Genomic predictors of brain metastases in breast cancer.

Abstract

2040 Background: Despite therapeutic advances in metastatic breast cancer (MBC), the rising incidence of brain metastases (BM) remains a major challenge, contributing to poor prognosis and significant morbidity. Due to the absence of consensus screening strategies for BM, they are often detected only after clinical symptoms emerge. There is therefore a pressing need for predictive biomarkers to identify breast cancer patients at risk of BM. Methods: This study included 3908 patients who underwent sequencing of primary tumor (n = 1885) or non-brain metastasis (n = 2023) with MSK-IMPACT, a custom tumor-normal next generation sequencing assay. First, we performed penalized logistic regression on a gene level to identify alterations in extracranial metastases or primary tumors associated with development of BM. We adjusted for multiple hypothesis testing using Benjamini-Hochberg. Lastly, we developed a lasso machine-learning (ML) model, incorporating baseline genomic and clinicopathologic features, to predict onset and timing of BM from initial diagnosis (for early stage cases) or metastatic disease (for MBC). Each analysis was stratified by receptor status, and repeated to account for loss of heterozygosity (LOH) of tumor suppressor genes. Results: Our cohort included 528 BM events over a median follow-up of 58 mos. Pathogenic variants in several genes were associated with subsequent BM development. In the HR+/HER2- subset (n = 2624), pathogenic variants in the following genes portended the onset of BM: RB1 (OR 2.59 [1.39 - 4.81], q = 0.011), NF1 (OR 2.22 [1.21 - 4.06], q = 0.039), TP53 (OR 1.91 [1.43 - 2.54], q < 0.001), PIK3CA (OR 1.47 [1.21-4.06], q = 0.028). Pre-existing LOH of RB1, in the absence of an RB1 functional alteration, was associated with BM development (OR 1.37 [1.03 - 1.83], q = 0.090). TP53 LoF .OR 5.14 [2.21 - 11.9], q < 0.001) was enriched in the BM group in HER2+ tumors, while amplification of CDKN2A (OR 11.6 [2.44 - 55.5], q = 0.01) or EGFR (OR 4.60 [1.54 - 13.8], q = 0.03) were enriched in TNBC. TP53 emerged as an important feature across all receptor subtypes in our machine-learning model; RB1 LoF was also selected as an important feature in the HR+/HER2- group. Validation of the ML model in an external cohort will be presented at the meeting. Conclusions: In a large cohort of genomically profiled breast cancer samples, we found several biologically plausible candidates for molecular harbingers of BM. For instance, the recurrent involvement of genes involved in cell cycle regulation ( RB1, CDKN2A, TP53) has been implicated as candidates for BM tropism in other cancer types. Our approach also uncovers several alterations for which targeted therapies exist or are actively in development (NF1, PIK3CA). Our clinically actionable multimodal model of BM risk is poised to facilitate the development of early detection strategies and guide-high risk patient selection for novel clinical trials to intercept this devastating complication.