Research map / Graph learning / Knowledge and temporal graphs
Multi-modal Topology-embedded Graph Learning for Spatially Resolved Genes Prediction from Pathology Images with Prior Gene Similarity Information
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 M2TGLGO, which predicts spatially resolved gene expression from H&E pathology images by fusing deep and handcrafted nuclear features in a GAT-based graph, adding a spatial neighborhood ranking loss, and using a GO-similarity gene graph for multi-gene prediction.
Problem
Predicting spatial transcriptomics gene expression from H&E pathology images is limited by single-type spot features, over-smoothing in GNNs, and treating each gene independently, ignoring gene-gene functional relationships.
Approach
Co-learns spot representations from deep (ResNet-101) and handcrafted nuclear morphology/topology features via within-modal GAT and inter-modal attention, adds a quartile-based spatial neighborhood ranking loss to counter over-smoothing, and predicts multiple genes with a GAT over a Gene Ontology similarity graph initialized with Gene2Vec.
Key results
- On DLPFC, M2TGLGO reaches HVG PCC 0.291 vs 0.252 for the best baseline BLEEP among top-50 HVG prediction.
- Across three datasets, M2TGLGO achieves HVG PCC of 0.291, 0.296, and 0.308 (DLPFC, ccRCC, BC), outperforming all compared ST-based methods.
- Ablation shows removing components lowers PCC; the full model reaches HEG PCC 0.217 on DLPFC versus 0.187 for the best partial variant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine multi-modal image features, spatial topology ranking, and GO-guided gene prediction | M2TGLGO · DLPFC HVG | PCC: 0.252 -> 0.291 | Helped |
| Combine multi-modal image features, spatial topology ranking, and GO-guided gene prediction | M2TGLGO · DLPFC HEG | PCC: 0.186 -> 0.217 | Helped |
| Combine multi-modal image features, spatial topology ranking, and GO-guided gene prediction | M2TGLGO · ccRCC HVG | PCC: 0.263 -> 0.296 | Helped |
| Combine multi-modal image features, spatial topology ranking, and GO-guided gene prediction | M2TGLGO · ccRCC HEG | PCC: 0.209 -> 0.256 | Helped |
| Combine multi-modal image features, spatial topology ranking, and GO-guided gene prediction | M2TGLGO · BC HVG | PCC: 0.254 -> 0.308 | Helped |
| Combine multi-modal image features, spatial topology ranking, and GO-guided gene prediction | M2TGLGO · BC HEG | PCC: 0.223 -> 0.269 | Helped |
| Add inter-modal embeddings (ablation) | M2TGLGO · DLPFC, ccRCC, and BC (HVG and HEG) | PCC: 0.090 -> 0.291; 0.073 -> 0.217; 0.215 -> 0.296; 0.183 -> 0.256; 0.162 -> 0.308; 0.115 -> 0.269 | Helped |
| Add within-modal embeddings (ablation) | M2TGLGO · DLPFC, ccRCC, and BC (HVG and HEG) | PCC: 0.236 -> 0.291; 0.182 -> 0.217; 0.229 -> 0.296; 0.204 -> 0.256; 0.260 -> 0.308; 0.213 -> 0.269 | Helped |
| Add SONRM (ablation) | M2TGLGO · DLPFC, ccRCC, and BC (HVG and HEG) | PCC: 0.252 -> 0.291; 0.187 -> 0.217; 0.258 -> 0.296; 0.222 -> 0.256; 0.261 -> 0.308; 0.232 -> 0.269 | Helped |
| Add GO prior knowledge (ablation) | M2TGLGO · DLPFC, ccRCC, and BC (HVG and HEG) | PCC: 0.163 -> 0.291; 0.102 -> 0.217; 0.216 -> 0.296; 0.160 -> 0.256; 0.117 -> 0.308; 0.223 -> 0.269 | Helped |
| Vary γ from 0 to 1 (ablation) | M2TGLGO · DLPFC, ccRCC, and BC | PCC | Helped |
| Vary SONRM quantile value from 3 to 6 (ablation) | M2TGLGO · DLPFC, ccRCC, and BC (HVG and HEG) | PCC: 0.285, 0.210, 0.298, 0.231, 0.300, 0.257; 0.291, 0.217, 0.296, 0.256, 0.308, 0.269; 0.289, 0.204, 0.281, 0.240, 0.295, 0 | No clear effect |
Limitations
- Evaluated only on top-50 HVG and HEG genes rather than the full transcriptome.
- Reported gains are modest in absolute PCC (around 0.29-0.31) and per-slide variance is large.
- Relies on a fixed KNN graph (k=6) and hand-set hyperparameters such as gamma and quantile count.
Experimental setup
Datasets: DLPFC ccRCC BC
Benchmarks: DLPFC ccRCC BC
Metrics: Pearson correlation coefficient mean squared error
Baselines: Hist2ST HisToGene THItoGene ST-Net BLEEP TRIPLEX TG-GATEs TSIEN GECMC MGNN TMEGL SI-MIL
Builds on: graph attention networks HoVer-Net Gene2Vec BLEEP
Closest papers by meaning
- MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images (CVPR 2025 · similarity 0.91)
- M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images (AAAI 2025 · similarity 0.89)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.88)
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.88)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.88)
- Global Context-aware Representation Learning for Spatially Resolved Transcriptomics (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.