Gene-Morphology Alignment via Graph-Constrained Latent Modeling for Molecular Subtype Prediction from Histopathology...

Area of Science:

• Computational biology
• Digital pathology
• Cancer genomics

Background:

• Molecular subtyping of cancer traditionally relies on transcriptomic data, but sequencing costs limit clinical application.
• Histopathology offers rich morphological data correlating with molecular states but lacks a direct link to gene expression.
• Bridging morphology and gene-level representations is crucial for accessible precision oncology.

Purpose of the Study:

• To develop a computational framework aligning histopathology morphology with molecular subtypes.
• To enable cancer subtyping using only routine histopathology slides, bypassing gene sequencing.
• To create a 'virtual transcriptomics' approach for broader clinical utility.

Main Methods:

• A graph-constrained learning framework was developed to link morphology-derived signals to a gene network.
• Hierarchical Monte Carlo screening identified a data-driven gene network.
• Random sampling derived gene sets for classification, using coexpression networks to enforce morphology-based learning without gene expression data.

Main Results:

• The model successfully predicted pancreatic cancer subtypes (Moffitt classification) using morphology alone.
• Achieved 85% AUC for subtype prediction in PANCAN and TCGA datasets.
• Demonstrated that morphology-based predictions, guided by gene network structures, yield high accuracy.

Conclusions:

• This framework establishes virtual transcriptomics by deriving molecular insights from histopathology.
• It bypasses the need for gene sequencing, potentially expanding precision oncology access.
• The approach offers a cost-effective and accessible method for cancer subtyping.