Predicting TP53 biomarkers from whole slide images across human solid tumours using weakly supervised learning

Accurate molecular profiling from routine histopathology slides could transform clinical oncology. A Vision Transformer (ViT)-based model was developed to jointly predict the TP53 biomarker, detect 32 solid tumour types, and predict survival directly from Whole Slide Images (WSIs). Over 11K primary tumour data were retrieved from the Pan-Cancer Atlas, along with corresponding somatic mutation, RNA-seq, and clinical outcome data. WSIs underwent tissue masking, quality control, stain normalisation, patch extraction, and feature embedding using a ViT encoder. Seven task heads were developed to generate predictions for cancer type, TP53 mutation status, TP53 RNA expression levels, overall survival, progression-free interval, and their corresponding event times. Model training proceeded in two stages: initial training on tumour-only patches at multiple magnifications, followed by fine-tuning on WSIs using a content-aware strategy. Model performance was evaluated on an independent validation set of 1,729 slides using evaluation metrics, including the area under the receiver operating characteristic curve (AUROC), regression metrics, and the concordance index. An AUROC of 0.766 for TP53 mutation detection on an independent validation set across 32 human solid tumours. In conclusion, the ViT-based model could simultaneously infer TP53 mutation status, TP53 RNA expression levels, and tumour taxonomy directly from WSIs, supporting the existence of reproducible morphologic correlates of TP53 alterations across human cancers, whereas prognostic risk prediction remained limited.