HE2Gene: image-to-RNA translation via multi-task learning for spatial transcriptomics data.

Area of Science:

• Computational biology
• Genomics
• Pathology

Background:

• Spatial molecular profiling links tissue morphology and gene expression but is costly and time-consuming.
• Existing computational tools have limited gene prediction capabilities and are often designed for bulk RNA-seq.

Purpose of the Study:

• To introduce HE2Gene, a novel multi-task learning method for predicting gene expression and pathological annotations from H&E-stained images.
• To overcome the limitations of current spatial molecular profiling technologies in terms of cost, time, and analytical scope.

Main Methods:

• HE2Gene utilizes multi-task learning to predict tens of thousands of spot-level gene expressions.
• The method analyzes standard Hematoxylin and Eosin (H&E)-stained images.
• It integrates pathological annotations with gene expression predictions.

Main Results:

• HE2Gene achieves performance comparable to state-of-the-art methods.
• The model demonstrates strong generalization on external datasets without re-training.
• It successfully preserves annotated spatial domains and identifies potential biomarkers.

Conclusions:

• HE2Gene provides a powerful and accessible tool for spatial gene expression prediction from histology images.
• The method has significant potential for advancing cancer diagnosis and exploring gene-disease associations.
• It democratizes spatial transcriptomics analysis by leveraging standard histopathology data.