DiffBulk: Enhancing Spatial Transcriptomic Prediction with Diffusion-Based Training.

Area of Science:

• Computational Biology
• Genomics
• Digital Pathology

Background:

• Spatial Transcriptomics (ST) is crucial for cancer diagnosis and precision medicine but is limited by high costs.
• Deep learning models can predict gene expression from histopathology images, but the complex morphology-gene relationship remains a challenge.

Purpose of the Study:

• To develop a novel framework, DiffBulk, for accurate gene expression prediction from histopathology images.
• To leverage conditional diffusion models for enhanced gene-image representation learning.

Main Methods:

• A two-stage framework utilizing conditional diffusion models.
• A permutation-invariant open-embedding gene encoder for unified gene panel training.
• Fusion of diffusion-derived features with pathology foundation model representations.

Main Results:

• DiffBulk consistently outperformed state-of-the-art baselines in gene expression prediction.
• The framework demonstrated effective bridging of the domain gap between image and gene expression data.
• Evaluation on Xenium ST data from HEST and CrunchDAO challenge confirmed robust performance.

Conclusions:

• DiffBulk offers a promising approach for diffusion-based gene-image representation learning.
• The findings suggest potential for cost-effective gene expression prediction in cancer research and clinical applications.
• This work opens new avenues for integrating imaging and genomic data in precision medicine.