Research map / Graph learning / Graph self-supervised learning
Unsupervised Gene-Cell Collective Representation Learning with Optimal Transport
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes scGCOT, an end-to-end deep graph clustering model for scRNA-seq that jointly learns cell and gene graph embeddings, uses a ZINB decoder, and aligns cluster distributions with optimal transport (Sinkhorn divergence).
Problem
Cell type identification from scRNA-seq data is hard due to high dimensionality and sparsity; existing deep clustering methods model cell correlations but ignore gene correlations and often lack principled noise/count modeling.
Approach
Jointly learns attention-based graph autoencoders over both the cell graph and gene graph, reconstructs the count matrix via the inner product of cell and gene embeddings with a ZINB likelihood, and replaces KL-based DEC self-training with an OT (Sinkhorn) alignment between the clustering distribution and its auxiliary target.
Key results
- scGCOT achieves average NMI 0.8226 and ARI 0.8084 on the three additional datasets' averages (Table 3), highest among all baselines.
- scGCOT obtains the best NMI on 11 of 15 datasets and best ARI on 10 of 15 datasets compared to state-of-the-art methods.
- Ablation: removing the ZINB loss Lz causes the largest drop (NMI 0.8226 to 0.7733, ARI 0.8084 to 0.7139).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use scGCOT instead of scNAME | scRNA-seq clustering · AVG (all datasets) | NMI: 0.7930 -> 0.8226 | Helped |
| Use scGCOT instead of scBGEDA | scRNA-seq clustering · AVG (all datasets) | ARI: 0.7496 -> 0.8084 | Helped |
| add gene graph reconstruction loss Lg (ablation) | scGCOT · AVG (all datasets) | NMI: 0.8104 -> 0.8226 | Helped |
| add gene graph reconstruction loss Lg (ablation) | scGCOT · AVG (all datasets) | ARI: 0.7864 -> 0.8084 | Helped |
| add count matrix reconstruction loss Lr (ablation) | scGCOT · AVG (all datasets) | NMI: 0.7990 -> 0.8226 | Helped |
| add count matrix reconstruction loss Lr (ablation) | scGCOT · AVG (all datasets) | ARI: 0.7658 -> 0.8084 | Helped |
| add ZINB loss Lz (ablation) | scGCOT · AVG (all datasets) | NMI: 0.7733 -> 0.8226 | Helped |
| add ZINB loss Lz (ablation) | scGCOT · AVG (all datasets) | ARI: 0.7139 -> 0.8084 | Helped |
| add OT alignment loss Ls (ablation) | scGCOT · AVG (all datasets) | NMI: 0.7976 -> 0.8226 | Helped |
| add OT alignment loss Ls (ablation) | scGCOT · AVG (all datasets) | ARI: 0.7648 -> 0.8084 | Helped |
| add graph attention (ablation) | scGCOT · AVG (all datasets) | NMI: 0.7906 -> 0.8226 | Helped |
| add graph attention (ablation) | scGCOT · AVG (all datasets) | ARI: 0.7544 -> 0.8084 | Helped |
Surprising findings
- Several deep embedded baselines (DCA, scDeepCluster, scziDesk) give unstable performance across datasets, which the authors attribute to relying only on the expression count matrix.
Limitations
- Some baselines (scGAE, scDFC, scGAC on certain datasets) ran out of memory on a single RTX 4090, so results are missing for those cases.
- Performance on several datasets is mixed; scGCOT does not always achieve best ARI (10 of 15 datasets).
Experimental setup
Datasets: p3cl Muraro Qx Limb Muscle Qs Diaphragm Qs Heart Qs Limb Muscle Qs Lung Young Adam Plasschaert Chen Xin
Benchmarks: p3cl Muraro Qx Limb Muscle Qs Diaphragm Qs Heart Qs Limb Muscle Qs Lung Young Adam Plasschaert Chen Xin
Metrics: NMI ARI
Baselines: K-means Spectral clustering DCA scDeepCluster GraphSCC scziDesk scGAE scDSC scGAC scNAME CAPKM++2.0 scDFC
Builds on: scDeepCluster DEC Graph Transformer Sinkhorn divergence scGAC
Closest papers by meaning
- Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data (NeurIPS 2024 · similarity 0.86)
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.85)
- sciLaMA: A Single-Cell Representation Learning Framework to Leverage Prior Knowledge from Large Language Models (ICML 2025 · similarity 0.85)
- A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity (ICML 2025 · similarity 0.84)
- Cell ontology guided transcriptome foundation model (NeurIPS 2024 · similarity 0.84)
- GENOT: Entropic (Gromov) Wasserstein Flow Matching with Applications to Single-Cell Genomics (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.