Deep Learning-Based Classification of NSCLC-Derived Extracellular Vesicles Using AFM Nanomechanical Signatures

Abstract

Nonsmall cell lung cancer (NSCLC) remains a leading cause of cancer-related mortality, with liquid biopsy emerging as a promising tool for noninvasive diagnostics. Extracellular vesicles (EVs) serve as molecular messengers of the tumor microenvironment, yet precise characterization methods remain limited. Using atomic force microscopy (AFM), we analyzed EVs from NSCLC subtypes (A549, PC9, PC9/GR) and nontumorigenic bronchial epithelial cells (BEAS-2B), revealing that A549-derived EVs exhibited significantly higher stiffness, likely due to KRAS mutation-associated lipid alterations. EGFR mutant EVs (PC9, PC9/GR) showed overlapping nanomechanical properties, correlating with their shared genetic background. To enhance classification, we implemented a DenseNet-based deep learning model for AFM image analysis, integrating nanomechanical and morphological features. This approach significantly improved diagnostic performance, achieving an AUC of 0.92, and notably, EVs from the A549 (KRAS mutant) cell line were classified with 96% accuracy. This study provides the first demonstration of the nanomechanical classification of NSCLC-derived EVs, highlighting the potential of deep learning-enhanced AFM analysis as a powerful tool for advancing liquid biopsy and precision diagnostics. Addressing sample variability and validating performance across clinical samples will be key to enabling its clinical translation.