Research map / Graph learning / Knowledge and temporal graphs
MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images
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 MERGE, a GAT-based GNN over a multi-faceted hierarchical graph (spatial and feature-space clustering with intra- and inter-cluster edges) plus SPCS gene-aware smoothing, to predict spatial gene expression from whole slide histology images.
Problem
Predicting spatially resolved gene expression from histology image patches is hard because existing methods do not model long-range interactions between morphologically similar but spatially distant tissue locations, and ST data is noisy and sparse with dropouts.
Approach
Cluster spots both spatially and in patch-embedding feature space, connect each spot to its cluster centroid (internal edges), and fully connect centroids across both clusterings (shortcut edges), so any two nodes are within three hops; combine with gene-aware SPCS smoothing of targets.
Key results
- On ST-Net, MERGE (hier) reaches PCC 0.6795 vs TRIPLEX 0.2320 and MSE 0.1347 vs 0.1472 in eight-fold cross-validation.
- On Her2ST, MERGE achieves PCC 0.5037 versus TRIPLEX 0.3927 with MSE 0.6422 vs 0.8982.
- Replacing 8n with SPCS smoothing improves MERGE PCC from 0.6645 to 0.679 on ST-Net; on SCC MERGE is outperformed by TRIPLEX in MSE (0.4891) while MERGE gets 0.5512 PCC vs TRIPLEX 0.5416.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the multi-faceted hierarchical graph instead of all 1-hop neighbors (ablation) | MERGE on ST-Net · ST-Net | PCC: 0.5789 -> 0.6795 | Helped |
| add spatial clustering to feature-space clustering (ablation) | MERGE on ST-Net · ST-Net | PCC: 0.6356 -> 0.6795 | Helped |
| add feature-space clustering to spatial clustering (ablation) | MERGE on ST-Net · ST-Net | PCC: 0.6719 -> 0.6795 | Helped |
| select the feature-centroid spot instead of a random spot as cluster representative (ablation) | MERGE on ST-Net · ST-Net | PCC: 0.666 -> 0.6795 | Helped |
| select the feature-centroid spot instead of the spatial-centroid spot as cluster representative (ablation) | MERGE on ST-Net · ST-Net | PCC: 0.6583 -> 0.6795 | Helped |
| increase cluster size from 25 to 100 (ablation) | MERGE on ST-Net · ST-Net | PCC: 0.6754 -> 0.6795 | Helped |
| increase cluster size beyond 100 to 150 (ablation) | MERGE on ST-Net · ST-Net | PCC: 0.6795 -> 0.673 | Hurt |
| replace SPCS smoothing with 8n smoothing (ablation) | MERGE on ST-Net · ST-Net | PCC: 0.679 -> 0.6645 | Hurt |
| use the hierarchical graph rather than the baseline methods | MERGE · ST-Net | PCC: 0.5221 -> 0.6795 | Helped |
| use the hierarchical graph rather than the baseline methods | MERGE · Her2ST | PCC: 0.4629 -> 0.5037 | Helped |
| use the hierarchical graph rather than the baseline methods | MERGE · SCC | PCC: 0.5416 -> 0.5512 | Helped |
| use SPCS smoothing instead of 8n smoothing (ablation) | ResNet+FCN on ST-Net · ST-Net | PCC: 0.2701 -> 0.5221 | Helped |
Surprising findings
- TRIPLEX performs better than MERGE on SCC in MSE (0.4891 vs 0.5353), and HisToGene attains the lowest MAE on SCC.
- Using 8n smoothing instead of SPCS actually helps TRIPLEX's MSE (0.0760 vs 0.1472 with SPCS), contrary to the other methods.
Limitations
- Cluster size is a sensitive hyperparameter; performance drops for clusters larger than 100 spots on ST-Net.
- Does not outperform all baselines on every metric (e.g., SCC MSE and MAE).
Experimental setup
Datasets: ST-Net Her2ST SCC TCGA-BRCA
Benchmarks: ST-Net Her2ST SCC
Metrics: MSE MAE PCC
Baselines: ResNet+FCN BLEEP HisToGene Hist2ST THItoGene TRIPLEX
Builds on: Graph Attention Network SPCS TRIPLEX BLEEP SimCLR
Closest papers by meaning
- Multi-modal Topology-embedded Graph Learning for Spatially Resolved Genes Prediction from Pathology Images with Prior Gene Similarity Information (CVPR 2025 · similarity 0.91)
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.89)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.89)
- M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images (AAAI 2025 · similarity 0.89)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.88)
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.88)
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.