Decoding Hidden Features in Near-Infrared Fluorescence Spectra of Single-Walled Carbon Nanotubes via Machine Learning for Multiplexed Virus Identification

Single-walled carbon nanotubes (SWCNTs) exhibit rich spectral diversity in their near-infrared (nIR) fluorescence, offering strong potential for multiplexed optical sensing via diverse signal features, even with a single sensor. However, conventional analytical methods primarily focus on overt spectral parameters such as the absolute values and relative shifts of peak intensity and wavelength, leaving numerous subtle yet critical multispectral features largely unexamined, particularly near detection limits. In this study, we developed a systematic analytical framework leveraging machine learning to decode analyte-specific hidden multispectral features within nIR spectra that are indistinguishable by traditional analytical methods. This technique boosts both sensitivity and specificity, enabling multiplexed detection at ultralow concentrations. Using corona phase molecular recognition of three pathogenic coronaviruses as a model system, we collected SWCNT nIR emission spectra below conventional detection limits and quantitatively assessed how individual wavelengths within the 900-1400 nm window contributed specifically to distinguishing each virus. Integration of these spectral data into our optimized model enabled accurate virus classification and quantitative assessment of virus adsorption rates. Furthermore, the model facilitated the identification of unknown viruses and optimized detection timing by capturing early stage spectral variations. Importantly, the model retained strong adaptability in complex biological environments such as human serum after minimal fine-tuning. This approach demonstrates the capability to fully leverage the multispectral properties of SWCNTs, significantly advancing sensitive and precise multiplexed optical sensing.