acceptodds
Under review as a conference paper at ICLR 2027

RIPPLE: A Self-Dimensioning Factor Model for Single-Cell Transcriptomics

Abstract

Single-cell atlases are becoming biology's reference maps, and almost every downstream analysis runs on a low-dimensional embedding of the counts rather than on the counts themselves. Such an analysis must discover the cell types and states of interest, so the representation has to be learned without labels. Two choices govern it and are usually left unprincipled: how many factors the representation carries, and whether its coordinates mean anything. We present RIPPLE, an unsupervised negative binomial factor model that settles both from one design. A sparsity-inducing half-Cauchy prior on the per-factor loading scales collapses the factors the data do not support, and the active dimension is read off the fitted posterior by a signal-to-noise rule with no threshold to tune. The decoder is linear, so every factor that survives is a gene program that can be read directly. Under mild conditions, we prove that this rule loses no true factor, and that the fitted signal carried beyond the true factors is asymptotically weightless. On a six-dataset, eight-method benchmark, RIPPLE outperforms every unsupervised baseline on every biology-preserving metric averaged over the datasets. On the 585k-cell Human Lung Cell Atlas, a single selected factor separates each of the 61 annotated cell types at a median AUROC of 0.96, independently of how rare the cell type is.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell, Variational Inference, Sparse Factor Analysis, Automatic Relevance Determination, Unsupervised Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dkGvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper introduces a generative model for single-cell data, which selects the dimensionality of the latent space by prior-fitting.

Critical strengths

Overall the paper is well-written and tackles an interesting problem.

Critical weaknesses

I am asking reviewers to assess two critical weaknesses:

W1: Relevancy of the problem setting:

In simple-terms, the procedure the authors propose seems to pick the top-k most informative latent dimensions. This dimensionality-reduction step is part of many real-world single-cell pipelines, and other literature has looked at automating this step with different heuristics, most prominently the 'elbow plot'. For different heuristics, see e.g.: Haotian Zhuang et al., findPC: An R package to automatically select the number of principal components in single-cell analysis, Bioinformatics 2022 How do these heuristics compare to the approach of the paper?

W2: Unrealistic benchmarking assumptions:

The authors benchmark with a ground truth number of cluster labels 'k'. This ground truth is not available in real-life scenarios, and changes the goal of the embedding. When analysing new single-cell data, biologists need to a) first define the available clusters in the data, b) assigning cell identities to them. A good embedding therefore supports both: selecting the number of clusters and cluster assignment. Knowing the ground truth number of clusters changes this task.

Minor comments:

  • The paper sometimes contains surprising formulations in the appendix, e.g. L1153 no per-dataset configuration file overrides any of them
Questions for authors
  • See Weakness 1: How were the dimensionalities for the datasets picked for the benchmarks? How does the method perform against established heuristics for picking the dimensionality?
  • See Weakness 2: How does not-specifing k in the benchmark impact performance?
  • What is the biological relevancy of the scenario in L373 (removing canonical markers):
  • What is the implication of the theoretical guarantees?
Reviewer_g9Z7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RIPPLE, an unsupervised negative-binomial factor model for single-cell transcriptomics that addresses two limitations of existing low-dimensional representations: 1 pre-specify the latent dimension, 2 the limited interpretability of nonlinear latent coordinates. RIPPLE uses factor-wise Bayesian shrinkage priors and mean-field variational inference, and selects active factors after fitting using a posterior signal-to-noise criterion. Its linear decoder also allows the retained factors to be interpreted as gene programs.

The paper provides theoretical results for the identifiable design-residual signal, including consistency of the fitted signal, asymptotic preservation of sufficiently strong true factor directions in the selected span, and asymptotically negligible fitted signal beyond the true rank. Empirically, RIPPLE is compared with 7 baselines on six public scRNA-seq datasets and achieves good performance across several clustering metrics. The method is further evaluated on the Human Lung Cell Atlas with approximately 585k cells, demonstrating scalability and stable, interpretable factors.

Critical strengths
  1. The paper is technically thorough and generally careful in calibrating its claims.
  2. The empirical evaluation is broad and unusually detailed for an interpretable factor model.
  3. The modeling choices are well aligned with the practical requirements of single-cell analysis.
Critical weaknesses
  1. The paper lacks a clean synthetic experiment with known ground-truth latent dimension and factors. This is particularly important because “self-dimensioning” is one of the central claims of the paper, while the true latent dimension is unknown in all real datasets. A controlled simulation could directly measure under-selection, over-selection, recovery of the true latent dimension, subspace recovery, and sensitivity to factor strength. The current real-data experiments demonstrate stability and usefulness of the selected factors, but they cannot establish whether the inferred dimension is actually correct.
  2. The contribution of the self-dimensioning mechanism is not fully isolated. RIPPLE combines shrinkage-based factor selection with neighborhood regularization and covariates, both of which contribute noticeably to performance. A controlled synthetic study with known latent dimension, together with a matched fixed-rank ablation, would make it much easier to determine whether the observed gains are specifically attributable to the proposed self-dimensioning mechanism.
Questions for authors

1 The theoretical guarantees do not directly cover the deployed optimization procedure. The analysis assumes an idealized maximizer of the variational objective, whereas the practical algorithm uses stochastic optimization and a stop-gradient update that need not be stationary for the analyzed objective. The authors clearly acknowledge this limitation, but optimization itself remains outside the theory. 2. The prior in this paper is like a ARD-style factor pruning plus half-Cauchy factor-scale prior. The novelty relative to classical Bayesian dimension-selection methods could be positioned more clearly. The core idea of over-specifying the latent dimension and shrinking unsupported factors is closely related to ARD/Bayesian PCA and subsequent Bayesian shrinkage factor models. While the paper briefly acknowledges this connection, a more explicit comparison with prior work on shrinkage-based factor-number selection would help clarify which aspects of RIPPLE constitute the main methodological advance. 3. The theorem is correct for an idealized estimator, but it does not directly certify the implemented optimizer. I did not identify an obvious mathematical error in the stated theoretical results. My concern is instead about the scope of the guarantees: they apply to an idealized variational maximizer under fixed-design assumptions, while the deployed stochastic optimizer and some data-dependent covariates fall outside the theory.

Reviewer_WmZZOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces RIPPLE, an unsupervised negative binomial factor model for single-cell transcriptomics. A halfCauchy prior on the factor loading scales shrinks factors the data do not support, and the active dimension is read off the fitted posterior by a signal-to-noise rule. A linear decoder makes the factors directly readable as gene programs, while a graph regularizer encourages similar cells to have similar embeddings. Under stated assumptions, the authors show that the selection rule loses no true factor and that the fitted signal beyond the true rank is asymptotically weightless. Experiments compare eight methods across six datasets and demonstrate a 24-factor representation of the Human Lung Cell Atlas, evaluated through celltype separation, factor stability, and concordance with pseudobulk differential expression.

Critical strengths

1.The paper addresses two practical challenges in single-cell representation learning: choosing the number of latent reduction dimension and making the resulting factors interpretable. 2.RIPPLE achieves strong average clustering performance across six datasets and scales to the Human Lung Cell Atlas. It demonstrates this practical utility of the algorithm in the clustering task and the stability of the fitted factors across random seeds.

Critical weaknesses

1.This paper claims that RIPPLE solves the number selection problem in dimension reduction task, however, it can also tune the number of factors by selecting among different SNRs. 2.Some messages delivered are non-informative, such as the paragraph of 1.1 notation. 3.Figure 1 evaluates embeddings using k-means at the true number of cell types. This protocol may favor particular embedding geometries, and differences in latent dimensionality and hyperparameter settings make it difficult to attribute the performance gains to the proposed method.

Questions for authors

1.What distinguishes RIPPLE from NMF, MOFA, and topic models in terms of factor selection and interpretability, and what evidence supports these advantages? 2.How sensitive are the relative rankings to the clustering algorithm and latent dimensionality? Were comparable hyperparameter-selection procedures used across methods? 3.What benefit does half-Cauchy shrinkage provide over alternative penalties, particularly comparing withlasso penalty? Can the existing experiments isolate this benefit from graph regularization and SNR selection?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.