SHEST: single-cell-level artificial intelligence from haematoxylin and eosin morphology for cell-type prediction and...

Area of Science:

• Computational Biology
• Genomics
• Cancer Research

Background:

• Cancer progression understanding requires integrating tissue morphology and spatial molecular data.
• Existing methods lack a unified approach for cell-type prediction and spatial gene expression reconstruction.

Purpose of the Study:

• To introduce SHEST, a multi-task framework for predicting cellular composition and reconstructing spatial gene expression from haematoxylin and eosin morphology.
• To bridge histopathology and spatial transcriptomics for comprehensive tissue characterization.

Main Methods:

• SHEST utilizes a quadruple-tile input and a neighbourhood-informed clustering algorithm.
• A shared morphological encoder with task-specific heads for cell-type classification and gene expression reconstruction.
• Multi-task optimization with cross-entropy and zero-inflated negative binomial losses for sparse spatial transcriptomic data.

Main Results:

• High accuracy in predicting tumour cells (F1: 0.97) and lymphocytes (F1: 0.91) in lung adenocarcinoma.
• Successful reconstruction of spatially resolved, cell-type-specific gene expression patterns.
• Preservation of spatial relationships and gene-level autocorrelation, reflecting tumour microenvironment structure.

Conclusions:

• SHEST provides a synergistic and cost-efficient method to integrate histopathology and spatial transcriptomics.
• Enables comprehensive tissue characterization and cell-level insights into tumour-immune ecosystems.
• Forms a foundation for precision oncology through spatially informed analysis.