Nanostructured materials, such as silver nanorod (AgNR) arrays fabricated via oblique angle deposition, offer enhanced electromagnetic fields for surface-enhanced Raman spectroscopy (SERS), enabling sensitive molecular detection. However, many bacterial biomarkers exhibit inherently low surface affinity to silver, resulting in weak or variable spectral signatures, posing a critical limitation in SERS-based biosensing. In this study, we explore how machine learning can be harnessed to extract meaningful signals from these complex and sometimes weak interactions.
We investigated six bacterial biomarkers2,3-dihydroxybenzoic acid, 2,5-dihydroxybenzoic acid, pyocyanin, lipoteichoic acid (LTA), enterobactin, and β-carotene by depositing them on AgNR substrates across varying concentration levels. The SERS spectra were collected under consistent conditions without the use of affinity-enhancing surface modifications, to isolate the substrate–analyte interaction dynamics. We applied convolutional neural networks (CNNs) to both classify and quantify these biomarkers from raw spectral data.
Despite the low affinity of some molecules particularly Enterobactin and β-carotene for bare Ag surfaces, CNNs achieved over 99.9% classification accuracy across all six biomarkers. Regression models yielded R2 values > 0.97 and MAEs < 0.27, outperforming support vector regression in all cases. These results highlight CNNs’ capacity to learn subtle spectral patterns that may be lost using traditional feature-based methods, especially when surface interactions are weak or inconsistent.
This work demonstrates that deep learning models can effectively overcome material-level limitations in SERS sensing caused by low analyte-substrate affinity. By combining nanostructured AgNR arrays with AI - driven analysis, we offer a robust strategy for reliable biomarker detection even under suboptimal surface binding conditions, supporting the development of smarter nanomaterial-based biosensors for clinical and environmental applications.
