C11 - Machine Learning-Driven Analysis of Surface-Enhanced Raman Scattering (SERS) Spectra for Multi-Faceted Disease Diagnostics

Introduction: : Surface-Enhanced Raman Spectroscopy (SERS) is a powerful, label-free technique that enables molecular fingerprinting of biomolecules with high sensitivity. The diagnostic potential of SERS is significantly enhanced when paired with gold nanoparticle-based nanotags (AuNPs). However, the high dimensionality and variability of SERS spectra complicate data interpretation and hinder clinical translation. We hypothesize that machine learning (ML) approaches could address these challenges by enabling the extraction of meaningful spectral features, reducing noise, and improving classification accuracy, which will be vital for SERS-based diagnostic applications. In this study, we employed a range of machine learning algorithms to classify SERS spectra from seven amino acids biomineralized onto AuNPs.

Materials and

Methods: : We synthesized SERS nanotags by biomineralizing AuNPs with seven α-amino acids: Glutamic acid, Serine, Glycine, Leucine, Proline, Phenylalanine, and Valine, respectively, and performed rigorous physiochemical characterization for each. For each AA-AuNPs, Raman spectra were collected and preprocessed through baseline correction and vector normalization to minimize background noise and ensure consistency across samples (Figure 1). To reduce the high dimensionality of the spectral data and enhance feature separation, we applied Principal Component Analysis (PCA) for variance retention, Linear Discriminant Analysis (LDA) for supervised class separation, PCA+LDA for sequential variance filtering and discrimination, and t-distributed Stochastic Neighbor Embedding (t-SNE) for non-linear visualization. Three supervised machine learning models were trained, LDA, PCA+LDA, and PCA+K-Nearest Neighbors (K-NN), and evaluated using cross-validation. We further conducted a single-blind study using four unlabeled spectra derived from the AA-AuNPs, evaluating the models' ability to accurately predict their corresponding amino acid classes based on learned spectral features. We also performed Ratiometric classification by combining spectra from two AA-AuNPs, Glutamic Acid and Serine to evaluate the models’ ability to distinguish between closely related spectral classes in a binary setting.

Results, Conclusions, and Discussions:: Despite the high dimensionality of the data, AA-AuNPs could be completely distinguished and grouped into separate, well-defined clusters reflecting unique spectral fingerprints using PCA and PCA+LDA, two visualization techniques that reduce noise and highlight meaningful differences (Figure 2b). All three machine learning models achieved 100% accuracy on labeled Train-Test data sets (Figure 2c-d). In the single-blind test: LDA achieved 82.50% accuracy, PCA+LDA achieved 77.50% accuracy, and PCA+K-NN performed best at 83.50% accuracy (Figure 2e-f). PCA+K-NN model underscores the advantage of combining variance filtering with instance-based classification, particularly in distinguishing spectrally similar amino acid signatures. In the Ratiometric analysis of Glutamic acid and Serine AA-AuNPs, PCA+LDA achieved strong class separation, while t-SNE provided the most distinct visual clustering due to its non-linear embedding.

We developed a reliable machine learning pipeline to classify AA-AuNPs SERS spectra with a high degree of accuracy. The integration of dimensionality reduction and supervised learning was essential for robust performance on both known and unknown data. Among the models evaluated, PCA+K-NN achieved the highest accuracy, offering a scalable and interpretable solution for biomarker identification in unknown samples. Future work will extend this approach to more complex biological targets such as proteins, bacterial strains, and patient-derived exosomes which carry disease-related molecular biomarkers. Down the line, we intend to further explore advanced ML methods and expand our repertoire involving Support Vector Machines (SVMs), Random Forests, and deep learning models like Convolutional Neural Networks (CNNs), to enhance diagnostic robustness and adaptability. These developments will help enable real-time, AI-driven SERS diagnostics in personalized medicine.

Acknowledgements and/or References (Optional):: We thank Texas Tech University for the financial support.