types and predicts TF–target edges in a held-out type. Target expression data participate in preprocessing and self-supervised adaptation, but target regulatory labels are withheld.

The authors evaluate seven human and mouse BEELINE datasets using leave-one-cell-type-out experiments, including a mixed-species setting. They report strong top-ranked precision for human and hematopoietic datasets, with weaker results for mouse embryonic stem cells and mixed cross-species transfer. Ablations examine gene selection, graph architecture, soft labels, and edge weighting. These support the proposed pipeline, although comparisons with baselines use different candidate gene spaces, making it difficult to isolate improvements in regulatory prediction from those due to gene selection.

Critical strengths

The paper addresses a practically important setting: transferring regulatory knowledge to a cell type without its regulatory annotations. Its shared graph combines pretrained gene representations with cell-type-specific expression context, providing a concrete mechanism for transfer across differing gene sets. Leave-one-cell-type-out experiments show strong top-ranked precision on several human and hematopoietic datasets, with main results averaged over ten seeds.

Critical weaknesses

The main baseline comparison does not establish that GeneBridge is a better regulatory predictor because the methods are evaluated on different candidate gene and edge spaces. GeneBridge uses its proposed selection procedure, whereas most baselines use TF+500 HVGs. This changes both the prediction problem and the prevalence of reference positives, directly affecting AUPRC and top-ranked precision. The dramatic collapse under HVG selection therefore does not isolate the value of semantic graph learning. Matched-candidate comparisons, including baselines supplied with GeneBridge’s selected genes, are essential before the reported superiority can be substantiated.

The evaluation does not adequately distinguish context-specific regulatory inference from recovery of regulatory edges already annotated in source cell types. This is particularly consequential because the strongest mouse results involve HSC subtypes with highly similar reference networks, and the training objective explicitly promotes edges annotated in other source types. The merged-HSC experiment addresses subtype redundancy but does not establish performance on target edges absent from all source networks. A source-edge transfer baseline and separate evaluations of shared versus target-specific edges are needed to demonstrate that the model learns more than reusable gene-pair associations. Without these controls, the central generalization claim remains insufficiently supported.

Questions for authors

The issue addressed in this paper is indeed very important; however, the current results fall far short of the standard required for publication. Specifically,

  1. Is Figure 1 incomplete or incorrectly exported? The three colored header bars are blank, the panel labels referenced in the caption are missing, and the central architecture contains unexplained empty regions.

  2. How do the baselines perform when supplied with GeneBridge’s selected genes and evaluated on exactly the same candidate edges? The current comparison changes both gene selection and prediction methodology, making their respective contributions difficult to distinguish.

  3. What proportion of target reference edges already appears in the source networks? Please compare against a simple baseline that ranks edges by their occurrence across source networks to establish how much improvement requires the proposed model.

  4. How accurately does GeneBridge predict target-positive edges absent from every source reference network? Separating these edges from shared positives would clarify whether the model discovers context-specific relationships or mainly recovers previously annotated gene pairs.

  5. Which data determine early stopping and the selection of the soft-label weight, KEGG fusion weight, and negative-sampling ratio? Were held-out target regulatory labels accessed at any point during these decisions?

  6. Why do several appendix experiments use sampled negatives for AUROC/AUPRC while the main results use the full candidate space? Do the preferred hyperparameters and reported conclusions remain consistent when all configurations are evaluated over the full space?

  7. What fraction of reference regulatory edges is excluded by semantic gene selection? Please report coverage alongside precision, since strong performance within a selected subset may conceal the exclusion of substantial portions of the target network.

  8. What evidence shows that frozen input embeddings and contextualized output embeddings occupy a compatible representation space? Sharing dimensionality does not establish that these representations can be substituted without introducing systematic differences between expressed and absent genes.

  9. Does the Graph Transformer outperform a simple MLP using the same gene-pair features, source supervision, and candidate space? This comparison would establish whether graph message passing contributes beyond the information already available in pretrained embeddings and source annotations.

Reviewer_X88rOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies GRN inference across cell types (CTs) when regulatory labels exist only for source CTs. GeneBridge has three steps: it selects candidate target genes for each TF, first by co-expression and then by fine-tuned GeneCompass attention; it builds a unified graph over all CTs, with foundation-model embeddings as node features and embedding similarity (optionally combined with KEGG) as edge weights; it trains an edge-aware Graph Transformer with soft labels from other source CTs, then scores the held-out CT. In leave-one-CT-out experiments on BEELINE datasets, it reports high precision among top-ranked edges, especially on hESC/hHep and the mHSC subtypes.

Critical strengths
  • The problem is relevant: transferring regulatory knowledge to unannotated cell types. The transductive protocol is stated explicitly.
  • Main results use multiple seeds, evaluate on the full candidate space and report enrichment over a random baseline. Code is released.
  • The authors are open about protocol differences across methods and about the weak mESC case.
Critical weaknesses
  1. Performance tracks label overlap, so transfer is not separated from label reuse.** In the released GT files, almost all GT edges of each held-out mHSC subtype also appear in the training CTs' GT; the overlap is moderate for hESC/hHep and mDC; it is very small for mESC, the one target where GeneBridge is near chance. Node features are largely gene-identity embeddings shared across CTs. Strong mHSC results therefore mostly reflect recovering pairs that were labeled positive in training. The merged-HSC experiment lowers the overlap yet keeps high precision. That is consistent with a hub-TF prior, since a few TFs (e.g., SPI1, MYC) have very large target sets. Please report the overlap for each setting, stratify metrics by whether the edge was seen or unseen in source GT, add trivial baselines that rank edges by source-GT membership and by the TF's out-degree in source GT, and report how many distinct TFs make up the top-ranked predictions, plus per-TF metrics.

  2. The main comparison is not like-for-like, and the gain seems to come from restricting the candidate space. Baselines are evaluated on TF+HVG genes, while GeneBridge is evaluated on its own selected genes (Table 3). When GeneBridge runs on the baselines' HVG space (Table 4), its top-ranked precision collapses to near zero, below simple PCC and MI. The selection step already ranks genes by co-expression and attention, so a selection-only scorer inside GeneBridge's space may explain much of the gain. Please evaluate all methods on the same candidate spaces and add this baseline. The claim of consistently outperforming existing methods within and across species is not supported; no baselines are reported for the cross-species setting; on mESC, several baselines do better.

  3. Baselines for the actual setting (supervised cross-CT transfer) are missing. Most of the reported baselines are unsupervised methods run within a single CT. Missing:

    • UGRN (Qi et al., 2026), which the paper itself says uses the same datasets;
    • supervised link predictors (GENELink, GNNLink, GCLink) trained on source CTs and applied to the target through shared gene IDs;
    • Meta-TGLink;
    • an FM-embedding + MLP edge classifier with no graph, to show what the graph and Graph Transformer contribute.
  4. Hyperparameters are tuned on held-out labels, and statistical support for the ablations is weak. The soft-label weight, KEGG weight, negative ratio and encoder are all chosen from sweeps scored on the held-out CTs (Appendix D), including species-specific settings. With only two human CTs this is optimistic, and a nested protocol or a separate validation CT is needed. The ablations use a single seed, yet results vary a lot across seeds and settings (mESC especially), so most ablation differences are not shown to be significant.

  5. Label definition and evaluation. (a) The paper describes the reference network as ChIP-seq, but the README and GT files also include non-specific ChIP-seq and STRING, and a sizeable share of mESC GT "regulators" are not TFs (e.g., GABBR1). STRING, KEGG and GeneCompass embeddings (pretrained with prior GRN and co-expression knowledge) carry overlapping information with these labels, so the "semantic" features and edge weights may partly encode the targets; (b) Labels are matched as unordered pairs in both training and evaluation. The task is therefore effectively undirected, and TF–TF pairs can be counted twice in the top-K. Please add a directed evaluation on cell-type-specific ChIP-seq only.

Questions for authors
  1. Please confirm the mDC validation issue and report corrected results.
  2. LapPE is computed on the unified graph for training, but recomputed at inference on the separately built held-out subgraph, with no sign-flip augmentation. Are these encodings meaningful across graphs? What happens if LapPE is removed?
  3. Each CT contributes only a fixed small subset of its positive pairs, and negatives are arbitrary node pairs rather than TF-anchored pairs. Why, and how sensitive are the results to these choices?
  4. Why does HVG selection keep a reasonable AUROC but give near-zero top-ranked precision? What do the top predictions look like in that case?

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

Related papers

Discussion (0)

Sign in to comment.