Biomineralized Surface-Enhanced Raman Scattering Nanotags Encode Biomolecular Identity into Machine Learning-Resolvable Plasmonic Fingerprints

Abstract

Surface-enhanced Raman scattering (SERS) nanotags provide highly sensitive platforms for in vitro diagnostics but often require complex, disease-specific customization that limits clinical translation. Biomineralization, in which biomolecules mediate inorganic material synthesis, offers a versatile yet underexplored strategy for generating functional SERS nanotags. Here, we demonstrate that biomolecule-directed biomineralization of gold nanoparticles (AuNPs) using amino acids and exosomes generates distinct nano-bio interfacial architectures that encode biomolecular identity into machine learning-resolvable SERS fingerprints through modulation of plasmonic coupling and Raman reporter organization. As a proof-of-concept system, amino acid-biomineralized AuNPs were synthesized using biomolecules with diverse physicochemical properties, including differences in size, polarity, and charge. The resulting nanotags were characterized using UV-Vis spectroscopy, SERS, fluorescence spectroscopy, dynamic light scattering (DLS), and transmission electron microscopy (TEM). Random forest and support vector machine (SVM) models successfully differentiated amino acid-dependent SERS signatures with near-perfect classification performance. Extending this approach to a biologically complex preclinical cancer model, exosome-biomineralized AuNP nanotags were generated using exosomes derived from clinically relevant pediatric patient-derived osteosarcoma and neuroblastoma tumors. Distinct exosome-dependent spectral fingerprints enabled SVM classification with 93.9% accuracy, while Shapley Additive exPlanations (SHAP) and t-distributed stochastic neighbor embedding (t-SNE) analyses identified diagnostically relevant spectral regions and visualized clustering between tumor classes. Collectively, this work establishes biomineralization as a strategy for transforming complex biomolecular and cellular information into computationally resolvable optical fingerprints, enabling scalable and label-free diagnostic classification of patient-derived biomolecular samples.