Abstract P3-01-05: Liquid biopsy methods and machine learning modeling to understand organ tropism and metastatization behavior of metastatic breast cancer

Background: Liquid biopsy provides a growing amount of real-time data about prognosis and the genomic landscape of metastatic breast cancer (MBC) and its comprehensive analysis is an emerging clinical need. Machine Learning (ML) data-driven models are able to “learn” information about a system and to adaptively improve their performance by directly observing its data. This enables the discovery of hidden patterns in complex heterogeneous and high dimensional data. The aim of this study was to explore the combination of clinical characteristics, circulating tumor DNA-detected aberrations (ctDNA) and CTC enumeration in estimating target organs more susceptible to MBC involvement using a ML modeling approach. Methods: The study retrospectively analyzed 88 MBC patients (pts) treated and characterized for CTCs and circulating tumor DNA (ctDNA) at Northwestern University (Chicago, IL) independently from treatment line. Blood samples were collected at baseline, concomitantly with imaging. CTCs were isolated through the CellSearch™ kit (Menarini Silicon Biosystems, PA), while ctDNA was analyzed using the Guardant360™ NGS-based assay (Guardant Health, CA). All features were normalized and included in a random forest algorithm implemented in Python (Scikit learn, BSD license), node splitting criterion for the decision tree classifiers was varied using Gini index and entropy. Hyperparameters of the random forest were then optimized including number of trees and the minimum leaf size by implementing hyperparameter grid search using 10-fold cross validation. Results: The median number of lines at baseline collection was 1 (interquartile range: 1-3), while the median number of metastatic sites was 3 (inter quartile range: 1-3) with the most observed sites being bone (37%), lymph nodes (29%), lung (27%) and liver (25%). The cohort consisted of 43% hormone receptor positive (HRpos), 32% TNBC, and 25% HER2-positive MBC. In the overall population, continuous CTC number (n_CTC), inflammatory breast cancer diagnosis (IBC), and aberrations in ESR1, KITand CDK4were the main features linked to liver metastases (AUC: 0.842), n_CTC, ESR1, PIK3CA, CCNE1and CDK6were the features linked to bone involvement (AUC: 0.770), while PIK3CA, METand MYC, were linked to lung organotropism (AUC: 0.701). Factors linked to the metastatization net combination pattern were then explored within each MBC subtype. Intriguingly, AR, n_CTC, TP53and ESR1were the main drivers in HRpos MBC (Mean per class error0.46), while EGFR, KITand NOTCH1were the main features in TNBC (Mean per class error 0.605). Consistently, n_CTC, ERBB2, PIK3CAwere the driving features among HER2 positive MBC pts (Mean per class error 0.87). Conclusions: This novel analysis demonstrates that liquid biopsy integrating both CTCs enumeration and genomic characterization by ctDNA could prove useful in a detailed description of the metastatic process, allowing a more tailored monitoring and therapeutic approach. Intriguingly, features linked to Epithelial to Mesenchymal transition were found to be a potential driver of the metastatization behavior, underlining the need to further elucidate the clinical impact of this process. Citation Format: Lorenzo Gerratana, Andrew A Davis, Maurizio Polano, Qiang Zhang, Ami N Shah, Chenyu Lin, Debora Basile, Giuseppe Toffoli, Firas Wehbe, Fabio Puglisi, Amir Behdad, Leonidas C Platanias, William J Gradishar, Massimo Cristofanilli. Liquid biopsy methods and machine learning modeling to understand organ tropism and metastatization behavior of metastatic breast cancer [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P3-01-05.