High-Quality-Factor Metasurfaces for Rapid Identification and Classification of Mycobacterium Tuberculosis Using Surface-Enhanced Raman spectroscopy

Rapid identification of mycobacterium tuberculosis (MTB) infection and antibiotic susceptibility testing (AST) is essential for effective patient treatment, prevention of community spread, and combating antimicrobial resistance. To date, culture-based techniques remain the gold standard for the identification of TB strains and determination of their antibiotic susceptibility. Unfortunately, such identification and AST can take several days to multiple weeks in well-equipped labs, due to the slow growth rate of MTB. One promising alternative to culture-based cellular fingerprinting is Raman spectroscopy, an optical characterization technique which probes vibrational modes of a sample. It relies upon the inelastic scattering of photons incident on a sample and measures their wavelength shift after interacting with a sample. Raman scattering offers three key advantages over alternative cellular fingerprinting approaches. First, it is label-free and generalizable to all kinds of cells; the intrinsic vibrations of the cells serve as native labels. Second, Raman can be fast, with sub-second acquisition times, and accurately measured with minimal sample preparation. Lastly, because peaks in a Raman spectrum are unique to the vibrational modes of particular biomolecules, Raman cell signatures can be used to assess and monitor the composition, structure, and vitality of that cell and uniquely differentiate it from another cell. Despite these advantages, obtaining and utilizing Raman spectra from cells presents a challenge due to (1) Raman’s inherently weak scattering efficiency and low signal-to-noise ratio and (2) its complexity in spectral interpretation.<br/><br/>To address these two issues, we present a combined nanophotonic-machine learning platform that uses resonant nanophotonic surfaces, known as high-quality factor (high-Q) dielectric metasurfaces, to provide critical Raman signal enhancement. Using FDTD simulations, we demonstrate a metasurface of subwavelength silicon nanoblocks that are biperiodic in width to generate a high Q resonance. These nanoblocks have a length of 425 nm, a biperiodic width of 105 and 95 nm, and a height of 200 nm. By tuning the nanoblocks’ dimensions, we place a high-Q resonance at the Raman excitation wavelength that will be used to enhance the local electric field amplitude as well as a broad magnetic dipole Mie resonance at the MTB’s Stokes-shifted Raman region to provide strong, broadband Raman signal enhancement . Collectively, these two resonances provide a maximum Raman signal enhancement of over 105 over a wide spectral range of 1500 cm-1.<br/><br/>We fabricate these metasurfaces using electron beam lithography and experimentally verify this enhancement by placing dried bacterial samples on the metasurfaces and measuring their Raman spectra. We use four unique strains of antibiotic-resistant Bacillus Calmette-Guérin (BCG), an attenuated form of Mycobacterium Bovis, as a model for Mycobacterium tuberculosis, as well as a control BCG strain. We classify the antibiotic resistant BCG strains according to their Raman spectra using logistic regression. Specifically, we use a 80/20 train-test split, in which 80% of the data is used to train the model and the remaining 20% is held out and tested on the model in order to evaluate the accuracy of the model. Through this machine learning model, we are able to identify and differentiate the 5 BCG strains with ~98% accuracy. Through this combined nanophotonic-machine learning platform, we have demonstrated a rapid and accurate platform for bacterial identification and classification using Raman spectroscopy.