HFFST: A Hierarchical Feature Fusion Algorithm for Spatial Gene Expression Prediction Using Histopathology Images

Area of Science:

• Computational biology
• Digital pathology
• Genomics

Background:

• Spatial transcriptomics advances disease understanding and personalized medicine but faces cost and complexity barriers.
• Predicting spatial gene expression from H&E images offers a promising, more accessible alternative.
• Existing methods often overlook the hierarchical nature of pathological image information.

Purpose of the Study:

• To develop and validate a novel hierarchical feature fusion algorithm (HFFST) for predicting spatial gene expression from H&E-stained pathology images.
• To improve the accuracy and utility of spatial transcriptomic profile prediction using histopathology data.
• To address limitations of current methods by exploiting multi-level image features.

Main Methods:

• Developed HFFST, a hierarchical feature fusion algorithm integrating multi-level feature extraction from whole-slide images.
• Employed a coarse-to-fine regression framework for predicting spatial transcriptomic profiles.
• Validated HFFST through cross-validation on five public datasets and external validation with high-resolution Visium data.

Main Results:

• HFFST demonstrated strong performance in predicting spatial gene expression profiles.
• The algorithm successfully identified distinct spatial regions within pathology images.
• HFFST showed advantages over existing state-of-the-art methods in predicting spatial transcriptomic data.

Conclusions:

• HFFST offers a promising approach for predicting spatial gene expression from H&E images, overcoming limitations of current techniques.
• The method enhances the potential of integrating digital pathology with transcriptomic data for biomedical research.
• HFFST provides a valuable tool for advancing disease mechanism studies and therapeutic target discovery.