HiST: Histological Images Reconstruct Tumor Spatial Transcriptomics via MultiScale Fusion Deep Learning

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Spatial transcriptomics (ST) offers insights into the tumor microenvironment but is limited by high costs.
• Integrating spatial molecular data with histological context is crucial for cancer research.

Purpose of the Study:

• To develop a cost-effective deep learning framework (HiST) for reconstructing spatial gene expression profiles (GEPs) from histological images.
• To enhance downstream clinical analyses, including tumor heterogeneity assessment and patient stratification.

Main Methods:

• A multi-scale convolutional deep learning framework, HiST, was developed to learn the relationship between GEPs and histological morphology.
• HiST utilizes ST data to train the model for predicting spatial GEPs from standard histology slides.

Main Results:

• HiST accurately predicts tumor regions across multiple cancer types (AUC: 0.96) and reconstructs spatial GEPs from histology images with high fidelity (avg. PCC: 0.74).
• The reconstructed GEPs enable robust tumor heterogeneity assessment, identification of tumor subtypes, and stratification of patient prognosis (e.g., breast cancer CI: 0.78).
• HiST outperforms existing models by approximately two-fold in GEP reconstruction and facilitates immunotherapy response prediction.

Conclusions:

• HiST provides a reliable and cost-effective molecular representation from histological images, significantly advancing spatial transcriptomics applications.
• This framework enhances tumor profiling, biomarker discovery, and clinical decision-making in oncology.