Raman and SERS analytical platform with machine learning for classifying wild type p53 and hotspot mutants R175H and R273H

The development of analytical approaches capable of detecting subtle protein conformational changes is of significant interest in biomedicine, particularly for disease diagnostics and biomolecular characterization. In this work, the tumor suppressor p53 protein was selected as a model system to evaluate an analytical platform based on Raman spectroscopy combined with surface-enhanced Raman spectroscopy (SERS) and supervised machine learning for the classification of closely related protein variants. Label-free Raman and SERS spectra of recombinant wild-type p53 and its hotspot mutants R175H and R273H were acquired on bare Al and Al coated with gold nanospheres, gold nanorods, and silver nanoparticles. Principal Component Analysis (PCA) revealed distinct spectral fingerprints among the p53 variants, with relevant contributions in the amide III and CH stretching regions, indicating subtle conformational differences. Supervised classification was performed using several machine learning algorithms under a nested, group-disjoint cross-validation protocol (holding out entire acquisition lines) to account for within-line spectral correlation. Among tested models, a linear Support Vector Machine (Linear-SVM) achieved the highest performance, reaching an accuracy of 92.9 ± 6.9% on AuNS@Al substrates. These results demonstrate that the integration of optimized nanostructured SERS substrates, Raman spectroscopy, and machine learning constitutes a robust analytical platform for the structural classification of wild-type and mutant p53 proteins. The proposed approach provides a versatile framework for studying conformational changes in biomolecules of biomedical relevance and supports its potential application in cancer biomarker detection and protein misfolding studies.