acceptodds
Withdrawn from ICLR 2027

scProto: Affinity-Guided Prototype Learning for Cross-Batch Metacell Construction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Resolving which cell states are present in a tissue is essential for characterizing cellular heterogeneity in development and disease. Single-cell and spatial transcriptomics profile many donors and sections at once, but individual cells are too sparse to analyze alone, so cells are aggregated into metacells. Current methods form metacells from expression proximity alone: a state rare in every batch but common overall cannot be formed, and a cell's spatial niche is either absorbed into cell-type signal or, if added to the affinity directly, overcorrects and splits one state across neighborhoods. We introduce scProto, a prototype-based deep clustering model that learns metacells across an entire experiment rather than within each batch, using the affinity graph as a training objective rather than clustering it directly: a conditional variational autoencoder learns a batch-corrected representation jointly with soft prototype assignment under a cohesion criterion that keeps each group internally dense via both direct and shared-neighborhood similarity. Assignments come from expression alone, so a niche separates cells only to the extent it changes their transcriptome. Across single-cell and spatial datasets, scProto recovers rare and niche-associated states missed by within-batch clustering, improving batch mixing and rare-population recovery while preserving neighborhood structure.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell transcriptomics, metacells, deep clustering, representation learning, graph-based community detection, batch effect correction, spatial transcriptomics, variational autoencoder

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

this paper provide a training framework that uses a cVAE that explicitly learning both gene expression values and batch information to align cells from batches, create meta-cells, and generate soft-assignments. the method is validate on several metrics including homogeneity, F1, and coverage, demonstrating its effectiveness.

Critical strengths

S1: batch effect is ubiquitous in gene-expression data analysis and aligning cells across batch is very meaningful to facilitate and advance relevant research.

Critical weaknesses

W1: the aim and scope of this paper is unclear. does it simply create meta-cells and that's all? is there any downstream applications in bioinformatics? how can such meta-cell construction method help their performance? do you mean clustering by prototyping, and does this present a prototype-based cell clustering solution? the impact of this paper should be made more concrete specific.

W2: the method section is not well organized. it is presented in a flat manner, enumerating all model or training component. but this is no hierachical structures that demonstrate the inter-relationship between these design components.

Questions for authors

Q1: in eq.8 the reconstruction is g(z,b) but in figure 1 the reconstruction is g(c,b), please clarify.

Q2: please clariy the concept "meta-cell" and "prototype". what is the discrepency and what is in common, in your scenario. unify the terminology if approriate.

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

This paper introduces scProto, a prototype-based framework for constructing metacells across batches. It combines a pretrained batch-conditional autoencoder with affinity-guided objectives that encourage cells with similar local and shared-neighborhood structure to receive similar prototype assignments. The framework also incorporates spatial context through the input affinity graph. Experiments on three multi-batch scRNA-seq datasets and one spatial transcriptomics dataset evaluate batch mixing, metacell coherence, rare-cell representation, and preservation of niche-associated expression programs.

Critical strengths
  1. The problem is well motivated. Cross-batch aggregation could improve the representation of rare cell populations that are insufficiently sampled within individual batches. The Pancreas results support this motivation: scProto improves rare-cell F1 and coverage over both SEACells and its own pretrained scPoli representation clustered with Leiden.
  2. The framework provides a flexible way to incorporate biological affinities. Separating expression-based encoding from affinity-guided grouping allows spatial context to influence metacell construction while retaining an expression reconstruction objective. The comparison with SEACells under increasing spatial affinity weights provides preliminary evidence that this design can better preserve cell-type identity when spatial context becomes dominant.
Critical weaknesses
  1. The spatial transcriptomics evidence is limited relative to its prominence in the paper. The spatial evaluation uses only one selected NSCLC slide from a dataset containing six slides. Its main result is a spatial-weight sweep against SEACells, showing that scProto deteriorates less as spatial influence increases. This establishes robustness to stronger spatial guidance more directly than it establishes a benefit from incorporating spatial information: scProto’s reported niche-program recovery decreases from 0.81 to 0.74 across the sweep. The highlighted CXCL9 example is illustrative, but does not establish systematic recovery of spatially meaningful biology. Validation across independent slides, together with a clear improvement over the expression-only configuration on spatial biological endpoints, is needed to substantiate this central claim.
  2. The scRNA-seq results do not establish a consistent advantage over strong existing approaches or the pretrained baseline. Improvements are clearest on Pancreas, whereas on Immune, rare-cell performance is comparable to or slightly below SEACells and SuperCell. On Lung, scProto underperforms scPoli plus Leiden in cell-type purity (0.72 versus 0.84), batch entropy (1.16 versus 1.41), and rare-cell F1 (0.55 versus 0.62); MetaCell2 also achieves higher rare-cell F1 and coverage. These results leave the practical benefit of the additional prototype-learning objectives uncertain. Higher cohesion alone is insufficient to resolve this concern because cohesion is directly optimized during training. Matched comparisons and component ablations demonstrating improvements on independent biological outcomes would strengthen the case for the method.
Questions for authors
  1. Does incorporating spatial information improve scProto over its expression-only configuration on spatial biological outcomes? The spatial-weight sweep primarily demonstrates less degradation than SEACells. Please clarify which results establish a positive benefit from spatial guidance.
  2. How consistent are the spatial findings across slides? Why was slide 28 selected, and do the conclusions hold on the other available NSCLC slides, particularly for niche-program recovery?
  3. Could you visualize the inferred metacell assignments in the original spatial coordinates, alongside cell-type and niche annotations, for both scProto and the relevant baselines? This would help assess whether the metacells capture coherent niche-associated states while preserving cell identity.
  4. Do the spatial findings generalize beyond the selected NSCLC slide? Comparisons of joint metacell construction across multiple slides—and, where available, donors or experimental batches—would be needed to substantiate the applicability of cross-batch aggregation to spatial transcriptomics. In particular, does joint analysis improve rare or niche-associated state recovery over analyzing each slide separately, without merging biologically distinct states?
  5. What explains the degradation on Lung relative to scPoli plus Leiden, including lower cell-type purity, batch entropy, and rare-cell F1? Can component ablations identify when affinity-guided prototype learning helps and when it harms the pretrained representation?
  6. Are the held-out cells used to select hyperparameters by purity separate from those used for the final evaluation? Please clarify the validation/test split and whether comparable tuning procedures were applied to the baselines.
  7. The cited ICLR 2025 LMRL workshop paper, “Interpretable Self-Supervised Prototype Learning for Single-Cell Transcriptomics,” already introduces a method named scProto that combines a scPoli-based conditional VAE, graph-guided prototype learning, cross-batch metacell construction, and rare-cell coverage. Could the authors explicitly delineate the methodological advances over this work and provide a controlled comparison with its batch-wise SwAV and propagation objectives? In particular, do the proposed community and cohesion objectives improve rare-cell recovery and biological preservation under matched architectures, prototype counts, data splits, and tuning budgets?
Reviewer_ZSbJOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents scProto, a prototype-based method for cross-batch metacell construction. It combines scPoli pretraining with affinity-based community and cohesion objectives, and extends the framework to spatial data.

Critical strengths
  • The paper addresses a useful problem: constructing metacells across batches while retaining rare cell populations.
  • On Pancreas, scProto improves rare-cell F1 from 0.33 to 0.54 and coverage from 0.40 to 0.77 over scPoli+Leiden.
Critical weaknesses
  • Limited methodological novelty. The method combines an existing batch-corrected encoder with prototype clustering and standard graph objectives. The two cohesion terms use direct affinity and shared-neighborhood similarity within a classical clustering objective. The paper does not establish a clear methodological advance beyond this combination.

  • The added training does not consistently improve biological structure preservation. Compared with scPoli+Leiden, cell-type purity decreases on all three datasets: 0.97→0.93 on Pancreas, 0.87→0.82 on Immune, and 0.84→0.72 on Lung. On Lung, rare-cell F1 also drops from 0.62 to 0.55, while batch entropy decreases from 1.41 to 1.16. Thus, the loss in purity is not accompanied by better batch mixing. The higher cohesion scores provide limited additional evidence, since they measure the same similarities directly optimized during training.

  • The model-selection protocol leaves a risk of test-label leakage. Appendix D selects the community and reconstruction loss weights by maximizing held-out cell-type purity, but does not explain whether this validation set is separate from the held-out cells used for final evaluation. Without this distinction, the node-holdout results cannot be treated as an independent test. The paper should also clarify whether baselines receive comparable access to cell-type labels for hyperparameter selection.

  • The contribution of each component remains unclear. Appendix D provides a diagnostic for gradient routing, but there is no component-wise ablation of the community, reconstruction, and two cohesion terms. Given that the pretrained encoder alone performs better on several biological metrics, these ablations are needed to show which additions help and whether the full objective is justified.

Questions for authors

No further questions, see weaknesses please.

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

Related papers

Discussion (0)

Sign in to comment.