Research map / Image generation / Autoregressive and latent generation
M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology 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 M2OST, a many-to-one Transformer that jointly predicts a spot's spatial gene expression from three pyramid levels of the same WSI patch, with decoupled intra-level, cross-level token and channel mixing.
Problem
Predicting spatial transcriptomics (ST) gene expression from histopathology whole slide images; existing patch-level regression methods ignore multi-scale pyramid information and inter-spot visual context, and slide-level methods are large and costly.
Approach
Treat ST prediction as many-to-one regression: multiple magnification levels of the same spot are fed through weight-decoupled streams with deformable patch embedding, cross-level attention and cross-level channel mixing, yielding a compact Transformer with fewer parameters and FLOPs.
Key results
- M2OST reaches mean PCC of 48.07% on HBC, 44.17% on HER2+ and 50.50% on cSCC, the best among compared methods.
- M2OST uses 6.81M parameters and 2.24 GFLOPs, versus 7.21M and 2.87 GFLOPs for ST-Net, and fewer than slide-level Hist2ST (675.5M parameters).
- Masking one random level in 75% of samples drops PCC by 7.75%, 7.65% and 3.14% on the three datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DPE instead of ordinary patch embedding (ablation) | M2OST · HBC, HER2+, and cSCC | PCC (%): 47.13 -> 48.07; 43.10 -> 44.17; 49.35 -> 50.50 | Helped |
| Use separate ITMMs rather than unified self-attention (ablation) | M2OST · HBC, HER2+, and cSCC | PCC (%): 47.03 -> 48.07; 42.99 -> 44.17; 49.48 -> 50.50 | Helped |
| Add CTMM instead of simple concatenation for cross-level fusion (ablation) | M2OST · HBC, HER2+, and cSCC | PCC (%): 46.34 -> 48.07; 42.57 -> 44.17; 48.81 -> 50.50 | Helped |
| Use CCMM instead of ordinary fully connected layers for channel mixing (ablation) | M2OST · HBC, HER2+, and cSCC | PCC (%): 46.12 -> 48.07; 42.55 -> 44.17; 48.66 -> 50.50 | Helped |
| Add level 1 and level 2 image patches to level 0 inputs (ablation) | M2OST · HBC, HER2+, and cSCC | PCC (%): 46.92 -> 48.07; 43.12 -> 44.17; 49.31 -> 50.50 | Helped |
| Mask one random input level in more samples (ablation) | M2OST on the fused dataset · HBC, HER2+, and cSCC | PCC (%): 39.46 -> 31.71; 39.97 -> 32.32; 37.33 -> 34.19 | Hurt |
| Apply many-to-one modeling with ResNet-50 patch embeddings (ablation) | ResNet-50 · HBC, HER2+, and cSCC | PCC (%): +1.71; +1.43; +1.46 | Helped |
| Apply many-to-one modeling with ST-Net patch embeddings (ablation) | ST-Net · HBC, HER2+, and cSCC | PCC (%): +1.12; +1.72; +1.39 | Helped |
| Apply many-to-one modeling with ViT patch embeddings (ablation) | ViT · HBC, HER2+, and cSCC | PCC (%): +1.95; +0.63; +1.81 | Helped |
Surprising findings
- Using only level 0 (one-to-one) M2OST does not surpass simple baselines such as ResNet-50 and ST-Net, so gains come from the many-to-one scheme.
- Slide-level methods (HisToGene, Hist2ST) underperform patch-level methods, suggesting inter-spot correlations add little for spot-level gene prediction.
Limitations
- As a spot-level method, it predicts one spot at a time and is less efficient than slide-level methods for generating full ST maps.
- End-to-end spot-level training perceives only limited nearby non-local context in a WSI.
- Performance drops notably when image levels are missing at test time.
Experimental setup
Datasets: HBC HER2+ cSCC
Benchmarks: HBC HER2+ cSCC
Metrics: PCC RMSE parameter count FLOPs
Baselines: ResNet50 ViT-B/16 Swin-T ConvNeXt-T CrossViT DeepSpaCE ST-Net HisToGene Hist2ST BLEEP HIPT/iStar
Builds on: ViT CrossViT Random Mask Self-Attention squeeze-and-excitation
Closest papers by meaning
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.92)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.89)
- Multi-modal Topology-embedded Graph Learning for Spatially Resolved Genes Prediction from Pathology Images with Prior Gene Similarity Information (CVPR 2025 · similarity 0.89)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.89)
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · similarity 0.89)
- MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images (CVPR 2025 · similarity 0.89)
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.