Abstract PS2-08-05: Multimodal integration of real world genomic and clinical data for the prediction of brain metastasis development in breast cancer

Abstract Background: Despite advances in metastatic breast cancer (MBC) treatment, brain metastases (BM) continue to pose a major challenge, driving significant morbidity and poor prognosis. In the absence of consensus screening strategies, BMs are typically diagnosed only after neurological sequelae have already developed. This underscores the urgent need for predictive tools to identify patients at risk for BM before clinical symptoms arise. Methods: This study included 3901 patients who underwent sequencing of primary tumor (n = 1874) or non-brain metastasis of samples (n = 2027) with MSK-IMPACT, a custom tumor-normal next generation sequencing assay. We developed lasso machine-learning models to predict onset and timing of BM in two different scenarios. First, we predicted time from to BM diagnosis from the initial biopsy confirming metastatic breast cancer. Second, we developed an “early-stage” model capturing time from curative breast surgery to BM onset. Both models exclusively incorporated clinicopathologic and genomic data available at time of initial metastatic disease and initial surgery, respectively. We developed validative gradient boosting machine (GBM) machine learning and Fine-Gray models, allowing for consideration of variable interactions and competing risks, respectively. Results: Our cohort included 529 BM events over a median follow-up of 50 mos. In metastatic samples, the lasso-based ML model effectively stratified the cohort into high, intermediate and low-risk groups: the high-risk group (25% of the cohort) exhibited a hazard ratio (HR) of 10.1 (95% C.I 6.83 - 17.23) while the intermediate group (50% of the cohort) exhibited an HR of 3.65 (95% C.I 2.30 - 5.78) relative to the low-risk group. The model was driven by several genomic features (functional variants in TP53, RB1, PTEN, MDM2) as well as several clinical features (low ER and PR percentage, number of metastatic sites, metastasis to lung, adrenal or distant lymph node, and pre-menopausal status). In early-stage samples, the ML model stratified the cohort into two groups, with the high-risk group exhibiting an HR of 5.7 (95% C.I 3.83 - 8.53) relative to the low-risk group. Feature importance in the early-stage setting was driven by stage, ER/PR status as well as TP53 mutation status. The predictive performance of these models, along with the relevance of these key variables, was preserved in the GBM and Fine-Gray models. Conclusions: Leveraging a large cohort of genomically profiled breast cancers, we developed a machine learning model that effectively stratifies patients by risk of developing brain metastasis. The most influential variables informing risk stratification were biologically and clinically plausible. For example, recurrent alterations in cell cycle regulation (RB1, TP53) have been linked to brain metastasis tropism in other cancer types and emerged as key features in our analysis. Our clinically actionable multimodal model allows for early identification of patients at risk for brain metastasis, enabling the study of personalized screening and intervention approaches to identify and intercept this debilitating complication. Citation Format: A. Safonov, D. Smith, S. Nandakumar, L. Boe, E. Ferraro, J. Shen, K. Tsai, I. Khatri, R. Kumar, J. An, J. Jee, M. Robson, N. Schultz, N. Moss, W. Chatila, L. R. Pike, P. Razavi. Multimodal integration of real world genomic and clinical data for the prediction of brain metastasis development in breast cancer 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 PS2-08-05.