Research map / 3D vision / 3D generation
ASIGN: An Anatomy-aware Spatial Imputation Graphic Network for 3D Spatial 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.
Introduces 3D spatial transcriptomics imputation from 3D WSI sections plus one 2D ST slide, using a graph network (ASIGN) with cross-layer label propagation.
Problem
Full 3D spatial transcriptomics is prohibitively expensive, and existing WSI-to-ST predictors work only on single 2D sections and generalize poorly across samples.
Approach
Builds a 3D spot graph by registering WSI layers (XFeat + ANTs) and linking spots across layers via overlap and similarity; a multi-level spatial attention graph network (MSAGNet) fuses multi-resolution patches and GAT/Transformer layers, and a cross-layer imputation block propagates labels from one known ST layer to unknown layers.
Key results
- ASIGN 3D reaches PCC 0.693 on HER2 versus 0.649 for the best 3D baseline (similarity) and 0.441 for His2ST's 2D setting, per the reported table.
- On DLPFC, ASIGN 3D achieves PCC 0.739 vs 0.663 for the best 3D baseline; on ST-Data, ASIGN 2D is competitive with 2D baselines.
- Cross-layer imputation raises HER2 PCC from 0.635 to 0.693 with 20% known labels, with performance plateauing at 20-30%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ASIGN 2D instead of the strongest 2D baseline, EGN | 2D WSI-wise ST prediction · HER2 | PCC: 0.442 -> 0.476 | Helped |
| use ASIGN 2D instead of the strongest 2D baseline, His2ST | 2D WSI-wise ST prediction · DLPFC | PCC: 0.764 -> 0.766 | Helped |
| use ASIGN 2D instead of the strongest 2D baseline, EGN | 2D WSI-wise ST prediction · ST-Data | PCC: 0.554 -> 0.552 | No clear effect |
| use ASIGN 3D instead of the strongest 3D baseline, Similarity | 3D sample-wise ST prediction · HER2 | PCC: 0.649 -> 0.693 | Helped |
| use ASIGN 3D instead of the strongest 3D baseline, Similarity | 3D sample-wise ST prediction · DLPFC | PCC: 0.762 -> 0.773 | Helped |
| use ASIGN 3D instead of the strongest 3D baseline, Similarity | 3D sample-wise ST prediction · ST-Data | PCC: 0.696 -> 0.739 | Helped |
| add cross-attention and GAT blocks (ablation) | ASIGN · HER2 | PCC: 0.659 -> 0.693 | Helped |
| add cross-attention (ablation) | ASIGN · HER2 | PCC: 0.686 -> 0.693 | Helped |
| add GAT blocks (ablation) | ASIGN · HER2 | PCC: 0.677 -> 0.693 | Helped |
| add CLI label propagation with 20% known labels instead of no CLI (ablation) | ASIGN · HER2 | PCC: 0.635 -> 0.693 | Helped |
| increase known-label proportion from 20% to 30% (ablation) | ASIGN CLI block · HER2 | PCC: 0.693 -> 0.708 | Helped |
| increase known-label proportion from 20% to 40% (ablation) | ASIGN CLI block · HER2 | PCC: 0.693 -> 0.703 | Helped |
Surprising findings
- Adding the 3D sample-level graph raised PCC on HER2 from 0.476 to 0.635, showing that spatial context across layers matters more than 2D modeling alone.
Limitations
- Evaluated on only three public datasets with small sample counts (3-23 samples).
- Requires one ST-profiled layer per sample as known labels; the 3D setting depends on accurate registration of WSI layers.
Experimental setup
Datasets: HER2 DLPFC ST-Data
Benchmarks: HER2 DLPFC ST-Data
Metrics: PCC MSE MAE
Baselines: ST-Net EGN HisToGene BLEEP His2ST PASTE-inspired overlap imputation PASTE-inspired similarity imputation
Builds on: TRIPLEX PASTE ST-Net BLEEP
Closest papers by meaning
- HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis (NeurIPS 2024 · similarity 0.87)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.87)
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.86)
- STINR: Deciphering Spatial Transcriptomics via Implicit Neural Representation (CVPR 2025 · similarity 0.86)
- M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images (AAAI 2025 · similarity 0.86)
- MEATRD: Multimodal Anomalous Tissue Region Detection Enhanced with Spatial Transcriptomics (AAAI 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.