Interpretable dual-wavelength SERS profiling of tumor-associated extracellular vesicles via EV-adapted plasmonic substrates and feature-fusion learning.

Label-free surface-enhanced Raman spectroscopy (SERS) profiling of extracellular vesicles (EVs) enables integrated analysis of biochemical differences associated with tumor-related biological states, but its performance can be limited by inefficient coupling of nanoscale vesicles with plasmonic hot spots, incomplete information from single-wavelength excitation, and overlapped Raman bands. Herein, we report an interpretable dual-wavelength EV-SERS framework integrating an EV-adapted porous anodic aluminum oxide (AAO)/Au nanosphere substrate with feature-fusion deep learning. The three-dimensional positively charged plasmonic architecture facilitates adsorption of isolated EV preparations and enhances SERS sensitivity. EVs derived from cisplatin-sensitive/resistant A549 cells and plasma EVs from glioma-bearing/healthy mice were analyzed under 633 and 785 nm excitation. A one-dimensional conditional generative adversarial network was used for training-set augmentation, and a dual-channel SpecKAT model fused wavelength-specific spectral features, with final performance evaluated on held-out experimentally acquired spectra. The framework achieved classification accuracies of 98% for cisplatin-resistant versus parental A549-derived EVs and 97% for plasma EV preparations from glioma-bearing versus healthy mice, outperforming single-wavelength models and direct spectral concatenation. SHAP analysis combined with Voigt peak fitting associated model-relevant Raman regions with protein-, lipid-, aromatic amino acid-, and amide-related vibrations. Overall, the framework demonstrates promising methodological performance in cell-derived EVs and a preclinical mouse plasma model, while validation in independent human plasma cohorts remains necessary to assess robustness and generalizability.