Generalizable biomarker prediction from cancer pathology slides with self-supervised deep learning: A retrospective...

Area of Science:

• Computational pathology
• Oncology
• Biomarker discovery

Background:

• Deep learning (DL) shows promise in predicting biomarkers like microsatellite instability (MSI) from colorectal cancer (CRC) histopathology.
• Generalizability and performance for other key biomarkers remain uncertain.

Purpose of the Study:

• To evaluate the performance of various DL architectures in predicting multiple biomarkers from CRC pathology slides.
• To assess the generalizability of DL predictions using a large external validation cohort.
• To compare DL performance against established clinical standards.

Main Methods:

• Acquired CRC tissue samples from two large multi-centric studies.
• Systematically compared six state-of-the-art DL architectures for predicting MSI, BRAF, KRAS, NRAS, and PIK3CA mutations.
• Utilized self-supervised, attention-based multiple-instance learning models.
• Validated model performance on a large external cohort.

Main Results:

• Self-supervised, attention-based multiple-instance learning models outperformed previous DL approaches.
• DL models achieved clinical-grade performance for predicting MSI and BRAF mutations.
• Prediction accuracy for PIK3CA, KRAS, and NRAS mutations was clinically insufficient.
• Explainable visualizations of indicative regions and morphologies were provided.

Conclusions:

• DL can accurately predict certain biomarkers (MSI, BRAF) from CRC histopathology slides with clinical-grade performance.
• Model generalizability is crucial and was demonstrated through external validation.
• DL holds potential for non-invasive biomarker prediction in CRC, but further development is needed for KRAS, NRAS, and PIK3CA mutations.