Research map / Image generation / Autoregressive and latent generation
GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow
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 GeneFlow, a rectified-flow conditional UNet with an attention-based RNA encoder that generates H&E/DAPI histology tiles from single- and multi-cell spatial transcriptomics, outperforming a diffusion baseline.
Problem
Generating realistic histopathology images from single- and multi-cell gene expression (the inverse of gene expression prediction from histology), which has been largely unexplored.
Approach
Condition a UNet with rectified flow on an attention-based RNA encoder (low-rank gene-gene factorization, gene attention, multi-head cell attention) to map transcriptomes to paired histology images, with high-order ODE solving at inference.
Key results
- Rectified flow single-cell FID 20.73 on Xenium C1 vs diffusion 171.06, with FID 3-6x lower across datasets.
- Pathologists preferred rectified flow images over diffusion in 86% of cases.
- On C1 with UNI2-h metrics, single-cell FID 39.27 for rectified flow vs 405.98 for diffusion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use rectified flow instead of diffusion | GeneFlow image generation · Xenium datasets, single- and multi-cell | FID: 213.60$\pm$17.20 -> 34.31$\pm$5.65; 132.09$\pm$57.05 -> 27.43$\pm$5.91; 189.08$\pm$13.40 -> 47.95$\pm$7.38; 171.06$\pm$8 | Helped |
| Replace UNet (MSE) with rectified flow | Xenium_C1 single-cell image generation · Xenium_C1 single-cell | FID Overall UNI2-h: 525.51 -> 39.27 | Helped |
| Remove gene attention (ablation) | Rectified Flow RNA encoder · Xenium_C1 single-cell | FID Overall UNI2-h: 39.27 -> 26.46 | Helped |
| Remove gene attention (ablation) | Rectified Flow RNA encoder · Xenium_C1 single-cell | Inception Feat. Dist.: 14.50±2.52 -> 14.89±2.52 | Hurt |
| Remove gene attention (ablation) | Rectified Flow RNA encoder · Xenium_C1 single-cell | UNI2-h Embedding Sim.: 0.983±0.003 -> 0.991±0.003 | Helped |
| Remove gene attention and gene-relation module (ablation) | Rectified Flow RNA encoder · Xenium_C1 single-cell | FID Overall UNI2-h: 39.27 -> 27.25 | Helped |
| Remove gene attention and multi-head cell attention (ablation) | Rectified Flow RNA encoder · Xenium_C1 single-cell | FID Overall UNI2-h: 39.27 -> 29.50 | Helped |
| Remove gene attention and multi-head cell attention (ablation) | Rectified Flow RNA encoder · Xenium_C1 single-cell | Inception Feat. Dist.: 14.50±2.52 -> 15.39±2.68 | Hurt |
| Use the transformer-based RNA encoder instead of a simple linear encoder (ablation) | Rectified Flow · Xenium C1, C2, and P1 | FID and Feature Distance | Helped |
| Train on melanoma-only rather than all cell types (ablation) | GeneFlow · Xenium melanoma cells | FID | Mixed |
| Train on non-melanoma-only rather than all cell types (ablation) | GeneFlow · Xenium non-melanoma cells | image generation performance | Hurt |
| Condition on multiple cells rather than a single cell (ablation) | GeneFlow · Xenium melanoma and non-melanoma evaluations | SSIM | Helped |
Surprising findings
- Models trained only on melanoma cells performed comparably to or better than models trained on all cell types.
- Diffusion showed slightly better SSIM than rectified flow on some samples (C2, P1).
Limitations
- Operates only on 256x256 tiles; no whole-slide synthesis.
- Relies on Xenium single-cell resolution data, limiting compatibility with other ST platforms.
- Cell inclusion in multi-cell patches is binary rather than weighted by overlap.
Experimental setup
Datasets: Xenium C1 melanoma Xenium C2 melanoma Xenium P1 melanoma HEST-1k
Benchmarks: Xenium P1 Xenium C1 Xenium C2 HEST-1k
Metrics: FID SSIM Feature Distance UNI2-h FID nuclear morphometric similarity pathologist similarity and preference
Baselines: Diffusion (transformer encoder) conditional UNet (MSE) simple RNA encoder ablation
Builds on: rectified flow diffusion models UNet AdamW Enrichr
Closest papers by meaning
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.92)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.90)
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · similarity 0.89)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.87)
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.87)
- Modeling Microenvironment Trajectories on Spatial Transcriptomics with NicheFlow (NeurIPS 2025 · similarity 0.87)
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.