Abstract A12: Combining the immunophenotyping of MDSCs and lymphocytes with artificial intelligence (AI) to predict early-stage breast cancer

Abstract

Abstract The purpose of this study was to analyze immunophenotyping flow cytometry data using artificial intelligence (AI) to identify individuals presenting with early-stage breast cancer (stage I/II). Peripheral blood was collected from 64 subjects diagnosed with a biopsy-confirmed breast cancer (BC; Stage I n = 42 and Stage II n = 22) along with 80 tumor-free control female donors (HD). Subjects were excluded if they had a previous history of cancer and/or a medical intervention for a breast pathology. For each subject, we profiled various myeloid and lymphoid cell populations using standard multiparametric flow cytometry. Importantly, the cell surface markers included would identify myeloid-derived suppressor cell (MDSC) subsets—known key contributors in supporting tumor progression and tumor escape. Previous studies have investigated whether measuring MDSCs can act as a liquid biopsy to detect tumor development, monitor progression, and/or predict therapeutic responses. Next, a series of feedforward artificial neural networks, or multilayer perceptrons (MLPs), were created in silico to classify samples as either BC or HD. The network inputs consisted of the channel values for each cell event from each fluorescent and scatter parameter and were used to construct three datasets: the training dataset—to “teach” two output categories (BC or HD) through backpropagation and parameter fitting; the validation dataset—to evaluate the fit to minimize overfitting; and the test dataset—to rank the trained networks against each other and estimate the classification performance. Finally, a naive testing set (i.e., never seen by the network) was used to determine the overall performance of the top-ranking networks after voting. With this approach, we were able to distinguish BC subjects from HD subjects with 92.3% sensitivity and 86.7% specificity (AUROC = 0.8987; CI95% 0.8047 to 0.9927). Additionally, we tested 14 samples collected from subjects with biopsy-confirmed ductal carcinoma in situ (DCIS). Even though they are clinically deemed as precancerous (stage 0), 11 out of 14 of these subjects were classified as BC, indicating its possible utility for detecting the existence of even a noninvasive cancerous lesion. By pairing supervised machine learning with the immunophenotyping of MDSCs and other leukocytes using flow cytometry, we have developed a novel method for possibly distinguishing breast cancer subjects from those who are tumor free with high levels of accuracy using a simple blood draw. Although further study is needed, we believe this could potentially prove clinically useful in combination with current screening methods, such as mammography, to reduce the number of unnecessary biopsies performed each year. Citation Format: George A. Dominguez, John Roop, Alexander Polo, Anthony Campisi, Dmitry I. Gabrilovich, Amit Kumar. Combining the immunophenotyping of MDSCs and lymphocytes with artificial intelligence (AI) to predict early-stage breast cancer [abstract]. In: Proceedings of the AACR Special Conference on Tumor Immunology and Immunotherapy; 2018 Nov 27-30; Miami Beach, FL. Philadelphia (PA): AACR; Cancer Immunol Res 2020;8(4 Suppl):Abstract nr A12.