Liquid biopsy of lung cancer by deep learning and spectroscopic analysis of circulating exosomes.

Abstract

e15532 Background: Lung cancer has a high mortality rate because of belated diagnosis at advanced stages beyond the treatable condition. Early detection of lung cancer can improve the survival rate. A liquid biopsy that detects tumor-related biomarkers in body fluids has a great potential for the purpose. Particularly, tumor-derived exosomes in blood have been proposed as a promising biomarker. The tumor-derived exosomes carry molecules of their parental cells; thus, they provide information about the tumor in the body. Unfortunately, exosomal markers conducive to the early detection of lung cancer are still obscure. Therefore, using the molecular fingerprint of exosomes markers can be useful to detect the tumor exosomes. Raman spectroscopy is one of the representative methods for the purpose. However, because the exosomes have a heterogeneous composition in blood, interpreting their spectroscopic signals is hard. Thus, we utilized a deep learning approach to analyze the spectroscopic signal of the exosomes for liquid biopsy of lung cancer. Methods: The basic concept was to evaluate how much the exosomes in human plasma resemble cancer cell exosomes. As a proof of concept, exosomes of 43 non-small cell lung cancer (NSCLC) adenocarcinoma patients and 20 healthy controls were isolated from plasma of peripheral blood. Also, cell exosomes were isolated from culture media of adenocarcinoma cell lines and a human pulmonary alveolar epithelial cell line. Then, the spectroscopic signals were detected using surface-enhanced Raman spectroscopy (SERS). Further, the deep learning algorithm was employed to classify the signals. Then, we calculated the relative similarity to cancerous exosomes against human plasma exosomes. Results: Our method was able to classify cancer and normal cell exosomes with 95% accuracy. Also, Raman signals of cancer patients’ exosomes were more similar to the cancer cell exosomes than those of healthy controls. Notably, the similarity was proportional to cancer stages. Importantly, our method even detected stage I patients. The area under the curve (AUC) of receiver operating characteristic (ROC) curves was 0.912 for stage I and II, and 0.910 for stage I. Conclusions: We reported a novel diagnostic method using deep learning analysis against spectroscopic signals of circulating exosomes. Our method that evaluates the similarity to cancer exosomes accurately identified lung cancer patients, even stage I with high accuracy.