acceptodds
Under review as a conference paper at ICLR 2027

scDeepFusion: Auditable Gene-Module Representation Learning with Typed-Expert Routing for Single-Cell RNA-seq Integration

Abstract

Single-cell RNA sequencing presents a challenging representation learning problem: extreme sparsity, batch effects, and limited interpretability hinder accurate characterization of cellular states, particularly for rare or transcriptionally similar subtypes. We present scDeepFusion, a gene-set-level representation learning framework that represents cells by pathway-informed features rather than isolated gene measurements. Three components define the framework: (1) a module-activity representation that aggregates gene fluctuations over curated gene sets, suppressing dropout noise while preserving pathway-level biology; (2) a typed-expert routing architecture in which a per-cell sparse router assigns cells to knowledge-source experts (curated pathways, perturbation signatures, and a gene-line expert), yielding embeddings whose knowledge provenance is directly auditable from the exported routing weights; and (3) a teacher-aligned training scheme in which a coarse alignment loss anchors the representation to a reference embedding while auxiliary reconstruction and denoising objectives act as regularizers. Across five datasets evaluated under the standard scIB protocol with multi-seed controls on the flagship dataset and pathway-scoring baselines, scDeepFusion matches or exceeds strong integration baselines on aggregate metrics, attaining the highest Overall score among integration baselines on all four established benchmarks, with particular strengths on rare and subtype-level populations in label-weak settings. The framework further scales to 400k-cell training runs on the >1M-cell Tabula Sapiens atlas. This work establishes a principled and auditable framework for using curated biological knowledge in single-cell representation learning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell data integration, interpretable representation learning, gene-set analysis, batch effect correction, mixture-of-experts, rare cell populations, scRNA-seq benchmarking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_E7CpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes scDeepFusion, a framework for single-cell RNA-seq integration that combines gene-expression fluctuations with module-level features derived from Reactome pathways and chemical/genetic perturbation signatures. A cell-specific top-2 router combines three knowledge-source experts and exports routing weights to track their contributions. The model is trained through alignment to a Harmony reference embedding, together with module reconstruction, gene denoising, and load-balancing objectives. The authors evaluate integration performance across five datasets and demonstrate training on up to 400,000 Tabula Sapiens cells. They report competitive performance, while ablations show that simpler teacher-aligned models achieve similar aggregate scores. The main claimed architectural contribution is therefore auditable knowledge provenance, explored through cell-type-specific routing patterns, replication across datasets, and post-training expert-masking experiments.

Critical strengths
  1. The model is able to track the use of the knowledge used for model representation explicitly through their routing strategy. It actually provides an explainable way for biologists who want to understand the model's internal predictive behavior.
  2. The authors explicitly acknowledge that simpler models achieve comparable aggregate performance, which provides useful empirical insight into the field with the adoption of teacher-aligned training.
Critical weaknesses

The author should demonstrate the biological relevance and practical value of the routing-based explanations. The innovative part of the work is the routing weights which offers potential interpretation to the predictive power, but the experiments do not quite establish whether these preferences align with independently supported biological mechanisms. Specifically, the replication of the myeloid–lymphoid routing pattern across datasets demonstrates consistency, but does not by itself establish biological aware understanding. Since these datasets contain similar immune-cell populations, the observed agreement could reflect shared expression patterns or biases in gene-set coverage rather than biologically meaningful expert specialization. It would be great to see some literatures or discoveries that could prove such preference shown in the model output. Known marker expression supports the identities of these populations, but does not explain why they should preferentially rely on Reactome or CGP experts. Therefore, I think unless independent biological evidence and controls for these alternative explanations can be verified, the author cannot claim the interpretable ability of the routing module.

Questions for authors
  1. What evidence supports interpreting the myeloid–lymphoid routing pattern as biologically meaningful expert specialization, rather than a consequence of shared expression patterns or differences in gene-set coverage? Can you connect these preferences to specific pathways or perturbation signatures supported by independent biological evidence?
  2. What does the expert-masking experiment establish beyond sensitivity of the weighted representation to removing a component? Does masking affect downstream integration quality, and are the routing patterns stable across training seeds?
  3. Given that the pathway-MLP, concatenation, and no-module controls achieve comparable aggregate performance, what demonstrated capability justifies the additional architecture? In particular, can you provide a concrete example in which the routing information helps identify a model failure or yields a biologically supported insight unavailable from the simpler controls?
1 comment

Thanks the authors for the detailed and comprehensive response and the new experiments conducted to address my comments. Although I appreciate that the author is willing to modify their claims on biological aware discovery, I think in general the novelty of the work could be reassessed given the current further analysis. From my point of view, I would expect the routing to show some biologically supported insight, since the central claim is combining gene expression with prior biological knowledge. Given on this, I would keep my current score. Thanks again for the authors’ effort on all the experiments.

Reviewer_Cqw3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents scDeepFusion, a single-cell RNA-seq integration framework combining gene-fluctuation features and curated gene-set activities through sparse routing over Reactome, perturbation-signature, and gene-expression experts. Training aligns a projected representation with a Harmony teacher, alongside reconstruction, denoising, and load-balancing objectives. The principal architectural contribution is per-cell knowledge-source provenance rather than improved aggregate integration scores.

Critical strengths
  1. The controls meaningfully separate performance from architectural complexity. Tables 6 and 8 show that simpler fusion, gene-only inputs, and established pathway scores with a distilled readout achieve similar aggregate performance on ifnb. The authors explicitly acknowledge that routing does not provide an aggregate-score advantage. These controls make the empirical findings more informative and support evaluating the architecture on its stated provenance objective.

  2. The provenance analysis goes beyond displaying routing weights. Section 3.3 and Tables 11–12 report cell-type-dependent responses to expert masking and a broadly replicated myeloid–lymphoid routing distinction on Immune Human, which lacks the stimulation condition of ifnb. This provides initial evidence that the source partition produces structured, biologically associated model behavior while preserving competitive integration performance.

Critical weaknesses

The practical significance of the provenance mechanism remains insufficiently established. The paper appropriately limits exact attribution to the fused combination uiu_i and describes weights as relative preferences. Table 11 supports sensitivity of this intermediate representation to expert removal, but does not establish the stronger claims in Section 3.3 that the attribution supplies an “actionable error signal” or that shared routing indicates rare-population consolidation is not a teacher artifact. Similar routing can coexist with teacher-driven organization, especially given the strong teacher dependence in Table 6. The cross-dataset evidence is encouraging, but the reported analyses do not yet show whether an audit conclusion is stable across training runs or helps identify a concrete integration failure. Since provenance is the main distinctive benefit over simpler controls, this limits the demonstrated significance of the contribution.

Questions for authors
  1. Table 13 reports score stability across three ifnb training seeds. Are the routing preferences and expert-masking conclusions in Section 3.3 also consistent across those checkpoints?
  2. What specific audit conclusion supports the “actionable error signal” claim?
Reviewer_a9x9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The manuscript presents a single-cell representation learning method designed to improve batch integration and interpretability. It first computes gene-module activity scores from curated pathways and perturbation signatures, then combines these features with gene-level representations using a sparse, cell-specific expert router.

The authors evaluate the method on several integration datasets and report competitive or reasonable performance. They also analyzed the learned routing weights, which show different preferences for pathway, perturbation-signature across types of immune cells.

Critical strengths

Competitive performance across multiple integration benchmarks. The method reports strong integration results on several datasets, including improvements over existing baselines, as shown in figure 1. The paper includes relatively detailed ablations and comparisons with simpler alternatives. The routing analysis also shows distinct preferences for knowledge-source experts across myeloid and lymphoid cells. This is an interesting direction for interpretable single-cell modeling.

Critical weaknesses

The main performance gains come from Harmony teacher distillation rather than the proposed architecture. In table 6, many of the variants with missing components from full model give better performance. Removing typed routing or the module branch (second and third rows) produces nearly identical overall scores to the full model (0.766 v.s. 0.767). Furthermore, a simple pathway-MLP control achieves essentially the same performance, and Table 8 shows that AUCell, GSVA, and UCell also achieve comparable results when trained with the same teacher. These results suggest that the teacher-alignment objective explains most of the reported performance, rather than the proposed module representation or typed-expert architecture. While the authors acknowledge this to some extent, it substantially weakens the methodological contribution of the paper.

Questions for authors
  1. Given that removing the module branch does not reduce aggregate performance on ifnb (table 6), what evidence supports the benefit of the proposed module-activity representation? Does it provide consistent advantages on other datasets or downstream evaluations?

  2. Can the authors provide stronger evidence that the routing weights capture biologically meaningful knowledge-source contributions? For example, comparisons against randomly partitioned gene sets or alternative expert groupings.

  3. Most evaluations outside ifnb use single runs. Can the authors provide additional evidence that the improvements remain meaningful across random seeds and without this unstable metric?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.