Nanoplasmonic Platforms for Machine Learning-enabled Sensing and Diagnostics

Abstract

Due to their unique optical properties, metallic nanoparticles have been applied to many fields including sensing and diagnostics. These nanoparticles enhance Raman scattering by a factor of 106 or higher, producing surface-enhanced Raman scattering (SERS). Nanoparticle optical and physical properties can be tuned by their shape and composition for different applications. Additionally, SERS yields sharp, fingerprinting peaks, allowing for high degrees of multiplexed detection over more common methods like fluorescence. The sensitivity and tunability of nanoplasmonic SERS-based sensing has led to their use in vitro and in vivo. This work comprises two main goals to advance SERS-based sensing and diagnostic techniques. First, machine learning (ML) tools were investigated for spectral analysis of multiplexed spectra and nanoparticle synthesis, including comparative analysis of different established models and application of best performing models to various nanosensor platforms. Second, nanoplasmonic biosensors were developed for in vivo sensing and in vivo diagnostics. Inverse molecular sentinel (iMS), a label-free miRNA biosensor, was further developed for in vivo sensing of miRNA in plant and mammalian cells. Spikey nanorattle sandwich assay and a portable device was developed for cancer diagnostics in vitro through mRNA sensing in tissue biopsy samples. SERS has narrow, fingerprinting spectral features suitable for multiplexed analysis. Traditional spectral unmixing of multiplexed SERS spectra like spectral decomposition, principal component analysis, and identification of unique peaks is limited by noise and other variability, especially at very high degrees of multiplexing. To investigate whether ML can improve over traditional tools, we compared the performances of spectral decomposition, support vector regression (SVR), random forest regression, partial least squares regression, and convolutional neural network (CNN) for SERS spectral unmixing from a mixture of 7 SERS-active nanorattles loaded with different dyes. We showed that CNN most accurately determined relative contributions of each distinct dye-loaded nanorattle. We also applied CNN for classification and quantification of multiplexed SERS spectra for GNS direct detection of poly aromatic hydrocarbons (PAHs). High dimensionality of SERS data poses a challenge for traditional ML techniques, which can be prone to overfitting and poor generalization. Thus, we integrated the dimensionality reduction technique non-negative matrix factorization (NMF) to the ML pipeline to compare the performances of common ML methods including CNN, SVR, and extreme gradient boosting combined with and without NMF for spectral unmixing of four-way multiplexed SERS spectra from iMS nanoplatform for miRNA detection. Again, CNN achieved high accuracy in spectral unmixing. NMF decreased memory and training demands without sacrificing model performance. Additionally, models were interpreted using gradient class activation maps and partial dependency plots to understand predictions. Besides spectral analysis, ML can also aid in synthesizing nanoplatforms. A two-model ML system was developed for optical feature prediction of nanostars synthesized by a low-cost, automated synthesis machine. SERS is suitable for in vivo applications due to its sensitivity, specificity, and non-destructive nature. We worked towards adapting iMS for in vivo applications in mammalian cells by adapting DNA probes and developing an endosomal escape strategy based on photothermal irradiation. MiRNAs are not only involved gene expression in animals but plants as well. Monitoring plant miRNAs are important for developing efficient and cost-effective biofuels. Traditional nucleic acid analysis such as RT-PCR require sample lysis and homogenization, resulting in loss of spatial resolution and sample destruction. We used the Inverse Molecular Sentinel (iMS) biosensor on unique silver-coated gold nanorods (AuNR@Ag) with a high-aspect ratio to penetrate plant cell walls for detecting miR397b within intact, living plant cells. SERS biosensing of nucleic acids is also advantageous for in vitro diagnostics, especially for point-of-care (POC) without access to laboratory equipment and skilled personnel. We developed a plasmonics-enhanced spikey nanorattle-based biosensor for direct surface-enhanced Raman scattering (SERS) detection of mRNA cancer biomarkers. Our method uses a sandwich hybridization approach with magnetic beads and SERS spikey nanorattles. We demonstrated the feasibility of this assay for multiplexed detection and built an enclosed device to automate the assay. A pilot study using clinical samples demonstrated the efficacy of our biosensor in distinguishing between tissue with positive or negative diagnosis for head and neck squamous cell carcinoma, highlighting its potential for rapid and sensitive cancer diagnostics in low to middle resource settings.