Identifying transcriptomic correlates of histology using deep learning.

Area of Science:

• Computational biology
• Genomics
• Histopathology

Background:

• Linking gene expression to biological phenotypes is crucial but challenging.
• Traditional methods rely on subjective interpretation or extensive gene perturbations.
• Current approaches struggle to bridge the gap between molecular data and tissue-level organization.

Purpose of the Study:

• To develop a Deep Learning (DL) model for objective identification of histological features.
• To establish quantifiable correlations between visual tissue phenotypes and gene expression profiles (transcriptomes).
• To enhance the interpretability and explainability of DL-derived biological insights.

Main Methods:

• Utilized a dataset of whole slide images and matching gene expression data from 39 normal tissue types.
• Developed a DL-based tissue classifier achieving 94% accuracy.
• Identified genes correlating with DL-inferred visual features and visualized these features.

Main Results:

• DL successfully classified tissue types with high accuracy.
• The model automatically derived biologically interpretable visual features correlated with transcriptomes.
• Visualizations of inferred features aligned with immunohistochemistry data, confirming biological relevance.

Conclusions:

• Deep learning offers a powerful, objective approach to link histology and genomics.
• This method facilitates the discovery of novel transcriptome-phenotype correlations.
• The study provides a foundation for bridging molecular and cellular levels of biological organization.