Research map / Graph learning / Graph self-supervised learning
DUSTED: Dual-Attention Enhanced Spatial Transcriptomics Denoiser
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 DUSTED, a self-supervised graph autoencoder with gene channel attention and graph attention that denoises spatial transcriptomics counts under NB/ZINB likelihoods without histology images.
Problem
Technical noise in spatially resolved transcriptomics (SRT) data blurs gene expression and spatial patterns; existing denoisers ignore spatial structure, gene-specific noise levels, or rely on histological images.
Approach
A graph autoencoder with a gene channel attention module (capturing gene-specific noise variability) plus GAT-based spatial attention, whose decoder outputs per-gene mean, dispersion and dropout parameters fitted with NB or ZINB likelihoods, requiring no auxiliary image data.
Key results
- On 24 simulated SRT datasets, DUSTED reached average MSE 0.3839, versus 0.3925 for DCA, 0.4004 for MAGIC and 1.0831 for raw data (about 69.6% reduction vs raw).
- On HOCWTA, DUSTED improved the PTPRC-CD45 Pearson correlation by 0.0264 (8.45%) and Spearman by 0.0426 (11.31%) over MAGIC, the second-best method.
- On DLPFC, median ARI 0.515, NMI 0.685 and HS 0.665 after DUSTED denoising, with ablation showing GCA raising mean ARI from 0.4475 to 0.5125.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DUSTED instead of DCA | DUSTED · Simulated SRT setting 1 | Avg MSE: 0.3925 -> 0.3839 | Helped |
| Use DUSTED instead of DCA | DUSTED · Simulated SRT setting 2 | Avg MSE: 0.3916 -> 0.3840 | Helped |
| Use DUSTED instead of DCA | DUSTED · Simulated SRT setting 3 | Avg MSE: 0.3929 -> 0.3842 | Helped |
| Use DUSTED instead of DCA | DUSTED · Simulated SRT setting 4 | Avg MSE: 0.4054 -> 0.3929 | Helped |
| Add Gene Channel Attention (GCA) (ablation) | DUSTED · Simulated datasets | Avg MSE: 0.3904 ± 0.0087 -> 0.3865 ± 0.0070 | Helped |
| Add Gene Channel Attention (GCA) (ablation) | DUSTED · HOCWTA | Pearson correlation: 0.30096 -> 0.33852 | Helped |
| Add Gene Channel Attention (GCA) (ablation) | DUSTED · HOCWTA | Spearman correlation: 0.40337 -> 0.42193 | Helped |
| Add Gene Channel Attention (GCA) (ablation) | DUSTED · DLPFC | Avg ARI: 0.4475 ± 0.0965 -> 0.5125 ± 0.0859 | Helped |
| Add Gene Channel Attention (GCA) (ablation) | DUSTED · DLPFC | Avg NMI: 0.6158 ± 0.0665 -> 0.6542 ± 0.0692 | Helped |
| Add Gene Channel Attention (GCA) (ablation) | DUSTED · DLPFC | Avg HS: 0.6158 ± 0.0665 -> 0.6542 ± 0.0692 | Helped |
Surprising findings
- Differences among the top denoisers on simulated data are small (DUSTED 0.3839 vs DCA 0.3925), with error bars overlapping.
Limitations
- Simulated benchmark relies on generated data whose realism is uncertain.
- Performance gains on simulated data over DCA are marginal.
- Relies on hyperparameter choices (residual weight alpha, loss choice) tuned per dataset.
Experimental setup
Datasets: SymSim-generated scRNA-seq simulations (scCube spatial assignment) HOCWTA (10x Visium ovarian cancer) DLPFC (12 sections)
Benchmarks: simulated SRT datasets (24) HOCWTA DLPFC
Metrics: MSE Pearson correlation Spearman correlation Global Moran's I ARI NMI Homogeneity score
Baselines: DCA MAGIC STAGATE Smoother Raw
Builds on: STAGATE DCA Graph Attention Networks CBAM
Closest papers by meaning
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.87)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.86)
- STINR: Deciphering Spatial Transcriptomics via Implicit Neural Representation (CVPR 2025 · similarity 0.86)
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.86)
- HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis (NeurIPS 2024 · similarity 0.86)
- Global Context-aware Representation Learning for Spatially Resolved Transcriptomics (ICML 2025 · similarity 0.85)
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.