Cross-linking breast tumor transcriptomic states and tissue histology

Area of Science:

• Computational pathology
• Genomics
• Bioinformatics

Background:

• Predicting gene expression from pathology images is crucial for cancer patient care.
• Challenges exist due to complex, correlated gene expression patterns.

Purpose of the Study:

• To develop a data-driven method for predicting cancer gene expression states from whole slide images (WSIs).
• To identify biologically meaningful gene groups for improved prediction accuracy and clinical utility.

Main Methods:

• Utilized a data-driven approach to identify co-dependent gene expression groups.
• Employed a bespoke graph neural network (GNN) to predict gene group status from WSIs.

Main Results:

• Successfully identified biologically meaningful gene groups that capture patient gene expression states.
• Enabled prediction of gene expression states directly from histopathology images, correlating with clinical phenotypes.

Conclusions:

• This method offers a novel way to gain biological insights from routine pathology imaging.
• The identified gene groups provide clinically relevant histopathological insights for therapeutic applications.