Research map / Label-efficient and robust learning / Recognition and long-tail
Accurate Spatial Gene Expression Prediction by Integrating Multi-Resolution Features
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 TRIPLEX, which predicts spatial gene expression from WSIs by fusing target-spot, neighbor-view, and global-view features with a fusion loss, evaluated via patient-disjoint cross-validation on three ST datasets and external Visium data.
Problem
Predicting spatial transcriptomics gene expression from histology whole slide images, where prior methods use only spot images or spot-level interactions and often evaluate with patient replicates leaking across train and test splits.
Approach
Jointly encodes three resolutions (the target spot at 224x224, a 1120x1120 neighbor view split into 25 tokens, and all spots in the WSI as global tokens via APEG-augmented transformer), fuses them with cross-attention on global queries, and applies a distillation-style fusion loss from the fused token to the individual token predictors.
Key results
- On the SCC dataset, TRIPLEX improves PCC(M) by 0.093 and PCC(H) by 0.102 over EGN in leave-one-patient-out style cross-validation.
- On BC1, TRIPLEX improves PCC(M) by 0.117 and PCC(H) by 0.169 over EGN.
- On external Visium data, TRIPLEX attains average PCC(H) of 0.293 and outperforms ST-Net, EGN and BLEEP across MSE, MAE, PCC(M) and PCC(H).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Integrate target, neighbor, and global features with TRIPLEX | Spatial gene expression prediction · BC1 | PCC(M): 0.197 -> 0.314 | Helped |
| Integrate target, neighbor, and global features with TRIPLEX | Spatial gene expression prediction · BC2 | PCC(M): 0.111 -> 0.206 | Helped |
| Integrate target, neighbor, and global features with TRIPLEX | Spatial gene expression prediction · SCC | PCC(M): 0.281 -> 0.374 | Helped |
| Use TRIPLEX instead of the local baseline models | Visium gene expression prediction · 10X Visium-1 | PCC(M): 0.106 -> 0.136 | Helped |
| Use TRIPLEX instead of the local baseline models | Visium gene expression prediction · 10X Visium-2 | PCC(M): 0.104 -> 0.155 | Helped |
| Use TRIPLEX instead of the local baseline models | Visium gene expression prediction · 10X Visium-3 | PCC(M): 0.114 -> 0.118 | Helped |
| Add the target encoding module (TEM) (ablation) | TRIPLEX · SCC | PCC(M): 0.352 -> 0.374 | Helped |
| Add the neighbor encoding module (NEM) (ablation) | TRIPLEX · SCC | PCC(M): 0.330 -> 0.374 | Helped |
| Add the global encoding module (GEM) (ablation) | TRIPLEX · SCC | PCC(M): 0.358 -> 0.374 | Helped |
| Use APEG instead of no positional encoding (ablation) | TRIPLEX · SCC | PCC(M): 0.360 -> 0.374 | Helped |
| Use APEG instead of conventional PEG (ablation) | TRIPLEX · SCC | PCC(M): 0.364 -> 0.374 | Helped |
| Use the fusion layer instead of summation (ablation) | TRIPLEX · SCC | PCC(M): 0.341 -> 0.374 | Helped |
Surprising findings
- Hist2ST achieves near-zero PCC(M) (0.004) on the SCC dataset with MSE 1.291, suggesting instability of global-context models under patient-disjoint evaluation.
- Many prior reported results are not reproduced under the authors' stricter protocol; the paper attributes differences to cross-validation strategy, normalization and metric calculation.
Limitations
- Small ST datasets (12 to 68 samples) limit generalization assessment.
- Performance on Visium external validation is modest (PCC(H) average 0.293).
Experimental setup
Datasets: BC1 breast cancer ST dataset BC2 breast cancer ST dataset (ST-Net) SCC skin cancer ST dataset 10x Genomics Visium breast cancer samples
Benchmarks: BC1 BC2 SCC 10X Visium-1 10X Visium-2 10X Visium-3
Metrics: MSE MAE PCC(M) PCC(H)
Baselines: ST-Net EGN BLEEP HisToGene Hist2ST TEM NEM GEM
Builds on: ST-Net HisToGene Hist2ST EGN BLEEP HIPT PEG (conditional position encoding)
Closest papers by meaning
- MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images (CVPR 2025 · similarity 0.87)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.86)
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.85)
- M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images (AAAI 2025 · similarity 0.85)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.85)
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · 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.