Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow

Mengbo Wang, Shourya Verma, Aditya Malusare, Luopin Wang, Yiyang Lu, Vaneet Aggarwal, Mario Sola, Ananth Grama, Nadia Atallah Lanman

TL;DR

GeneFlow tackles the inverse problem of translating spatial transcriptomics into histopathology images by leveraging rectified flow to learn a continuous, bijective mapping from gene expression to image space. An attention-rich RNA encoder with two cell-context pathways conditions a UNet-based generator to synthesize high-resolution H&E/DAPI images, with a high-order ODE solver enabling stable generation. Across Melanoma Xenium datasets, GeneFlow outperforms diffusion baselines on FID, SSIM, and feature-distance metrics, and expert pathologists favor its outputs while confirming diagnostic coherence. The work enables visualizing transcriptomic perturbations and intercellular interactions, with potential for perturbation-guided diagnostics and biomarker discovery, while highlighting avenues for scaling to whole-slide images and broader gene panels.

Abstract

Spatial transcriptomics (ST) technologies can be used to align transcriptomes with histopathological morphology, presenting exciting new opportunities for biomolecular discovery. Using ST data, we construct a novel framework, GeneFlow, to map transcriptomics onto paired cellular images. By combining an attention-based RNA encoder with a conditional UNet guided by rectified flow, we generate high-resolution images with different staining methods (e.g. H&E, DAPI) to highlight various cellular/tissue structures. Rectified flow with high-order ODE solvers creates a continuous, bijective mapping between transcriptomics and image manifolds, addressing the many-to-one relationship inherent in this problem. Our method enables the generation of realistic cellular morphology features and spatially resolved intercellular interactions from observational gene expression profiles, provides potential to incorporate genetic/chemical perturbations, and enables disease diagnosis by revealing dysregulated patterns in imaging phenotypes. Our rectified flow-based method outperforms diffusion-based baseline method in all experiments. Code can be found at https://github.com/wangmengbo/GeneFlow.

GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow

TL;DR

GeneFlow tackles the inverse problem of translating spatial transcriptomics into histopathology images by leveraging rectified flow to learn a continuous, bijective mapping from gene expression to image space. An attention-rich RNA encoder with two cell-context pathways conditions a UNet-based generator to synthesize high-resolution H&E/DAPI images, with a high-order ODE solver enabling stable generation. Across Melanoma Xenium datasets, GeneFlow outperforms diffusion baselines on FID, SSIM, and feature-distance metrics, and expert pathologists favor its outputs while confirming diagnostic coherence. The work enables visualizing transcriptomic perturbations and intercellular interactions, with potential for perturbation-guided diagnostics and biomarker discovery, while highlighting avenues for scaling to whole-slide images and broader gene panels.

Abstract

Spatial transcriptomics (ST) technologies can be used to align transcriptomes with histopathological morphology, presenting exciting new opportunities for biomolecular discovery. Using ST data, we construct a novel framework, GeneFlow, to map transcriptomics onto paired cellular images. By combining an attention-based RNA encoder with a conditional UNet guided by rectified flow, we generate high-resolution images with different staining methods (e.g. H&E, DAPI) to highlight various cellular/tissue structures. Rectified flow with high-order ODE solvers creates a continuous, bijective mapping between transcriptomics and image manifolds, addressing the many-to-one relationship inherent in this problem. Our method enables the generation of realistic cellular morphology features and spatially resolved intercellular interactions from observational gene expression profiles, provides potential to incorporate genetic/chemical perturbations, and enables disease diagnosis by revealing dysregulated patterns in imaging phenotypes. Our rectified flow-based method outperforms diffusion-based baseline method in all experiments. Code can be found at https://github.com/wangmengbo/GeneFlow.

Paper Structure

This paper contains 59 sections, 57 equations, 14 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. GeneFlow Architecture
  5. Gene Relation Network:
  6. Global Gene Attention:
  7. Multi-Head Cell Attention:
  8. UNet Architecture:
  9. Generative Modeling With Rectified Flow
  10. Results
  11. Datasets
  12. Quantitative Image Analysis
  13. Gene Importance Analysis
  14. Qualitative Image Analysis
  15. Discussion
  16. ...and 44 more sections

Figures (14)

  • Figure 1: Mapping from histopathology images to gene expression (forward problem) and vice versa (inverse problem) through GeneFlow.
  • Figure 2: Architecture of the GeneFlow model for mapping transcriptomes to histology images.
  • Figure 3: Evaluation by 3 ACVP board certified pathologists and 2 residents.
  • Figure 4: Rectified Flow model cross-dataset performance evaluation
  • Figure 4: Overlap significance of influential genes
  • ...and 9 more figures