acceptodds
Under review as a conference paper at ICLR 2027

scCGNet: Jointly Optimizing Copula Mixtures and Graph Autoencoders for Single-Cell Clustering

Abstract

Clustering is central to single-cell RNA sequencing (scRNA-seq) analysis but remains difficult: the data are high-dimensional, sparse, and correlated. Recent deep methods increasingly model this correlation to learn better representations- for example, through gene-gene structure or masked reconstruction. The clustering step, however, does not. After embedding, cells are grouped by K-means, deep embedded clustering, or contrastive objectives that rely on distances to cluster centroids. These objectives treat the latent dimensions as independent within a cluster, and cannot capture the non-Gaussian, asymmetric, and tail dependence that transcriptomic data exhibit. We propose scCGNet, which models dependence at the clustering stage. A variational graph autoencoder encodes a cell-cell k-nearest-neighbor graph into topology-aware latent representations. A Gaussian mixture copula model then clusters cells by explicitly modeling the dependence structure within each cluster. A zero-inflated negative binomial layer accounts for sparsity and over-dispersion. The three components share one objective and are trained jointly. On fifteen real scRNA-seq datasets, scCGNet attains the highest and most consistent accuracy, and ablations show that the copula-based cluster model drives the gain.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
deep clustering, variational graph autoencoder, single-cell RNA-seq, copula models

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_M5Z8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

scCGNet combines a variational graph autoencoder, a Gaussian mixture copula model, and ZINB reconstruction in one training framework for single-cell clustering. The authors test it on 15 real datasets and nine simulated datasets, and compare GMCM with GMM both during joint training and on fixed embeddings.

Critical strengths

The overall design is easy to follow: graph structure, expression reconstruction, and clustering are trained together. The experiments cover a fairly large number of datasets. It is also useful that the paper reports cases where GMCM performs worse than GMM on fixed embeddings, since this gives a more balanced view of the method.

Critical weaknesses

The paper does not give enough details about how the GMCM density is computed in the actual implementation. Equation 7 subtracts joint and marginal log densities, which can become unstable when the likelihood values are very small. A few implementation details are needed to show how this issue is handled during training. The practical cost of the method is also unclear. For large datasets, training time and peak GPU memory are important, but these numbers are not reported. The paper also does not provide a directly comparable cost measurement for the baseline methods.

Questions for authors

How do you keep the computation in Equation 7 numerically stable during training?

Could you report the wall-clock training time and peak GPU memory on representative datasets, together with the same measurements for the baseline methods?

Reviewer_yUU6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes scCGNet, combining a variational graph autoencoder, a Gaussian mixture copula model (GMCM) over latent representations, and a zero-inflated negative-binomial reconstruction head in a joint objective. It evaluates 15 real scRNA-seq datasets and nine simulated datasets, reporting mean real-data ARI 0.8702 and NMI 0.8443. The paper argues that copula-based clustering drives the gains, but fixed-embedding tests do not consistently support that mechanism. In addition, the appendix states that checkpoints are selected by ground-truth ARI, which risks label leakage into model selection and materially weakens the reported evaluation.

Critical strengths
  1. There is evidence of sound statistical motivation. Copulas separate marginal distributions from dependence, directly addressing a limitation of centroid-based clustering, although the paper does not establish that this component drives the gains.

  2. The reconstruction objectives reflect distinct properties of scRNA-seq data. Graph reconstruction preserves cell-neighbourhood structure, while the ZINB objective models sparse, overdispersed counts, providing distinct signals for representation learning. These objectives give the representation multiple data-grounded constraints rather than relying on cluster labels alone

  3. Data is tested on a large scale. Fifteen real datasets test performance across varied data, while synthetic marginal transformations test robustness to distributional changes.

Critical weaknesses
  1. Appendix K selects the epoch with the highest ARI against ground-truth labels. This leaks evaluation-label information into model selection and may inflate reported clustering results.

  2. Copula contribution is not isolated convincingly. GMCM underperforms GMM on most fixed real embeddings, while the joint-training ablation covers only two datasets. The evidence is insufficient to establish the copula as the primary source of improvement.

  3. The parameter K is set to the annotated number of cell types. This is common in benchmarks but does not test cluster-number discovery and may simplify the task.

  4. Results lack seed-level uncertainty and contain a numerical inconsistency. The main tables report point estimates without standard deviations. Also, Section 4.2 gives scRCL mean ARI as approximately 0.724, whereas Table 1 reports 0.7436. This discrepancy affects the stated performance margin over the baseline and should be reconciled.

Questions for authors
  1. Appendix K states that checkpoints are selected by the highest ground-truth ARI. Can you re-evaluate using a label-free model-selection/stopping rule and report how much the results change?

  2. Given that GMCM performs worse than GMM on most fixed real embeddings and the joint-training ablation is limited, what evidence supports the claim that the copula component is the primary contributor? Can you include a matched GMM-in-the-loop ablation with the same warm start, training schedule, and checkpoint-selection rule?

  3. How sensitive are results to under- or over-specifying the parameter K? Please distinguish the standard benchmark setting where K is known from the more general unsupervised setting.

  4. Can you report per-dataset means and standard deviations across seeds for the main real-data benchmark, and reconcile the scRCL ARI discrepancy (approximately 0.724 in Section 4.2 vs 0.7436 in Table 1)?

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

This paper proposes an unsupervised framework for scRNA-seq clustering called scCGNet. scCGNet consists of a variational graph autoencoder, a Gaussian mixture copula model, and a ZINB decoder. Specifically, VGAE uses a cell-cell KNN graph to learn cell representations, GMCM uses a copula-based clustering objective to capture dependencies among representations, and the ZINB decoder accounts for sparsity and overdispersion in gene expression counts. The authors evaluate scCGNet on 15 real scRNA-seq datasets against eight baselines, reporting ARI and NMI. Additional experiments include evaluations on simulated datasets, parameter sensitivity analysis, ablation studies, and biological analyses based on cell-type recovery and marker gene expression.

Critical strengths

scCGNet has good clustering performance, it achieves the highest average ARI and NMI across 15 real scRNA-seq datasets compared with eight baseline methods, showing its effectiveness for single-cell clustering.

Critical weaknesses
  1. scCGNet lacks novelty, the main contribution is applying the existing GMCM to scRNA-seq clustering, while VGAE and ZINB are already widely used. The overall framework mainly combines existing techniques, without introducing much methodological innovation.

  2. scCGNet lacks sufficient evaluation of its key contribution. Since GMCM is the main novelty, its advantages over existing clustering methods should be better demonstrated. The end-to-end comparison with GMM is only conducted on two datasets. Moreover, the standalone GMCM performs worse than GMM on three of the four real datasets in Appendix J. These results raise questions about whether GMCM is necessary and how much it contributes to the overall performance.

Questions for authors

Why is GMCM better suited for scRNA-seq clustering than existing methods? The authors only compare GMCM with GMM on two datasets in the ablation study. Also, Appendix J shows that standalone GMCM performs worse than GMM on three of the four real datasets. Could the authors explain why this happens and provide more evidence to show the benefit of using GMCM?

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

The paper proposes scCGNet, a framework for single-cell RNA-seq clustering that trains a variational graph autoencoder, a Gaussian mixture copula clustering head, and a ZINB count decoder jointly under a single objective. The central argument is that existing methods already model correlations among genes and cells during representation learning, yet at the clustering step they fall back on centroid-distance criteria such as K-means or deep embedded clustering and thereby discard the dependence structure they have learned. The authors therefore introduce a copula at the clustering stage so that marginal distributions and dependence structure are modeled separately.

Critical strengths

S1. Dependence modeling in the representation-learning stage has become fairly mature over the years, whereas clustering criteria have long remained geometric. Shifting the attention from the representation to the statistical assumptions of the clustering criterion itself is a worthwhile reframing.

S2. Conventional Gaussian mixture copula fitting relies on pseudo-EM together with numerical inversion of the marginal CDFs, which is slow and hard to keep differentiable everywhere. The authors observe that the mixture is defined directly on the latent space, so the inverse transform reduces to the identity, and the copula density can be evaluated as the joint density divided by the product of the per-dimension marginal densities. The whole path is then analytically differentiable with respect to both the network parameters and the mixture parameters. This is what allows a copula clustering head to be embedded in end-to-end gradient training for the first time, and it is the technical basis on which the claim of being the first end-to-end copula-based graph clustering framework rests.

S3. The three heads constrain one another in a well-motivated way. Graph reconstruction pulls topologically adjacent cells together in the latent space, the KL term pushes the latent space toward a standard normal so that it remains continuous and without holes, and the ZINB term forces the latent representation to reproduce the observed counts so that it cannot drift toward directions unrelated to biology. Together they provide a constraint boundary for the clustering term.

S4. The evaluation spans fifteen real datasets and nine SPARSim simulated datasets, with eight baselines covering centroid-based, graph-based, and mixture-based families.

S5. The authors do not stop at reporting clustering accuracy but design multiple layers of analysis around the core component. Moving from reported performance to the underlying mechanism and to the conditions under which it holds is an approach that is not common in method papers.

Critical weaknesses

W1. The incremental novelty of this work is limited, as it mainly stacks existing modules. The variational graph autoencoder is essentially the original formulation of Kipf and Welling (2016), ZINB count modeling is the standard configuration established by scDeepCluster (2019), and placing a mixture model on top of a graph autoencoder for clustering was already done by scGMM-VGAE (2023). What genuinely belongs to this paper is the GMCM head. Yet by Eq. 9 the marginal denominator in Eq. 7 is common to all mixture components and therefore cancels, so the responsibilities reduce to the ordinary mixture form. As a consequence, for identical mixture parameters the hard assignments of the version with the copula and the version without it coincide point by point, and the copula cannot change the assignment of any cell at inference. This is in clear tension with the repeated emphasis in the abstract, the contribution list, and the conclusion that the copula is the main source of the performance gain.

W2. The sensitivity analysis and related experiments are validated only on QS-Diaphragm, where performance is already close to the ceiling, which makes them unreliable and raises the concern of selective reporting. QS-Diaphragm is the smallest and easiest of the fifteen datasets. I have myself reproduced several baselines on this dataset and likewise observed that all methods tend to obtain good results there. Fixing k=15 and two thousand highly variable genes for all fifteen datasets on the basis of a single easy dataset is not convincing.

W3. The paper emphasizes scalability but provides no supporting runtime or memory measurements.

Questions for authors

Q1. I would like the authors to clarify whether the objective in Eq. 8 is bounded from below. If the mean of one component is placed exactly on a data point, its variance is driven to zero, and its weight is driven to zero at the rate needed for that component to remain dominant at that point, then the copula log-density at that point grows at the rate (1-d) times the log weight. The clustering loss is the negative mean of this quantity, so for latent dimension d of at least 2 the loss is unbounded below. In my own numerical experiments the loss can be pushed to nearly minus forty at d=32, and the paper uses latent dimensions up to 512. I noticed that the code clamps the log variance to plus or minus ten, but this is not mentioned in the paper, and since it constrains the variance rather than the weights it does not appear to block this path. Could the authors explain how the existence of an optimum is guaranteed in theory, whether component collapse was observed during training, and whether that clamp should be presented as a formal regularizer in the method section.

Q2. The code in the supplementary material does not appear to match the method described in the paper. Line 254 of main.py fixes cluster_type to the string GMM when constructing the model, while lines 187 to 190 of vgae_model.py show that the copula clustering head is instantiated only when this argument is GMCM, and that the value GMM instantiates a plain Gaussian mixture head without the marginal normalization term. Running python main.py as instructed in the README would therefore produce the ablation variant used as a control in Sec. 4.4 rather than the scCGNet proposed in the paper. Could the authors confirm the actual script configuration that produced the scCGNet row in Table 1 and Table 2, and provide code corresponding to the reported results.

Q3. The symbols theta and pi are each defined twice. I suggest unifying them.

Q4. Figure 1 is not very good-looking.

Q5. Typo errors. l.315 sumarized -> summarized l.797 has a significant in improving -> has a significant effect in improving l.809 scCGNet still obtain comparble results -> still obtains comparable results l.936 latent embedding on six datasets -> on four datasets

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.