Abstract Background: Clinical criteria, such as Chompret 2015, guide genetic testing for Li-Fraumeni Syndrome (LFS) and TP53 mutations. However, these guidelines may overlook hereditary breast cancer risk in genetically diverse populations. In Brazil, the founder mutation TP53 p.R337H is highly prevalent, presenting with atypical tumor patterns and age of onset, often leading to underdiagnosis by conventional criteria. This study aims to assess whether machine learning (ML) and explainable artificial intelligence (XAI) can enhance the prediction of hereditary breast cancer subtypes, particularly TP53-related, in a real-world Brazilian cohort. Methods: We curated a dataset of 576 probands (515 with complete cancer history data) subjected to genetic testing and with structured clinical and family information.Three multiclass machine learning models were trained to predict three groups: TP53-negative, TP53 R337H, and other TP53 mutations. Performance was evaluated using ROC-AUC, and SHAP values were employed to interpret variable importance. Among the tested algorithms, CatBoost achieved the highest ROC-AUC values, although Random Forest was selected for interpretability. Kaplan-Meier analysis assessed age at diagnosis across groups. Results: Among 515 probands, 70.1% had a personal cancer history. Among those with cancer, 32.7% carried pathogenic mutations, with TP53 variants being the most frequent: 7.8% for R337H and 6.1% for other TP53 mutations, totaling 13.8%. Other recurrent mutations included BRCA2 (3.0%), BRCA1 (2.5%), CHEK2 (1.4%), ATM (1.1%), and NF1 (0.6%). The remaining 67.3% had no identifiable pathogenic variant. The model achieved robust discrimination (AUC: 0.73 for TP53 R337H, 0.84 for other TP53, 0.85 for TP53-negative), outperforming the Chompret clinical criteria (AUC = 0.67). Unlike Chompret, the ML model distinguished between R337H and other TP53 mutations, offering a more tailored and data-driven approach to hereditary cancer risk stratification in genetically diverse populations. SHAP analysis identified key predictors for TP53 p.R337H, including early-onset maternal cancer, proband's breast cancer, and early-onset brain tumors. Notably, TP53 p.R337H cases showed a significantly earlier median age at diagnosis (43 years) compared to TP53-negative cases, yet later than classical TP53 mutations. The model also accurately revealed that several TP53-negative individuals carried pathogenic variants in BRCA1, BRCA2, and ATM, with early-onset breast cancer mirroring TP53 carriers, highlighting the limitations of TP53-centric screening. Conclusion: Our findings reveal that Chompret criteria may miss substantial hereditary breast cancer risk associated with LFS, especially in populations with founder mutations like TP53 p.R337H. Explainable ML models integrating comprehensive clinical and familial data provide accurate and interpretable risk stratification, demonstrating superior performance in identifying diverse hereditary cancer patterns. These insights support the use of multigene panel testing and population-tailored triage strategies for improved genetic counseling and surveillance in hereditary breast cancer care, particularly in regions with high genetic heterogeneity. Citation Format: J. Casali da Rocha, A. C. Ricciardi, G. B. Pinheiro1, D. S. Pegos, R. A. Romero. Explainable machine learning reveals hidden hereditary risk of breast cancer beyond TP53: insights from Brazilian families with high prevalence of the p.R337H mutation abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-11.
Rocha et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: