Pith. sign in

REVIEW 4 major objections 7 minor 47 references

Soft Graph Clustering for single-cell RNA Sequencing Data

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Replacing hard binary cell graphs with soft weighted edges gives scSGC roughly 4% average gains in ACC, NMI, and ARI over the runner-up across ten scRNA-seq datasets.

desk verdict A well-described but empirically under-verified soft-graph clustering method; the real-data benchmark needs a protocol audit before the 4% gain claim can be believed. read the letter →

arxiv 2507.09890 v1 pith:OLNILE2G submitted 2025-07-14 cs.LG cs.AIq-bio.GN

Ping Xu , Pengfei Wang , Zhiyuan Ning , Meng Xiao , Min Wu , Yuanchun Zhou This is my paper · ORCID
classification cs.LGcs.AIq-bio.GN
keywords scRNA-seqclusteringsoftgraphneuralnetworkzero-inflatednegativebinomialoptimaltransportnormalizedcutcelltypeannotationsingle-cellRNAsequencing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes scSGC, a clustering framework for single-cell RNA sequencing data whose central idea is to replace the binary (0/1) cell-to-cell edges used by graph neural network baselines with soft, continuously weighted edges. It argues that hard thresholded graphs discard transitional similarity information and create spurious inter-cluster connections that mislead message passing, and that soft graphs avoid both failure modes. The framework couples a zero-inflated negative binomial autoencoder for dropout and sparsity, a dual-channel soft graph embedding trained with a joint normalized-cut objective, and an optimal transport clustering module. On ten datasets spanning pancreas, liver, and muscle cells, the paper reports that scSGC beats thirteen compared methods on accuracy, normalized mutual information, and adjusted Rand index by about 4% on average, with better cell-type annotation and lower runtime.

What carries the argument

The load-bearing object is the dual-channel cut-informed soft graph embedding module. It builds two weighted adjacency matrices—inner-product similarity and cosine similarity—then minimizes a jointly normalized cut objective $\mathrm{Tr}(Z^T(\alpha L_1 + (1-\alpha)L_2)Z)$ with an orthogonality penalty on $Z$, fusing both graphs' structural information before clustering. The other two modules, the ZINB autoencoder and the optimal transport (Sinkhorn) clustering step, handle feature sparsity and assignment balance, but the soft graph is what differentiates scSGC from hard-graph GNN baselines.

What would settle it

Run scSGC with its soft graph replaced by a thresholded binary graph, keeping all other modules and hyperparameters identical, on the same ten datasets; the paper's claim predicts a clear drop in ACC, NMI, and ARI. Separately, rerun with the ground-truth cluster count $k$ supplied instead of estimated by K-means; if the margin over baselines moves by more than a few points, the reported advantage depends on the evaluation protocol rather than the soft graph construction.

Watch

Extended reading notes

Core claim

The central discovery is that soft graph construction—keeping similarity as non-binary edge weights instead of thresholding to 0/1—gives graph-based clustering a consistent edge for scRNA-seq data. The paper shows this through a dual-channel design: a feature-similarity graph and a cosine-similarity graph, whose normalized Laplacians are fused by minimizing a joint normalized cut, so the continuous structure of cell-cell similarity drives the embedding. On top of that embedding, a ZINB autoencoder models dropout and an optimal transport step refines cluster assignments. The reported result is an average improvement of 4.06% (ACC), 3.725% (NMI), and 4.03% (ARI) over the runner-up method, with the largest margins in the imbalanced, high-dropout simulated datasets.

Load-bearing premise

The load-bearing premise is that the evaluation was genuinely unsupervised: per-dataset hyperparameters and the cluster count $k$ were chosen without using the true cell-type labels, so the reported ~4% margins reflect method quality rather than selection.

Editorial extensions

If this is right

  • If the reported margins hold, cell type annotation from scRNA-seq data becomes more accurate without extra experimental cost, and the method runs faster than nine compared approaches on the tested datasets.
  • The joint normalized cut gives a principled way to combine multiple similarity measures, so adding or reweighting graph channels could further improve the embedding without changing the rest of the framework.
  • The optimal transport module prevents degenerate, unbalanced cluster assignments, which is directly relevant to datasets with rare cell types or strong class imbalance.
  • The modular design separates feature modeling, graph structure, and assignment optimization, and the authors state the framework is intended to extend to other high-dimensional omics data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A strict ablation that replaces the soft graph with a thresholded binary graph while keeping the ZINB and optimal transport modules identical would directly test whether the soft edges are the source of the reported gain.
  • Because the paper reports per-dataset hyperparameter tuning and estimates the cluster count $k$ from the embeddings, the 4% advantage could shrink if any of those choices implicitly used the true labels.
  • The same dual-Laplacian joint NCut design could be applied to other continuous-similarity clustering problems beyond single-cell data, where thresholded graphs are also common.
  • The simulated-data results suggest the largest gains occur in imbalanced, high-dropout settings, implying the method's practical value may be greatest for noisy datasets with rare cell populations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes scSGC, a soft-graph clustering method for single-cell RNA-seq data, combining a ZINB-based autoencoder, a dual-channel cut-informed soft graph embedding module, and an optimal-transport clustering optimization module. The authors claim that on ten real scRNA-seq datasets, scSGC outperforms 13 state-of-the-art baselines by average margins of about 4% in ACC, NMI, and ARI, and that it is also computationally efficient. The central contribution is empirical: the method is a modular combination of known components, and its claimed superiority rests on the experimental results in Tables 3 and 5 and the associated qualitative analyses.

Significance. If the reported results are unbiased, scSGC would be a practically useful addition to the scRNA-seq clustering toolkit: the soft-graph formulation addresses a real limitation of hard-thresholded graph constructions, the ZINB autoencoder is well matched to dropout-heavy count data, and the simulated-data experiment with fixed hyperparameters (§3.7) is a sound practice that partially addresses generalization concerns. The paper also reports ablation studies for each module and promises a public code repository, which are strengths that support reproducibility. However, the headline claim is empirical, and the evaluation protocol has unresolved label-leakage risks; the significance of the contribution therefore hinges on correcting those risks.

major comments (4)
  1. [§3.1.4 and Table 2] The paper states that the authors 'conduct meticulous parameter tuning across different datasets to ensure optimal model performance' and reports dataset-specific values for α, β, γ, weight decay, and the number of highly expressed genes, but no validation split or nested cross-validation procedure is described. With at least five loss-balancing hyperparameters plus weight decay and gene-count selection, tuning directly on the test labels could manufacture the reported 4% margin. Please specify exactly how each hyperparameter value in Table 2 was selected, and state explicitly whether any ground-truth cell-type labels were used at any point in the selection process. If an unsupervised validation criterion was used, name it and give the chosen value for each dataset.
  2. [§3.1.3] The sentence 'we apply K-means to the embeddings to obtain the optimal value of k' is under-specified because K-means requires k as an input. No unsupervised criterion (e.g., silhouette score, eigengap, stability analysis) is described. If the annotated number of cell types was supplied, or if k was selected by maximizing ACC/NMI/ARI against ground truth, then all reported metrics in Table 3 are optimistically biased. Please provide the exact rule used to choose k for each dataset and confirm that it does not use label information.
  3. [§2.2.3, Eqs. (10)–(13)] The optimal-transport module derives its target distribution P from the model's own intermediate assignments: Eq. (10) constrains P to match cluster proportions π 'estimated from intermediate clustering results,' and Eq. (13) then minimizes KL(P || Q) where Q is the model's own assignment matrix. This is a self-training loop that can reinforce existing assignments without adding new information. The ablation in §3.4.3 compares this module against DEC, but both are self-training mechanisms. Please clarify what prevents P from merely copying Q, and provide evidence that the OT module changes the optimization landscape rather than simply re-weighting the model's own output, for example by comparing against a target distribution that is independent of the current Q or by initialization from random assignments.
  4. [Table 3 and §3.2.1] The headline claim that 'scSGC outperforms all other models across all three metrics' is supported only by means over ten runs, with no paired significance tests or multiple-comparison correction. Some differences are large, but the table also shows high variance for many baselines, and the paper does not report whether the observed 4% average margins are statistically distinguishable from noise under paired testing. Please add paired tests (e.g., Wilcoxon signed-rank or paired bootstrap) across the ten datasets, or report effect sizes with confidence intervals, for each metric before claiming universal superiority.
minor comments (7)
  1. [Abstract, Background, Conclusion] The number of baselines is inconsistent: the abstract says 13 state-of-the-art models, while the Background and Conclusion say eleven; Table 3 actually contains 13 baselines plus scSGC. Please unify the count.
  2. [Tables 1 and 2] The dataset name 'Mauro Human Pancreas cells' is a typo for 'Muraro Human Pancreas cells'; the same typo appears in Table 2 and in the text.
  3. [§3.3.1 and Fig. 4] The text says that the overlap of top 100 DEGs was computed for ten methods, but Fig. 4 shows only eight panels; please correct the count or add the missing panels.
  4. [Fig. 3 caption] The caption refers to 'DEG' in the title, but the compared method is DEC (Deep Embedded Clustering); please correct the acronym.
  5. [§3.1.4] The sentence 'the hyperparameters λ, θ and µ are initially set to 5, 1, and 1e-3' introduces θ, which is not listed as a tunable hyperparameter in §3.5 and appears to be the ZINB dispersion parameter estimated by the decoder. Please clarify whether θ is a fixed hyperparameter or a learned parameter.
  6. [Eq. (5)] The expression A2 = |cos(X, X)| · |cos(X, X)|^T is ambiguously typed: cos(X, X) is not defined as a matrix operation, and the absolute value and transpose need elementwise definitions. Please define the notation precisely.
  7. [§6.3] The data and code availability statement says the code 'will be made available' at a GitHub URL, but the manuscript does not include a versioned repository or DOI. Please provide an accessible repository at submission time so that the experiments can be reproduced.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the method's equations are explicit and the headline empirical claim is tested on external public benchmarks; only minor self-referential design (optimal-transport target derived from own assignments) and a disclosed scCDCG ancestry are noted.

full rationale

The paper's contribution is an empirical architecture, and the derivation chain is largely self-contained. The ZINB loss (Eq. 4), the dual-channel soft-graph NCut loss (Eqs. 5-9), and the total loss (Eq. 14) are all stated explicitly, so the method does not depend on an unstated external result. The optimal-transport module (Eqs. 10-13) is a self-training loop: the target distribution P is obtained by Sinkhorn normalization of the current assignment Q, with cluster proportions π estimated from intermediate clustering results, so P is a sharpened and balanced version of Q rather than an independent label source. I do not count this as circularity of the central claim, because it is an optimization objective, not an external prediction, and the reported ACC/NMI/ARI values are measured against ground-truth cell-type labels of ten public datasets. The paper also discloses that scSGC was developed from the authors' own scCDCG framework (Section 3.7) and compares against scCDCG, but this self-citation is transparent and does not carry the argument: the equations are given and the baselines are independently implemented. The only real concern is the unspecified rule for choosing k ('we apply K-means to the embeddings to obtain the optimal value of k', Section 3.1.3) and the per-dataset tuning in Table 2; if ground-truth labels were used in those selections, the reported 4% margin could be inflated. That is a protocol-correctness risk, not an exhibited circular reduction, so it is noted but not scored as circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a ZINB distributional assumption, a similarity-is-informative assumption, standard spectral and optimal transport mathematics, and self-training. The free parameters are all hyperparameters tuned per dataset, which creates a risk of test-set leakage.

free parameters (6)
  • alpha (graph balance weight) = 0.1 to 0.75 per dataset (Table 2)
    Balances the two soft graphs in Eq. 8; tuned separately for each dataset.
  • beta (orthogonality regularization) = 10 to 30 per dataset (Table 2)
    Controls the orthogonality penalty in Eq. 9; tuned per dataset.
  • gamma (ZINB loss weight) = 20 to 100 per dataset (Table 2)
    Weights the ZINB reconstruction loss in Eq. 14; tuned per dataset.
  • mu (KL loss weight) = 1e-3 initially, 1e-3 recommended
    Weights the clustering KL loss in Eq. 14; set by hand and adjusted per dataset.
  • lambda (entropy regularization in optimal transport) = 5 initially, 50 recommended
    Controls Sinkhorn smoothing in Eq. 11; recommended value changes between sections.
  • number of highly expressed genes n = 1500 to 8000 per dataset (Table 2)
    Selects the top n genes during preprocessing (Section 3.1.1); varies per dataset and affects all downstream results.
assumptions (5)
  • domain assumption ZINB distribution models dropout and overdispersion in scRNA-seq counts
    Used in Eq. 1-4 as the autoencoder loss; this is a standard but unproven statistical assumption about scRNA-seq data generation.
  • domain assumption Cell-cell similarity in raw expression space is informative for cell types
    The soft graphs G1 and G2 are constructed from inner products and cosine similarities of the preprocessed data (Section 2.2.2), assuming these reflect biological similarity.
  • standard math Spectral relaxation of the joint normalized cut is a valid clustering objective
    Eq. 7-9 follow the standard spectral clustering relaxation; the orthogonality is enforced only as a soft penalty, so the solution is an approximation.
  • standard math Sinkhorn iterations converge to the entropy-regularized optimal transport plan
    Eq. 12 applies the standard Sinkhorn fixed-point algorithm; convergence is assumed without discussion.
  • domain assumption Self-training with the model's own confident assignments improves clustering
    The OT module uses proportions pi estimated from intermediate clustering results (Eq. 10-13), a common but not guaranteed assumption in deep clustering.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Soft Graph Clustering for single-cell RNA Sequencing Data." pith.science (2026). https://pith.science/paper/OLNILE2G

@misc{pith2026250709890,
  author       = {Pith},
  title        = {Pith review of: Soft Graph Clustering for single-cell RNA Sequencing Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OLNILE2G}},
  note         = {Machine review of arXiv:2507.09890}
}
read the original abstract

Clustering analysis is fundamental in single-cell RNA sequencing (scRNA-seq) data analysis for elucidating cellular heterogeneity and diversity. Recent graph-based scRNA-seq clustering methods, particularly graph neural networks (GNNs), have significantly improved in tackling the challenges of high-dimension, high-sparsity, and frequent dropout events that lead to ambiguous cell population boundaries. However, their reliance on hard graph constructions derived from thresholded similarity matrices presents challenges:(i) The simplification of intercellular relationships into binary edges (0 or 1) by applying thresholds, which restricts the capture of continuous similarity features among cells and leads to significant information loss.(ii) The presence of significant inter-cluster connections within hard graphs, which can confuse GNN methods that rely heavily on graph structures, potentially causing erroneous message propagation and biased clustering outcomes. To tackle these challenges, we introduce scSGC, a Soft Graph Clustering for single-cell RNA sequencing data, which aims to more accurately characterize continuous similarities among cells through non-binary edge weights, thereby mitigating the limitations of rigid data structures. The scSGC framework comprises three core components: (i) a zero-inflated negative binomial (ZINB)-based feature autoencoder; (ii) a dual-channel cut-informed soft graph embedding module; and (iii) an optimal transport-based clustering optimization module. Extensive experiments across ten datasets demonstrate that scSGC outperforms 13 state-of-the-art clustering models in clustering accuracy, cell type annotation, and computational efficiency. These results highlight its substantial potential to advance scRNA-seq data analysis and deepen our understanding of cellular heterogeneity.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 43 canonical work pages

  1. [1]

    Cell Research, 1–16 (2024)

    Yang, X., Liu, G., Feng, G., Bu, D., Wang, P., Jiang, J., Chen, S., Yang, Q., Miao, H., Zhang, Y., et al.: Genecompass: deciphering universal gene regulatory mecha- nisms with a knowledge-informed cross-species foundation model. Cell Research, 1–16 (2024)

    work page 2024

  2. [2]

    Nature Reviews Genetics 20(5), 273–282 (2019)

    Kiselev, V.Y., Andrews, T.S., Hemberg, M.: Challenges in unsupervised clustering of single-cell rna-seq data. Nature Reviews Genetics 20(5), 273–282 (2019)

    work page 2019

  3. [3]

    Advanced Science, 2500870 (2025)

    Wang, P., Liu, W., Wang, J., Liu, Y., Li, P., Xu, P., Cui, W., Zhang, R., Long, Q., Hu, Z., et al.: sccompass: An integrated multi-species scrna-seq database for ai-ready. Advanced Science, 2500870 (2025)

    work page 2025

  4. [4]

    Nature methods 13(10), 845–848 (2016)

    Haghverdi, L., B¨ uttner, M., Wolf, F.A., Buettner, F., Theis, F.J.: Diffusion pseu- dotime robustly reconstructs lineage branching. Nature methods 13(10), 845–848 (2016)

    work page 2016

  5. [5]

    Cell 162(1), 184–197 (2015)

    Levine, J.H., Simonds, E.F., Bendall, S.C., Davis, K.L., El-ad, D.A., Tadmor, M.D., Litvin, O., Fienberg, H.G., Jager, A., Zunder, E.R., et al.: Data-driven phe- notypic dissection of aml reveals progenitor-like cells that correlate with prognosis. Cell 162(1), 184–197 (2015)

    work page 2015

  6. [6]

    Cell 174(3), 716–729 (2018)

    Van Dijk, D., Sharma, R., Nainys, J., Yim, K., Kathail, P., Carr, A.J., Bur- dziak, C., Moon, K.R., Chaffer, C.L., Pattabiraman, D., et al.: Recovering gene interactions from single-cell data using data diffusion. Cell 174(3), 716–729 (2018)

    work page 2018

  7. [7]

    Nature biotechnology 36(5), 411–420 (2018)

    Butler, A., Hoffman, P., Smibert, P., Papalexi, E., Satija, R.: Integrating single- cell transcriptomic data across different conditions, technologies, and species. Nature biotechnology 36(5), 411–420 (2018)

    work page 2018

  8. [8]

    Genome biology 18(1), 1–11 (2017)

    Lin, P., Troup, M., Ho, J.W.: Cidr: Ultrafast and accurate clustering through imputation for single-cell rna-seq data. Genome biology 18(1), 1–11 (2017)

    work page 2017

Show all 47 references
  1. [9]

    Proteomics 18(2), 1700232 (2018)

    Wang, B., Ramazzotti, D., De Sano, L., Zhu, J., Pierson, E., Batzoglou, S.: Simlr: A tool for large-scale genomic analyses by multi-kernel learning. Proteomics 18(2), 1700232 (2018)

    2018

  2. [10]

    Bioinformatics 34(12), 2069–2076 (2018)

    Park, S., Zhao, H.: Spectral clustering based on learning similarity matrix. Bioinformatics 34(12), 2069–2076 (2018)

    2018

  3. [11]

    Nature Machine Intelligence 1(4), 191–198 (2019)

    Tian, T., Wan, J., Song, Q., Wei, Z.: Clustering single-cell rna-seq data with a model-based deep learning approach. Nature Machine Intelligence 1(4), 191–198 (2019)

    2019

  4. [12]

    PLOS Computational Biology 19(11), 1011641 (2023)

    Li, S., Guo, H., Zhang, S., Li, Y., Li, M.: Attention-based deep clustering method 21 for scrna-seq cell type identification. PLOS Computational Biology 19(11), 1011641 (2023)

    2023

  5. [13]

    BMC medical genomics 11, 65–72 (2018)

    Gan, Y., Li, N., Zou, G., Xin, Y., Guan, J.: Identification of cancer subtypes from single-cell rna-seq data using a consensus clustering method. BMC medical genomics 11, 65–72 (2018)

    2018

  6. [14]

    Scientific reports 9(1), 1–12 (2019)

    Traag, V.A., Waltman, L., Van Eck, N.J.: From louvain to leiden: guaranteeing well-connected communities. Scientific reports 9(1), 1–12 (2019)

    2019

  7. [15]

    Scientific reports 11(1), 20028 (2021)

    Luo, Z., Xu, C., Zhang, Z., Jin, W.: A topology-preserving dimensionality reduc- tion method for single-cell rna-seq data using graph autoencoder. Scientific reports 11(1), 20028 (2021)

    2021

  8. [16]

    Bioinformatics 38(4), 1037–1044 (2022)

    Ciortan, M., Defrance, M.: Gnn-based embedding for clustering scrna-seq data. Bioinformatics 38(4), 1037–1044 (2022)

    2022

  9. [17]

    Nature communications 12(1), 1882 (2021)

    Wang, J., Ma, A., Chang, Y., Gong, J., Jiang, Y., Qi, R., Wang, C., Fu, H., Ma, Q., Xu, D.: scgnn is a novel graph neural network framework for single-cell rna-seq analyses. Nature communications 12(1), 1882 (2021)

    2021

  10. [18]

    Briefings in Bioinformatics 23(2), 018 (2022)

    Gan, Y., Huang, X., Zou, G., Zhou, S., Guan, J.: Deep structural clustering for single-cell rna-seq data jointly through autoencoder and graph neural network. Briefings in Bioinformatics 23(2), 018 (2022)

    2022

  11. [19]

    arXiv preprint arXiv:2406.16992 (2024)

    Zhou, Y., Wang, P., Dong, H., Zhang, D., Yang, D., Fu, Y., Wang, P.: Make graph neural networks great again: A generic integration paradigm of topology- free patterns for traffic speed prediction. arXiv preprint arXiv:2406.16992 (2024)

    2024 arXiv

  12. [20]

    In: Pacific-Asia Conference on Knowledge Discovery and Data Mining, pp

    Ning, Z., Qiao, Z., Dong, H., Du, Y., Zhou, Y.: Lightcake: A lightweight framework for context-aware knowledge graph embedding. In: Pacific-Asia Conference on Knowledge Discovery and Data Mining, pp. 181–193 (2021). Springer

    2021

  13. [21]

    Theoretical Computer Science 389(1-2), 219–236 (2007)

    Alekseev, V.E., Boliac, R., Korobitsyn, D.V., Lozin, V.V.: Np-hard graph prob- lems and boundary classes of graphs. Theoretical Computer Science 389(1-2), 219–236 (2007)

    2007

  14. [22]

    In: 2021 Interna- tional Conference on Cyber-Physical Social Intelligence (ICCSI), pp

    Tang, Y., Huang, Z., Cheng, J., Zhou, G., Feng, S., Zheng, H.: Graph neural network-based node classification with hard sample strategy. In: 2021 Interna- tional Conference on Cyber-Physical Social Intelligence (ICCSI), pp. 1–4 (2021). IEEE

    2021

  15. [23]

    Frontiers of Computer Science 19(4), 194311 (2025)

    WU, Y., HUANG, H., SONG, Y., JIN, H.: Soft-gnn: towards robust graph neural networks via self-adaptive data utilization. Frontiers of Computer Science 19(4), 194311 (2025)

    2025

  16. [24]

    arXiv preprint arXiv:2209.13964 (2022)

    Ning, Z., Wang, P., Wang, P., Qiao, Z., Fan, W., Zhang, D., Du, Y., Zhou, 22 Y.: Graph soft-contrastive learning via neighborhood ranking. arXiv preprint arXiv:2209.13964 (2022)

    2022 arXiv

  17. [25]

    arXiv preprint arXiv:2503.06635 (2025)

    Ning, Z., Wang, Z., Zhang, R., Xu, P., Liu, K., Wang, P., Ju, W., Wang, P., Zhou, Y., Cambria, E., et al.: Deep cut-informed graph embedding and clustering. arXiv preprint arXiv:2503.06635 (2025)

    2025 arXiv

  18. [26]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Chen, D., Lin, Y., Li, W., Li, P., Zhou, J., Sun, X.: Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 3438–3445 (2020)

    2020

  19. [27]

    arXiv preprint arXiv:2505.05533 (2025)

    Ning, Z., Wang, P., Qiao, Z., Wang, P., Zhou, Y.: Rethinking graph contrastive learning through relative similarity preservation. arXiv preprint arXiv:2505.05533 (2025)

    2025 arXiv

  20. [28]

    General mathematics notes 21(2) (2014)

    Thumbakara, R.K., George, B.: Soft graphs. General mathematics notes 21(2) (2014)

    2014

  21. [29]

    arXiv preprint arXiv:2405.09591 (2024)

    Wang, Z., Wang, P., Liu, K., Wang, P., Fu, Y., Lu, C.-T., Aggarwal, C.C., Pei, J., Zhou, Y.: A comprehensive survey on data augmentation. arXiv preprint arXiv:2405.09591 (2024)

    2024

  22. [30]

    In: International Conference on Database Systems for Advanced Applications, pp

    Xu, P., Ning, Z., Xiao, M., Feng, G., Li, X., Zhou, Y., Wang, P.: sccdcg: Effi- cient deep structural clustering for single-cell rna-seq via deep cut-informed graph embedding. In: International Conference on Database Systems for Advanced Applications, pp. 172–187 (2024). Springer

    2024

  23. [31]

    arXiv preprint arXiv:2505.12626 (2025)

    Xu, P., Ning, Z., Li, P., Liu, W., Wang, P., Cui, J., Zhou, Y., Wang, P.: scsiamese- clu: A siamese clustering framework for interpreting single-cell rna sequencing data. arXiv preprint arXiv:2505.12626 (2025)

    2025

  24. [32]

    Cell systems 3(4), 346–360 (2016)

    Baron, M., Veres, A., Wolock, S.L., Faust, A.L., Gaujoux, R., Vetere, A., Ryu, J.H., Wagner, B.K., Shen-Orr, S.S., Klein, A.M., et al.: A single-cell tran- scriptomic map of the human and mouse pancreas reveals inter-and intra-cell population structure. Cell systems 3(4), 346–...

    2016

  25. [33]

    Cell systems 3(4), 385–394 (2016)

    Muraro, M.J., Dharmadhikari, G., Gr¨ un, D., Groen, N., Dielen, T., Jansen, E., Van Gurp, L., Engelse, M.A., Carlotti, F., De Koning, E.J., et al.: A single-cell transcriptome atlas of the human pancreas. Cell systems 3(4), 385–394 (2016)

    2016

  26. [34]

    Nature communications8(1), 14049 (2017)

    Zheng, G.X., Terry, J.M., Belgrader, P., Ryvkin, P., Bent, Z.W., Wilson, R., Ziraldo, S.B., Wheeler, T.D., McDermott, G.P., Zhu, J., et al.: Massively parallel digital transcriptional profiling of single cells. Nature communications8(1), 14049 (2017)

    2017

  27. [35]

    Nature communications 9(1), 4383 (2018)

    MacParland, S.A., Liu, J.C., Ma, X.-Z., Innes, B.T., Bartczak, A.M., Gage, B.K., 23 Manuel, J., Khuu, N., Echeverri, J., Linares, I., et al.: Single cell rna sequenc- ing of human liver reveals distinct intrahepatic macrophage populations. Nature communications 9(1), 4383 (2018)

    2018

  28. [36]

    Nature 583(7817), 590–595 (2020)

    A single-cell transcriptomic atlas characterizes ageing tissues in the mouse. Nature 583(7817), 590–595 (2020)

    2020

  29. [37]

    BMC bioinformatics 17, 1–11 (2016)

    ˇZurauskien˙ e, J., Yau, C.: pcareduce: hierarchical clustering of single cell tran- scriptional profiles. BMC bioinformatics 17, 1–11 (2016)

    2016

  30. [38]

    Interdisciplinary Sciences: Computational Life Sciences 13, 83–90 (2021)

    Wang, H.-Y., Zhao, J.-p., Zheng, C.-H.: Suscc: secondary construction of feature space based on umap for rapid and accurate clustering large-scale single cell rna-seq data. Interdisciplinary Sciences: Computational Life Sciences 13, 83–90 (2021)

    2021

  31. [39]

    In: International Conference on Machine Learning, pp

    Xie, J., Girshick, R., Farhadi, A.: Unsupervised deep embedding for clustering analysis. In: International Conference on Machine Learning, pp. 478–487 (2016). PMLR

    2016

  32. [40]

    BMC bioinformatics 22(1), 280 (2021)

    Ciortan, M., Defrance, M.: Contrastive self-supervised clustering of scrna-seq data. BMC bioinformatics 22(1), 280 (2021)

    2021

  33. [41]

    PLOS Computational Biology 18(12), 1010772 (2022)

    Wang, H., Zhao, J., Zheng, C., Su, Y.: scdssc: Deep sparse subspace clustering for scrna-seq data. PLOS Computational Biology 18(12), 1010772 (2022)

    2022

  34. [42]

    Bioinformatics 38(6), 1575–1583 (2022)

    Wan, H., Chen, L., Deng, M.: scname: neighborhood contrastive clustering with ancillary mask estimation for scrna-seq data. Bioinformatics 38(6), 1575–1583 (2022)

    2022

  35. [43]

    NAR genomics and bioinformatics 2(2), 039 (2020)

    Chen, L., Wang, W., Zhai, Y., Deng, M.: Deep soft k-means clustering with self- training for single-cell rna sequence data. NAR genomics and bioinformatics 2(2), 039 (2020)

    2020

  36. [44]

    Genome biology 19, 1–5 (2018)

    Wolf, F.A., Angerer, P., Theis, F.J.: Scanpy: large-scale single-cell gene expression data analysis. Genome biology 19, 1–5 (2018)

    2018

  37. [45]

    Journal of machine learning research 3(Dec), 583–617 (2002)

    Strehl, A., Ghosh, J.: Cluster ensembles—a knowledge reuse framework for com- bining multiple partitions. Journal of machine learning research 3(Dec), 583–617 (2002)

    2002

  38. [46]

    1073–1080 (2009)

    Vinh, N.X., Epps, J., Bailey, J.: Information theoretic measures for clusterings comparison: is a correction for chance necessary? In: Proceedings of the 26th Annual International Conference on Machine Learning, pp. 1073–1080 (2009)

    2009

  39. [47]

    Genome biology 18(1), 174 (2017) 24

    Zappia, L., Phipson, B., Oshlack, A.: Splatter: simulation of single-cell rna sequencing data. Genome biology 18(1), 174 (2017) 24

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.