Predicting TP53 Biomarkers from Whole Slide Images across Human Solid Tumors Using Weakly Supervised Learning.

Area of Science:

• Computational pathology
• Digital oncology
• Artificial intelligence in cancer research

Background:

• Accurate molecular profiling from histopathology slides can advance clinical oncology.
• Current methods for analyzing tumor characteristics from slides are limited.

Purpose of the Study:

• To develop and evaluate a Vision Transformer (ViT)-based model for comprehensive analysis of Whole Slide Images (WSIs).
• To jointly predict TP53 biomarker status, detect 32 solid tumor types, and assess survival directly from WSIs.

Main Methods:

• Utilized over 11,000 primary tumor WSIs from the Pan-Cancer Atlas with associated molecular and clinical data.
• Employed a ViT encoder for feature extraction after preprocessing WSIs.
• Trained the model in two stages: initial training on tumor patches and fine-tuning on WSIs.

Main Results:

• Achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.766 for TP53 mutation detection across 32 tumor types on an independent validation set.
• The model successfully inferred TP53 mutation status, RNA expression levels, and tumor taxonomy from WSIs.
• Prognostic risk prediction accuracy was limited.

Conclusions:

• A ViT-based model can simultaneously predict TP53 mutation status, RNA expression, and tumor type from WSIs.
• Findings suggest reproducible morphologic correlates of TP53 alterations exist across human cancers.
• The model demonstrates potential for transforming molecular profiling in clinical oncology.