Construction And Application Of Machine Learning-Based Multichannel Surface-Enhanced Raman Spectroscopy

The rapid identification and analysis of drug mechanisms play an extremely important role in expediting drug discovery and development.Cellular phenotypic profiling is a screening approach that relies on gene signatures and cell imaging.It can be used to monitor cellular responses induced by drug candidates,providing important evidence for the determination of drug targets and the exploration of drug mechanisms.These methods require multi-step processing of cells to extract the intracellular biomarkers,thus limiting their widespread application in rapid screening of drug candidates.Inspired by sensory organs,scientists have developed pattern-recognitionbased sensor arrays using nanoparticle-fluorescent protein conjugates.These sensor arrays could enable to identify physicochemical changes on drug-treated cell surfaces based on the “artificial nose” approach.After interacting with drug-treated cells,these sensor arrays could produce characteristic fluorescent signals which can be regarded as references for categorizing drugs with diverse mechanisms.Nevertheless,the indirect measurements and limited number of output channels make these pattern recognition sensor arrays difficult to accurately distinguish subtle differences of cell surface phenotypes in highly heterogeneous and complex biological systems.Label-free surface-enhanced Raman spectroscopy(SERS)can directly fingerint the unique physicochemical properties of biomolecules in complex biological systems without using external pre-labeling or pre-existing biomolecules.Recently,selfassembled monolayers(SAMs)have been introduced into label-free SERS sensors.Here,SAMs could function as the mildly selective interface layer to regulate the interactions between the SERS substrate and test analytes at the interface,thus increasing the dimensionality of output data and improving the identification accuracy to enable discrimination of closely related biological samples.In Chapter 2 of this thesis,we elaborately selected multiple SAMs to functionalize label-free SERS active substrates and successfully constructed a multi-channel SERS sensor array,which can directly high dimensionality fingerprint anticancer drug-induced molecular changes on cellular surfaces.The high-dimensional fingerprint spectra were applied to gain insight on the rich information of cell phenotype from multiple dimensions.This SERS sensor consists of stellate-like fractal gold nanoparticles(Au NPs)that generate unique SERS signatures upon interacting with biomolecules in the tested samples(i.e.cells,blood,urine,et.al.).Eight types of SAMs with different molecular characteristics(such as different end functional groups and carbon chain lengths)were applied to functionalize the SERS substrate,thus generating multidimensional spectral datasets from a single measurement.In Chapter 3,based on the constructed multi-channel SERS sensor array and the powerful convolutional neural network(CNN)model,high-dimensional fingerprint spectra were recognized and analyzed by CNN to accurately predict unknown mechanism of anticancer drugs action,providing an efficient analysis strategy for phenotype-based drug screening.We built a CNN model according to multidimensional SERS fingerprint datasets.This CNN has the ability to extract feature information of the fingerprint spectra and learn by itself,seeking the nonlinear relationship between characteristic peaks and mechanisms of drug action.Due to quantitatively interpret the rich information present within the highly overlapped signatures,the CNN analysis with the spectral input in the matrix form could thus better benefit from the full data distribution and provide an improved identification accuracy(～98%).In summary,we have developed a deep learning-based multichannel SERS platform that can enable high-precision rapid identification of cancer drug mechanisms.This strategy of using SERS sensor array and CNN models not only speeds up the discovery of new drugs based on phenotype analysis but also provides a powerful platform for developing new clinical therapies.