Exploring the Application of Surface-enhanced Raman Scattering-based Biosensing of Individual sEVs in Disease Diagnosis and Therapeutics

We use surface enhanced Raman scattering combined with machine learning to tackle cancer early diagnostics, and also deliver biomedical applications. A key experimental challenge is the inherent heterogeneity in exosome samples, the highly complex, high dimensional spectroscopic data, which is further compounded by low signal to noise ratio. To begin, pipette approximately five microliters of the small extracellular vesicle sample onto the designated plasmonic substrate.

Place the substrate in a desiccator to let the droplet dry completely for approximately 15 minutes. Then transfer the dried substrate to a confocal Raman microscope. Using the instrument software WiRE version 4.4, initiate data collection.

Set the excitation laser to 785 nanometers at five milliwatts and calibrate the system. To acquire spectra from the gold substrate, perform a scouting scan over a 300 by 300 micrometer area to locate single vesicles using static mode with a 10 micrometer step size, 0.1 second exposure, and with 50%laser power. Once the vesicles are located, conduct high resolution mapping in static mode within a five by five grid using a one micrometer step size, 0.2 second exposure per point, and 50%laser power.

Next, acquire spectra from the graphene substrate by performing the following scans. Screen the initial raw spectra and exclude any that display excessive noise or abnormal spectral shapes. Apply a Savitsk-Golay filter to each spectrum to reduce fluorescence background and smooth the spectral data.

Calculate the signal to noise ratio for each spectrum using either the peak to baseline method or the standard deviation method. For each spectrum, calculate the signal to noise ratio as the ratio of the peak intensity after baseline subtraction at a representative Raman band to the standard deviation of the noise. Estimate noise by subtracting a Savitsky-Golay smoothed version of the spectrum, using a window size of 11 and a polynomial order of three from the original spectrum.

Rank all spectra in ascending order of signal to noise ratio and evaluate the persistence of diagnostic peaks across the ranked spectra to determine the threshold. Then set the threshold at the point where any of these peaks disappear into baseline noise corresponding to a signal to noise ratio of 28. Exclude spectra that fall below the signal to noise ratio threshold of 28 to ensure only high quality spectra are retained for downstream analysis.

And normalize each remaining spectrum to either its maximum peak intensity or the total area under the curve. Assign a label to each spectral measurement based on its origin, such as a gastric cancer patient or a healthy control. Input the labeled spectral data into a support vector classifier with a linear kernel and apply a stratified shuffle split for fivefold cross validation to ensure balanced representation in each fold.

Now, average the accuracy metrics across all folds to assess overall model performance. Finally, use linear discriminant analysis, or LDA, to reduce the dimensionality of the surface enhanced Raman scattering dataset for visualization. Employ the trained linear discriminant analysis model to perform direct classification between the sample groups.

Raman spectra with the corresponding average spectra revealed distinct molecular signatures across each sample type acquired from single small extracellular vesicles isolated from tissue, blood, and saliva samples. Linear discriminant analysis revealed clear clustering of extracellular vesicle data according to their source tissue, blood, or saliva. Binary classification using a support vector machine classifier achieve the highest accuracy for tissue samples at 90.1%followed by blood at 70.9%and saliva at 60.7%LDA based classification splits clearly separated healthy and gastric cancer samples in tissue derived extracellular vesicle data.

Similar LDA based classification splits for blood and saliva derived extracellular vesicle data showed less optimal separation between healthy and cancer samples. Raman spectra of doxorubicin incubated vesicles without graphene showed consistent peaks at approximately 1, 081, 1, 206, and 1, 440 inverse centimeters. With graphene, an additional D peak at approximately 1, 350 inverse centimeters, and a G peak at approximately 1, 580 inverse centimeters were observed serving as internal standards.

The ratio of the doxorubicin peak at 442 inverse centimeters to the graphene G peak increased with higher drug concentrations and longer incubation times. We have established a single vesicle biochemical fingerprinting to distinguish disease sources showing potential for early diagnosis. Three key advantages of our technique.

One, it's highly sensitive. Two, it's non-invasive. and three, it does not require lysing, or physically breaking down the particles.

In the future, we will correlate spectral profile to the proteomic profile, which, in turn, would help to correlate exome content with specific functions, such as cell communication and the transportation.