SERS mapping combined with explainable deep learning for exosome analysis to enhance lung cancer detection

Abstract

Exosomes are critical biomarkers for early cancer diagnosis and prognosis due to their rich biological information. Nevertheless, analyzing exosomal biomarkers comprehensively remains challenging. Surface-enhanced Raman scattering (SERS) has been employed to detect exosomes due to its high sensitivity and reliable fingerprint. However, most Raman signals originate from surface molecules rather than exosomal cargo, as the SERS effect decreases significantly beyond 10 nm from the metal surface, while exosomes have a lipid bilayer of approximately 5 nm thickness. Herein, we demonstrate the enhanced detection accuracy of lung cancer cells by exhaustively analyzing SERS signals of exosomes, including surface and internal biomarkers, using a smart and explainable deep learning model. Specifically, gold nanocube superlattices (GNSs) were prepared by the Marangoni effect-driven self-assembly to obtain SERS mapping signatures of lung cancer-derived exosomes. The gradient-based category activation mapping (Grad-CAM) augmented-deep learning model was then constructed to recognize the signal patterns of exosomes to identify the presence of lung cancer and simultaneously visualize crucial features in the SERS spectra that contributed to lung cancer detection. The model was trained using SERS signals from both surface and internal biomarkers derived from normal and lung cancer cells, achieving a classification accuracy of 98.95%. In contrast, when trained solely on surface biomarkers, the model achieved an accuracy of 96.35%. Moreover, Grad-CAM highlighted interpretable molecular signatures in the SERS spectral data, reflecting the network's decision-making logic. These findings demonstrate the power of combining SERS mapping of exosomal biomarkers with explainable deep learning, bridging the gap between model performance and human-understandable explanations.