Alzheimer's Disease Diagnosis Using SERS and Advanced Data Analysis

Early diagnosis of Alzheimer’s disease (AD) is critical for disease prevention and cure, but no method to do so has yet been developed that had the required sensitivity and specificity. Computational methods are increasingly being applied in these efforts. One such method is "deep learning." We propose here a convolutional neural network-based AD diagnosis approach using surface enhanced Raman spectroscopy (SERS) fingerprints of human cerebrospinal fluid (CSF). To realize the testing, we further prove the reliability of a SERS hybrid platform from its quantification capability and orientation dependence. Analysis using Amyloid beta (Aβ) to prove the biological feasibility and test the specificity of the platform is also done. We report results demonstrating the reproducibility and accuracy of this novel SERS data analysis platform. We have achieved 100% reproducibility in double blind experiments and 92% accuracy in disease diagnosis. Comparison of the SERS-neural network approach with single biomarker tests shows it is more accurate, thus it may have substantial value in the differential diagnosis of AD as well as other neurodegenerative disorders.We also show here that surface-enhanced Raman spectroscopy (SERS) coupled with principal component analysis (PCA) readily distinguishes small biological differences: Aβ40 and Aβ42. We show further, through comparison of assembly-dependent changes in secondary structure and morphology, that the SERS/PCA approach readily and unambiguously differentiates closely related assembly stages not readily differentiable by circular dichroism spectroscopy, electron microscopy, or other techniques.To test the substrate feature, we demonstrate, using a biologically relevant test analyte, the amyloid β-protein (Aβ), a seminal pathologic agent of Alzheimer's disease (AD), that linear relationships exist between (a) peak intensity and concentration at a single plasmonic hot spot smaller than 100 nm, and (b) frequency of hot spots with observable protein signals, i.e. the co-location of an A protein and a hot spot. We demonstrate the detection of Aβ at a concentration as low as 10-18 M after a single 20 �l aliquot of the analyte onto the hybrid platform.Orientation dependence is also proved by analyzing the standard deviation of spectral feature. The standard deviation in the intensity of individual Raman peaks diminishes for protein size larger than 13 amino acids. Secondary structure of protein (such as protein-protein interaction) remains unchanged regardless of protein orientation. Numerical simulation studies corroborate the experimental observation in that the SERS spectral features of biomedically relevant protein (of larger than 13 amino acids in size, which represent all human protein types) are not affected by the orientation of amino acids randomly dispersed on SERS-active surfaces