69P Deep learning prognostication through prediction of TP53 gene mutation status on breast cancer hematoxylin and eosin slides

TP53 gene plays a crucial role in cell functions, such as proliferation, survival, and maintaining genomic integrity. In breast cancer, TP53 mutations are associated with worse prognosis. The pivotal role of p53 in tumor development suggests that morphological features in H&E-stained tumor tissue sections may be distinctive for the presence of mutation. The aim of this study is to develop an artificial intelligence (AI) model for predicting TP53 mutations to be used for patient prognostication. We trained an ER-stratified deep learning model on surgical H&E-stained tissue sections for predicting TP53 gene mutation status on TCGA-BRCA (N=943). We validated its ability to predict TP53 mutations and its prognostic performance on an external validation cohort (N=207 patients primarily treated in Uppsala, Sweden, 1987-1989). TP53 mutations were observed in 30.9% (NGS) of the individuals, with rates of 19% and 72.7% in the ER+ and ER- subgroups in TCGA; external cohort mutation rates were 22.7% overall (Sanger sequencing), 17.4% for the ER+ and 47.4% for the ER- subgroups. A Convolutional Neural Network pretrained on ImageNet was used to extract features from the tissue slides, while a multiple instance learning classifier was trained for mutation status prediction based on bags of instances of the extracted imaging features. ROC curve analysis was used for optimal cut-off estimation. The model achieved an overall AUC of 0.76 on the validation set, reaching 0.74 and 0.77 for ER+ and ER- subgroups respectively. Digital TP53 prediction was significantly prognostic with respect to overall survival in univariate (HR=1.7, 95% CI: 1.1-2.5, p=0.019) and multivariate analyses (HR=2.1, 95% CI: 1.3-3.4, p=0.002) after adjusting for ER, age, size, nodal and HER2 status. Additional results on external validation cohorts are ongoing and will be presented. We developed and validated an AI model that showed promise as a cost-effective and robust alternative surrogate method to predict TP53 mutation status on H&E-stained tissue sections in breast cancer. The model's prognostic value independent of established prognostic clinical factors highlights its potential clinical relevance as a digital biomarker.