Interpretability-driven deep learning for SERS-based classification of respiratory viruses

Abstract

Respiratory viruses, such as influenza A/B, RSV, SARS-CoV-2 and its variants, continue to be a major global health threat, highlighting the need for rapid and accurate variant-level diagnostics. Herein, we have developed a diagnostic platform for several respiratory viruses by integrating surface-enhanced Raman scattering (SERS) signals from three-dimensional (3D) plasmonic nanopillar substrates with interpretability-driven deep learning. The 3D plasmonic nanopillar array enables robust and reproducible capture of viral components, enhancing the SERS signal for virus-specific molecular fingerprinting. A one-dimensional convolutional neural network (1D-CNN) has been trained on SERS spectra from 13 respiratory virus types, including SARS-CoV-2 variants and sublineages, achieving over 98 % classification accuracy. To further improve model transparency, gradient-weighted class activation mapping (Grad-CAM) has been applied, revealing consistent Raman shift regions critical for virus discrimination across various media conditions. The platform has demonstrated reliable performance even in complex clinical samples, confirming its applicability for real-world diagnostics. The present approach offers a scalable and label-free solution for rapid virus detection, with potential for point-of-care applications and epidemiological surveillance.