acceptodds
Under review as a conference paper at ICLR 2027

scROBM: Role-Guided Bottleneck Model for Single-Cell Latent Representation Learning

Abstract

Single-cell expressions mix cell identity, condition response, inter-individual background, and technical variation, and reconstruction alone gives no account of which latent coordinates carry which role. We present scROBM , a role-objective-guided bottleneck model for expression-only single-cell representation learning. scROBM assigns three cell-level latent variables to cell identity, perturbation semantics, and donor context, together with a shared donor-level center, and organizes training around two coupled bottlenecks: a label-free count decoder that forces every label-dependent reconstruction signal through the assigned latents, and a conditional-invariance (CI) penalty that suppresses condition information in the joint background means within supported donor–cell-type strata while leaving donor and cell-type differences unconstrained. We evaluate scROBM on four single cell data sets. Donor-associated structure shows the clearest separation, whereas cell-type and condition remain in a high, dataset-dependent range; because every latent is supervised for a single role, the three exports stay interpretable, and the donor-context export resolves donors while exposing a cell-type-dependent condition structure concentrated in the treatment's target compartment.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
distribution prior, representative learning, single cell data

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_8zHWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a role-guided latent representation model for single-cell RNA-seq data. The main motivation is that single-cell expression jointly reflects multiple sources of variation, including cell identity, experimental condition, and donor-specific background, while conventional shared representations do not explicitly specify which latent variables encode which factors.

Critical strengths

The motivation of explicitly organizing experimental semantics in the latent representation is meaningful. I appreciate the paper’s attempt to move beyond a generic shared embedding and explicitly allocate different latent variables to biologically and experimentally meaningful factors. Besides, the label-free decoder, role-specific supervision, donor-ranking objective, and conditional-invariance constraint are each motivated by a particular aspect of the proposed role allocation.

Critical weaknesses
  1. The main contribution of paper is framed around learning role-specific latent representations for cell identity, condition, and donor context. However, several closely related single-cell disentanglement methods, including contrastiveVI, biolord, and TarDis, are discussed in the related-work section but are not included in the experimental comparison. Although these methods do not necessarily target exactly the same donor-aware setting as scROBM, they are substantially more aligned with the paper’s central claim of role-specific latent disentanglement than generic representation-learning baselines such as PCA or scVI.

  2. The experimental validation is not sufficiently comprehensive for the proposed representation-learning framework. Reconstruction is an important component of the model, but the paper does not report reconstruction or held-out count-modeling quality. In addition, the evaluation of condition-related representations is limited to binary condition classification, leaving it unclear whether the learned representation can distinguish richer experimental or perturbation conditions. More comprehensive evaluation would be needed to support the broader representation-learning claims.

Questions for authors

In addition to the concerns raised above, I have the following questions:

  1. What exactly is meant by “cross-dataset generalization” in Section 6.3? Is a model trained on one dataset transferred to another dataset, or is scROBM trained separately on each dataset?

  2. The exported representations are computed from expression alone, but training explicitly uses cell-type, condition, and donor metadata for role supervision, conditional invariance, and donor ranking. How should the term “expression-only representation learning” be interpreted?

  3. Two of the evaluated datasets contain only three and four donors, respectively, while donor-associated representation is the setting where scROBM shows its clearest empirical advantage. How robust are the donor-level conclusions given the relatively small number of donors in several datasets?

  4. Since the label-free count decoder and reconstruction bottleneck are central components of scROBM, could the authors report reconstruction or held-out count-modeling performance?

  5. The stimulus representation is described as encoding perturbation semantics, but all evaluated datasets appear to involve only binary condition contrasts. Can learned representation distinguish among multiple distinct perturbations when more than two perturbation conditions are present?

  6. Could the authors clarify the meaning of terms such as “health-validation artifact” and “declared final split”? These expressions are used to qualify the main experimental comparison, but they are neither clearly defined nor supported by references, making it difficult to understand their precise meaning and implications for the evaluation. I also encourage the authors to improve the overall clarity of the writing.

Reviewer_hwMtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents scROBM, a supervised generative model for learning role-guided representations of single-cell RNA-sequencing data. It assigns three cell-level latent blocks to cell identity, condition semantics, and donor-associated variation, together with shared donor-level centers. The model combines a serial expression-only encoder, a hierarchical Gaussian prior, and a ZINB count decoder that receives the latent variables and a fixed library-size exposure but no cell-type, condition, or donor labels. Training uses role-specific classification losses, cross-condition donor-ranking supervision within cell types, and an MMD-based conditional-invariance penalty on the joint deterministic basal and donor-context exports within supported donor–cell-type strata. The authors report experiments on four datasets covering cell-type and condition prediction, donor matching, retrieval, and clustering. The clearest reported distinction concerns donor-associated structure, and the ablation identifies donor-ranking supervision as the largest contributor to matching and retrieval. The paper also provides qualitative observations of the three representations in an anti-PD-1 dataset. For the main scBloodNL-MTB benchmark, the authors explicitly qualify the baseline comparisons as descriptive because the scROBM export is a “health-validation artifact,” while the fresh baselines use their declared final split.

Critical strengths
  1. The paper addresses a relevant representation-learning problem with a coherent design. For studies targeting repeated-donor correspondence, retaining donor-associated structure across conditions is a relevant objective. Conditioning the invariance penalty on donor and cell type is a sensible way to encourage alignment within supported strata: CI directly compares conditions within a stratum and does not directly match distributions across different donors or cell types. The expression-only inference path also avoids directly supplying the prediction targets as encoder inputs when exporting cell representations.
  2. The donor-ranking ablation provides meaningful evidence for the retrieval objective. Table 2 reports Recall@1 of 0.6477 and MRR of 0.7811 for the Gaussian configuration, compared with 0.4318 and 0.6523 after removing donor-ranking supervision. Within the reported ablation, these point estimates support explicitly learning cross-condition donor correspondence. The authors also acknowledge that removing CI or role supervision can improve some clustering measures, which helps delimit the empirical claims.
Critical weaknesses
  1. The main comparative results do not yet establish an advantage under a clearly matched held-out protocol. Lines 350–352 state that the scROBM export is a “health-validation artifact,” while the fresh baselines use their declared final split, and explicitly describe the comparisons as descriptive. I therefore cannot treat Table 1 as a matched final-test comparison without clarification of the evaluation provenance. The main text does not identify which donors, eligible strata, or candidate sets, if any, differ between the reported artifacts. Section 5.1 also mentions “22 eligible baseline donors” and “a separate four-donor scROBM preview” for GSE236581 without sufficiently explaining how those descriptions relate to the comparisons in Figure 3. This uncertainty matters because donor-associated performance is the main empirical distinction emphasized by the paper.
  2. The results in the main paper leave the achieved effect and incremental value of conditional invariance incompletely established. The largest matching and retrieval ablation effect comes from donor-ranking supervision. Relative to the Gaussian configuration, removing CI changes Recall@1 from 0.6477 to 0.6023 and MRR from 0.7811 to 0.7631, while clustering ARI increases from 0.0613 to 0.0837. These results support a possible retrieval benefit, but the donor-performance metrics alone do not directly quantify residual condition information in the joint background export. Section 6.4 describes condition separation in the donor-context visualization within one donor and the T-cell group. The authors explicitly acknowledge that residual condition information can remain, so the degree of invariance achieved in this example requires clarification. More generally, I find the contribution of the complete construction beyond the donor-ranking supervision insufficiently isolated by the main-text comparisons. The main text also does not quantify the uncertainty of the smaller component effects.
Questions for authors
  1. Evaluation provenance and comparability. Section 5.3 states that checkpoints are selected using development RQ3 matching AUROC. What exactly does “health-validation artifact” mean in relation to that selection and the evaluation reported in Table 1? Please specify the evaluated donors and query/candidate construction for the full model, baselines, and ablations, and point to any existing common-protocol comparisons and uncertainty estimates. Please also clarify how the GSE236581 donor descriptions in Section 5.1 map to the results in Figure 3.
  2. Contribution beyond donor-ranking supervision. Table 2 already ablates CI, role supervision, and donor-ranking supervision. Do any existing results additionally isolate the contribution of the shared donor centers or the serial/hierarchical architecture, for example through a simpler representation model trained with comparable donor and role supervision? Such evidence would help assess how much of the observed retrieval benefit depends on the full construction.
  3. Achieved conditional invariance and Figure 4. Please point to any quantitative measurements of condition predictability from the joint background export within supported donor–cell-type strata, with and without CI, particularly on held-out data and using a probe capable of detecting nonlinear interactions. For the donor–T-cell example described in Section 6.4 and Figure 4, which CI configuration was used, and was the relevant stratum supported during training? How does the example relate to the intended invariance target?
  4. Kernel-specific interpretation. Taking Eq. 5 as the intended population target, please provide the exact formulas for the three kernels referenced in Eq. S12, particularly the block combination and bandwidth scaling in the linear mode. For each kernel, what would zero population MMD imply about the conditional distributions, and under what assumptions? How should these distinctions inform the interpretation of finite-training-set CI optimization, given the residual-condition-readability limitation acknowledged at lines 481–485?
Reviewer_ggkDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes scROBM, a VAE based model for single cell expression representation learning. It assigns three cell level latent variables to distinct roles. These roles are cell identity basal perturbation or condition semantics stimulus and donor context. The model also includes a shared donor level center. Training couples two bottlenecks. One is a label free ZINB count decoder. It forces label dependent reconstruction signal through the latents for architectural constraint. The other is a conditional invariance MMD penalty. It suppresses condition information in background means within donor cell type strata for semantic constraint. The training objective further includes role cross entropy and a donor ranking InfoNCE loss. On four donor structured datasets cell type and condition tasks are near saturated for all methods. Its method achieves strong performance on donor matching retrieval and clustering against PCA scVI scANVI and scDisInFact.

Critical strengths

1.Role aware latent allocation with expression only inference avoids label circularity at test time. The open set unseen donor evaluation protocol is sensible.

2.Stratified Horvitz Thompson minibatching analytic donor center KL and the unbiased finite set CI estimator from Lemmas 1 to 3 are technically solid and well documented.Component contributions are cleanly isolated. Donor ranking dominates.

Critical weaknesses

1.In the main benchmark beats scVI on matching AUROC only marginally while losing matching AUPRC. RQ1 and RQ2 are saturated for all methods and thus non informative. Donor clustering ARI remains very low even for scROBM.

2.DisInFact contrastiveVI biolord and TarDis are close to this work. They assign latent factors to conditions and covariates. They only appear in related work and are left out of experiments. Simple baselines such as kNN on raw and HVG expression are also missing. The paper does not show the weaknesses of these prior approaches. It is unclear which practical limitation this new combination addresses.

3.The paper itself acknowledges in Section 6.1 that the scROBM export is a health validation artifact while the fresh baselines use their declared final split. So the headline comparison is descriptive rather than controlled.

Questions for authors

See Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.