Research map / Image generation / Autoregressive and latent generation
Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching
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 STFlow, a whole-slide flow matching model with an E(2)-invariant local spatial attention denoiser that predicts spatial gene expression from H&E histology, outperforming spot- and slide-based baselines on HEST-1k and STImage-1K4M.
Problem
Predicting spatially resolved gene expression from whole-slide H&E images, where prior methods predict each spot independently (ignoring cell-cell interaction) and slide-level encoders are memory-heavy given >10,000 spots per slide.
Approach
Reformulate spot-level gene expression regression as whole-slide flow matching starting from a ZINB prior, iteratively refining joint expression predictions, with a frame-averaging E(2)-invariant transformer that attends only to k-nearest spatial neighbors and uses expression differences in attention.
Key results
- STFlow averages 0.415 Pearson on HEST-Bench vs 0.395 for the best baseline (TRIPLEX) and 0.344 for UNI, about 18% relative improvement over pathology foundation models.
- On STImage-Bench, STFlow averages 0.293 Pearson vs 0.269 for the best baseline (Gigapath).
- Removing flow matching (w/o FM) reduces UNI-based average from 0.415 to 0.405 on HEST-Bench.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use STFlow instead of TRIPLEX | UNI-based gene expression prediction · HEST-1k average | Pearson correlation: 0.395 -> 0.415 | Helped |
| use STFlow instead of Gigapath | gene expression prediction · STImage-1K4M average | Pearson correlation: 0.269 -> 0.293 | Helped |
| add flow matching and iterative refinement (ablation) | STFlow with UNI encoder · GATA3, ERBB2, UBE2C, and VWF biomarker prediction | Pearson correlation: 0.842_{.005} -> 0.860_{.005}; 0.822_{.002} -> 0.844_{.001}; 0.750_{.008} -> 0.772_{.008}; 0.570_{.001} -> 0.666_{.004} | Helped |
| add flow matching (ablation) | STFlow with UNI encoder · HEST-1k average | Pearson correlation: 0.405 -> 0.415 | Helped |
| add flow matching (ablation) | STFlow with Ciga encoder · HEST-1k average | Pearson correlation: 0.351 -> 0.361 | Helped |
| add frame-averaging transformations (ablation) | STFlow with UNI encoder · HEST-1k average | Pearson correlation: 0.406 -> 0.415 | Helped |
| add frame-averaging transformations (ablation) | STFlow with Ciga encoder · HEST-1k average | Pearson correlation: 0.348 -> 0.361 | Helped |
| use 2 refinement steps instead of 1 (ablation) | STFlow · HEST-1k average | Pearson correlation: 0.405 -> 0.417 | Helped |
| increase refinement steps from 5 to 10 (ablation) | STFlow · HEST-1k average | Pearson correlation: 0.415 -> 0.412 | Hurt |
| replace the frame-averaging Transformer with EGNN (ablation) | STFlow with UNI encoder · HEST-1k average | Pearson correlation: 0.415 -> 0.400 | Hurt |
| replace the frame-averaging Transformer with E2CNN (ablation) | STFlow with UNI encoder · HEST-1k average | Pearson correlation: 0.415 -> 0.363 | Hurt |
| replace ZINB prior with Gaussian prior (ablation) | STFlow with UNI encoder · HEST-1k average | Pearson correlation: 0.407 -> 0.415 | Hurt |
Surprising findings
- Gigapath-slide, which aggregates whole-slide information, does not outperform Gigapath on most tasks, possibly because its pretraining targets slide-level tasks.
- Patient-level splits make spot-based methods such as STNet and HisToGene underperform the UNI foundation model with a linear head.
Limitations
- ZINB prior hyperparameters are chosen by grid search rather than estimated from data.
- Gains over baselines are modest on some datasets (e.g., Skin and Mouth on STImage-Bench where STFlow is not best), and the approach relies on frozen pathology foundation models.
Experimental setup
Datasets: HEST-1k STImage-1K4M
Benchmarks: HEST-Bench STImage-Bench HEST-1k biomarker prediction (GATA3, ERBB2, UBE2C, VWF)
Metrics: Pearson correlation inference time GPU memory
Baselines: Ciga UNI Gigapath STNet BLEEP Gigapath-slide HisToGene TRIPLEX EGNN E2CNN
Builds on: flow matching frame averaging EGNN TRIPLEX BLEEP
Closest papers by meaning
- GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow (NeurIPS 2025 · similarity 0.92)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.92)
- HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis (NeurIPS 2024 · similarity 0.90)
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · similarity 0.90)
- MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images (CVPR 2025 · similarity 0.89)
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · 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.