acceptodds
Withdrawn from ICLR 2027

RepGene: Auditing Condition-Specific Fusion Biases for Multi-View Gene Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

This study evaluates how multi-view gene representations should be compared under a matched contract. Each of 18,460 genes is described by five aligned views: DNA sequence-derived features (D), RNA sequence-derived features (R), protein language-model embeddings (P), curated text knowledge (K), and single-cell expression (X). The contract fixes the gene universe, the 256-D output, label-free outer-training-only fitting, one random-forest probe, and task-equal Macro-F1 over nine registered binary tasks, and varies only the fusion operator across four protocol-defined input states: Core (DNA+RNA+protein), Core+K, Core+X, and Core+K+X. Under these states, the top-scoring operator changes with the available view combination and all task-bootstrap intervals overlap, establishing a condition-dependent rather than universal ranking pattern. An independently configured external RF track provides a separate check of the numerical range, while a gene-to-cell boundary diagnostic shows that gene-level differences attenuate toward a dimension-matched random control under expression-weighted aggregation, as predicted by a dilution identity. Historical log-derived multiplicative-interaction screening is retained only as magnitude-only supplementary evidence because fold-level records are unavailable. The contribution is therefore the matched contract and its audit trail: an evidence-graded protocol for turning fusion claims into comparable, condition-bounded measurements.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
gene representations, multi-view representation learning, fusion operators, label-free 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_ibq7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The main experiments cover 15 gene-level tasks, with an additional analysis of attention-based methods and external random forest (RF) configurations on nine binary classification tasks. The paper also examines sensitivity to probes, evaluation metrics, decision thresholds, and training budgets. Expression-weighted aggregation is used to assess how performance declines when gene representations are transferred to cell-level tasks.

Critical strengths

Equation (5) on PDF page 4 and the description on page 17 require all preprocessing and representation learning to use only the outer-training data, with refitting for every task–condition–operator–seed–fold combination. This helps prevent leakage from test-set feature statistics.

The authors acknowledge that all intervals overlap, attention shows no consistent advantage, method rankings depend on the metric, probe, and threshold, and gene-level gains do not necessarily transfer to cell-level tasks. The claim–evidence ledger in Appendix 15 and the limitations in Appendix 21 explicitly identify conclusions that the current results do not support.

Critical weaknesses

The caption of Table 7 on PDF page 18 states that “All operators share the view adapters,” whereas the text on the same page says that the three baselines “never touch the trained adapters.” These descriptions need to be reconciled.

Equation (7) on PDF page 18 uses an undefined function F. The text describes direct concatenation after budget allocation, but Table 8 lists “concatenation + projection.” Please clarify whether F is the identity mapping or an additional projection.

The text in Figures 4–14 on PDF pages 38–41 is generally too small to read comfortably. Figure 14 also presents pancreas and myeloid results, although its caption states that documentation for these datasets is outside the scope of the current report.

Questions for authors

Could the authors use the existing task-, fold-, and seed-level records to report paired difference intervals for the top two methods under each condition, together with the frequency with which each method ranks first? If differences smaller than the variability across seeds and folds are treated as ties, how many tasks still show a winner switch?

For tasks with only about 84 outer-training samples, how does PCA add produce 256 principal components? How does PCA then concat handle a requested dimension that exceeds the sample rank? What exactly does F in Eq. (7) denote?

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

RepGene benchmarks methods for combining five types of gene representations: DNA, RNA, protein, text knowledge, and single-cell expression. It fixes the output dimension, data splits, and downstream evaluation to compare fusion methods across four input combinations. The main finding is that method rankings depend on the available views and evaluation choices, while gene-level benefits do not necessarily transfer to cell-level tasks.

Critical strengths

The controlled evaluation protocol addresses a practical problem: fusion methods are often compared under inconsistent settings.

Critical weaknesses

The contribution is primarily descriptive. No new architecture is claimed, and the study does not yet provide a validated rule for choosing fusion methods. The aggregation bound also relies on uncorrelated residuals, whose relevance to real gene embeddings needs empirical verification.

Questions for authors

Do the findings persist under homology-aware splits and across the remaining eligible tasks?

Can shared-adapter baselines and classical multiview methods clarify which differences arise specifically from fusion?

Can the cross-gene residual correlations be measured to test whether the proposed dilution bound explains the observed cell-level results?

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

The paper introduces RepGene, a controlled evaluation protocol, for experimentally studying multi-view gene representation. The main question is how to combine gene representations into a single representation, and whether the best way to combine them depends on the available views (e.g., missing views). In addition, they study the transfer to cells. Following the protocol, they compare several fusion strategies with a random forest probe across several tasks. The experimental results show that no single method stands out: the differences are small, and the intervals overlap.

Critical strengths
  • The main finding of the paper is that the best fusion operator depends on the available views. Nevertheless, it is a weak finding.
  • The controlled protocol is also a good point in this application domain to study correctly multi-view gene representation.
Critical weaknesses
  • The main weakness and major concern is that the contribution is unclear, and the writing and formalization of the paper make it very difficult to read and understand. For instance, the notion of a contract is not clear. The authors just proposed a controlled evaluation protocol.
  • The technical and methodological novelty is very low.
  • The benchmark is small.
Questions for authors

The authors should substantially improve the paper's writing by stating its objectives and motivations more clearly and by removing unnecessary formalism and protocol-specific vocabulary (e.g., “contract”, “register”, evidence grades, unvalidated or out-of-scope derivations), which currently obscure rather than clarify the main contribution.

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

Related papers

Discussion (0)

Sign in to comment.