Predicting Mutational Status of Driver and Suppressor Genes Directly from Histopathology With Deep Learning: A...

Area of Science:

• Computational pathology
• Genomics
• Oncology

Background:

• Deep learning advances allow inferring genetic alterations from histopathology slides.
• Previous studies focused on individual genes in specific tumor types.
• Genetic alterations in solid tumors impact cell behavior via signaling pathways.

Purpose of the Study:

• To train deep learning networks to predict gene and pathway alterations across diverse solid tumors.
• To investigate if genetic changes are detectable directly from routine histology images.
• To compare the morphological signature strength of single genes versus pathways.

Main Methods:

• Manual outlining of tumor tissue in 7,829 patient H&E-stained slides (23 tumor types).
• Training convolutional neural networks (CNNs) end-to-end to detect gene/pathway alterations from images.
• Utilizing data from The Cancer Genome Atlas for training and validation.

Main Results:

• Deep learning successfully detected alterations in 12 out of 14 clinically relevant pathways.
• Numerous single gene alterations were also predicted, many novel.
• Prediction performance was higher for single gene alterations than for pathway alterations.

Conclusions:

• Genetic alterations in solid tumors are predictable directly from routine H&E histology images using deep learning.
• Individual genes possess stronger morphological signatures than genetic pathways.
• This approach holds potential for non-invasive genetic profiling in cancer research and diagnostics.