Explainable Machine Learning to Unveil Detection Mechanisms with Au Nanoisland-Based Surface-Enhanced Raman Scattering for SARS-CoV-2 Antigen Detection

Abstract

In this study, we introduce a simplified surfaceenhanced Raman scattering (SERS) nanobiosensor for precise detection of a SARS-CoV-2 antigen, leveraging supervised machine learning approaches. The biosensor was made with Au nanoislands conjugated with a 4-aminothiophenol Raman reporter and an anti-SARS-CoV-2 antibody. Through the integration of feature selection and learning algorithms, namely, logistic regression, linear discriminant analysis, and support vector machine, we achieved high accuracies ranging from 96 to 100% in antigen detection. Furthermore, we identified the underlying detection mechanisms by employing the concept of multidimensional calibration space, which is based on decision trees and random forest algorithms. This analysis with explainable machine learning allowed us to gain insights into the reasons why our simplified nanobiosensor exhibits lower sensitivity compared with that of the previous sandwich-type immunosensors for SARS-CoV-2. The results presented here emphasize the potential of supervised machine learning in SERS biosensing, which can be applied to any type of diagnostics.