Research map / Graph learning / Knowledge and temporal graphs
Global Context-aware Representation Learning for Spatially Resolved Transcriptomics
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 Spotscape, a self-supervised GNN for spatially resolved transcriptomics that adds a global Similarity Telescope relation-consistency loss, prototypical contrastive learning, and a similarity scale matching loss for multi-slice batch-effect correction.
Problem
Graph-based representation learning for spatially resolved transcriptomics yields poor spot representations near spatial domain boundaries because local neighbors are overemphasized, and multi-slice integration suffers from batch effects.
Approach
Instead of relying on local neighbor attention, Spotscape enforces invariance of the global pairwise cosine-similarity matrix across two augmented graph views (Similarity Telescope), and matches top-k similarity scales between intra- and inter-slice spots to reduce batch effects, with prototypical contrastive learning for multi-slice grouping.
Key results
- Spotscape consistently outperforms baselines across all 16 slices in four single-slice datasets in terms of ARI, NMI, and CA.
- In heterogeneous integration on MTG, removing the similarity scaling loss causes a significant degradation in clustering performance.
- On DLPFC single-slice clustering with MLP encoder replaced, ARI drops from 0.48 to 0.20 (Spotscape ARI 0.48, NMI 0.64, CA 0.61).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a GCN encoder instead of an MLP encoder (ablation) | Spotscape · DLPFC | ARI: 0.20 (0.01) -> 0.48 (0.02) | Helped |
| use a GCN encoder instead of an MLP encoder (ablation) | Spotscape · DLPFC | NMI: 0.30 (0.01) -> 0.64 (0.01) | Helped |
| use a GCN encoder instead of an MLP encoder (ablation) | Spotscape · DLPFC | CA: 0.42 (0.02) -> 0.61 (0.02) | Helped |
| use a GCN encoder instead of an MLP encoder (ablation) | Spotscape · Postnatal Mouse Brain | ARI: 0.58 -> 0.61 | Helped |
| use a GCN encoder instead of an MLP encoder (ablation) | Spotscape · Postnatal Mouse Brain | NMI: 0.65 -> 0.68 | Helped |
| use a GCN encoder instead of an MLP encoder (ablation) | Spotscape · Postnatal Mouse Brain | CA: 0.67 -> 0.74 | Helped |
| add the Similarity Telescope relation consistency loss (ablation) | Spotscape · single-slice and multi-slice tasks | clustering performance | Helped |
| add the reconstruction loss (ablation) | Spotscape · single-slice and multi-slice tasks | performance | No clear effect |
| add prototypical contrastive learning (ablation) | Spotscape · MTG heterogeneous integration | clustering performance | Helped |
| add similarity scaling (ablation) | Spotscape · MTG heterogeneous integration | clustering performance | Helped |
| add prototypical contrastive learning (ablation) | Spotscape · single-slice tasks | clustering performance | Helped |
Surprising findings
- Oracle edge weights (removing heterophilic edges) give only marginal boundary-spot clustering gains over the original SNN graph GAE, suggesting local-view limitations persist even with ideal attention.
Limitations
- PCL loss is omitted in single-slice settings because of its marginal gain relative to running time.
- Hyperparameter search for baselines is more extensive than for Spotscape, which tunes only the learning rate.
Experimental setup
Datasets: DLPFC MTG Mouse Embryo NSCLC Breast Cancer (Visium) Breast Cancer (Xenium)
Benchmarks: DLPFC MTG Mouse Embryo NSCLC
Metrics: ARI NMI CA Silhouette Batch kBET Graph Connectivity PCR comparison LTARI Spearman correlation (trajectory) RMSE
Baselines: SEDR STAGATE SpaCAE SpaceFlow GraphST PASTE STAligner CAST PASTE2 SLAT
Builds on: STAGATE GraphST SpaceFlow PASTE SLAT
Closest papers by meaning
- STINR: Deciphering Spatial Transcriptomics via Implicit Neural Representation (CVPR 2025 · similarity 0.87)
- SToFM: a Multi-scale Foundation Model for Spatial Transcriptomics (ICML 2025 · similarity 0.87)
- SUICA: Learning Super-high Dimensional Sparse Implicit Neural Representations for Spatial Transcriptomics (ICML 2025 · similarity 0.87)
- SpotDiff: Spatial Gene Expression Imputation Diffusion with Single-Cell RNA Sequencing Data Integration (AAAI 2025 · similarity 0.86)
- Multi-modal Topology-embedded Graph Learning for Spatially Resolved Genes Prediction from Pathology Images with Prior Gene Similarity Information (CVPR 2025 · similarity 0.86)
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.86)
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.