As new infectious diseases emerge and spread, one of the best shots against novel pathogens is finding new medicines or vaccines.
But before drugs can be used as potential cures, they have to be painstakingly screened for composition, safety and purity, among other things.
Thus, there is an increasing demand for technologies that can characterize chemical compounds quickly and in real time.
Addressing this unmet need, researchers at Texas A&M University have now invented a new technology that can drastically downsize the apparatus used for Raman spectroscopy, a well-known technique that uses light to identify the molecular makeup of compounds.
“Raman benchtop setups can be up to a meter long depending on the level of spectroscopic resolution needed,” said Dr. Pao-Tai Lin, assistant professor in the Department of Electrical and Computer Engineering and the Department of Materials Science and Engineering.
“We have designed a system that can potentially replace these bulky benchtops with a tiny photonic chip that can snugly fit within the tip of a finger.”
In addition, Lin said that their innovative photonic device is also capable of high-throughput, real-time chemical characterization and despite its size, is at least 10 times more sensitive than conventional benchtop Raman spectroscopy systems.
A description of their study is in the May issue of the journal Analytical Chemistry.
The basis of Raman spectroscopy is the scattering of light by molecules. When hit by light of a certain frequency, molecules perform a dance, rotating and vibrating upon absorbing the energy from the incident beam.
When they lose their excess energy, molecules emit a lower-energy light, which is characteristic of their shape and size. This scattered light, known as the Raman spectra, contains the fingerprints of the molecules within a sample.
Typical benchtops for Raman spectroscopy contain an assortment of optical instruments, including lenses and gratings, for manipulating light.
These “free-space” optical components take a lot of space and are a barrier for many applications where chemical sensing is required within tiny spaces or locations that are hard to reach. Also, benchtops can be prohibitive for real-time chemical characterization.
As an alternative to traditional lab-based benchtop systems, Lin and his team turned to tube-like conduits, called waveguides, that can transport light with very little loss of energy.
While many materials can be used to make ultrathin waveguides, the researchers chose a material called aluminum nitride since it produces a low Raman background signal and is less likely to interfere with the Raman signal coming from a test sample of interest.
To create the optical waveguide, the researchers employed a technique used by industry for drawing circuit patterns on silicon wafers. First, using ultraviolet light, they spun a light-sensitive material, called NR9, onto a surface made of silica.
Next, by using ionized gas molecules, they bombarded and coated aluminum nitride along the pattern formed by the NR9. Finally, they washed the assembly with acetone, leaving behind an aluminum waveguide that was just tens of microns in diameter.
For testing the optical waveguide as a Raman sensor, the research team transported a laser beam through the aluminum nitride waveguide and illuminated a test sample containing a mixture of organic molecules.
Upon examining the scattered light, the researchers found that they could discern each type of molecule within the sample based on the Raman spectra and with a sensitivity of at least 10 times more than traditional Raman benchtops.
Lin noted since their optical waveguides have very fine width, many of them can be loaded onto a single photonic chip. This architecture, he said, is very conducive to high-throughput, real-time chemical sensing needed for drug development.
“Our optical waveguide design provides a novel platform for monitoring the chemical composition of compounds quickly, reliably and continuously. Also, these waveguides can be easily manufactured at an industrial scale by leveraging the already existing techniques to make semiconductor devices,” said Lin.
“This technology, we believe, has a direct benefit for not just pharmaceutical industries but even for other industries, like petroleum, where our sensors can be put along underground pipes to monitor the composition of hydrocarbons.”
Raman spectroscopy is a vibrational spectroscopy technique that is used for the assessment of the chemical composition of samples. Even complex biological samples can be analyzed in a non-destructive and label-free manner and classified using their specific molecular fingerprints assessed by this method1–4.
However, intense and strongly varying backgrounds, e.g. due to autofluorescence (with or without photobleaching), detector etaloning effects and ambient light, are an often occurring challenge in Raman spectroscopy. If the Raman intensity is too low relative to the background intensity, Raman bands are hard to discern or are masked completely.
Although it cannot be excluded that an autofluorescence background contains useful information, background correction procedures are state-of-art in Raman spectroscopy of biological material.
Autofluoerescence contributions in Raman spectra seem to be sensitive, but specificity might be problematic due to bleaching and quenching effects, which are prone to variations and lack proper reproducibility.
Therefore, different approaches were suggested to tackle the challenge of intense and varying backgrounds. One option for autofluorescence is the destruction of the fluorophores by photobleaching5–7, which is rather time-consuming and might cause side effects, such as sample contamination with chemiphotobleaching agents or thermal stress due to extended laser exposure of the sample.
A review divided other techniques roughly into two groups: computational and instrumental background correction methods8. Examples of typically used computational background correction algorithms are extended multiplicative signal correction9,10 (EMSC), multiplicative signal correction11 (MSC), rubberband12,13, sensitive nonlinear iterative peak14 (SNIP), and polynomial fittings15.
These approaches often require high computational effort and need experienced personal for the data analysis. Cordero et al. corrected a high fluorescence background in Raman spectra of bladder biopsies using EMSC16. For in vivo Raman spectra of colorectal tissue Bergholt et al. corrected autofluorescence background by a high-order polynomial fitting17.
The in vivo acquired Raman spectra of brain cancer by Desroches et al. were also background corrected using a polynomial18. Galli et al. found a high fluorescence background in Raman spectra of brain biopsies, where 88.4–96.5% of the collected intensities were attributed to fluorescence.
For separating the background-free Raman signal and the fluorescence profile a baseline estimation toolkit was used. The authors concluded that the classification was best, when both information were used19.
Instrumental background correction methods for fluorescence rejection are time-gating approaches, where the fast Raman scattering is detected before the slower fluorescence emission, and phase or wavelength modulated techniques, where the Raman scattering changes according to the wavelength or phase modulation whereas the fluorescence emission does not8. Another promising method is shifted excitation Raman difference spectroscopy (SERDS)20.
SERDS belongs to the instrumental background correction methods, which uses two slightly shifted excitation wavelengths to acquire two Raman spectra consecutively at the same lateral position.
The shift in excitation wavelength is chosen to be only a few nanometers, leading to two slightly shifted Raman spectra with the same fluorescence background profile, since the same fluorophores are excited. After subtraction of the shifted Raman spectra from each other, the resulting difference spectrum is ideally free of background contributions and only contains Raman information.
Furthermore, other constant spectral contributions such as ambient light or the system transfer function (e.g. detector etaloning effects) can be suppressed21. Proof of principle studies demonstrated SERDS using several combinations of solvents and dyes as model analytes, especially for the introduction of new lasers with two or more excitation wavelengths20,22–25.
Sowoidnich and Kronfeldt analyzed different laser wavelengths for SERDS experiments on parts of beef and pork tissues like fat, connective tissue, bone and meat26. Noack et al. conducted SERDS measurements to measure algae cultivation samples and monitor sulfated exopolysaccharides (EPS) concentrations in the reactors. For this, 10 raw spectra were averaged, smoothed and baseline corrected before the subtraction.
A principle component analysis (PCA) and different regression models were then applied to the smoothed difference spectra to determine the EPS concentration, which worked poorly for the partial least squares regression (PLSR) model, but very well for the support vector regression (SVR) model27.
Martins et al. studied molar teeth ex vivo and human skin in vivo using SERDS with an excitation wavelength of 830 nm/830.5 nm and regular Raman spectroscopy at 1,064 nm as a control. For data analysis the difference spectra were integrated to reconstruct the Raman spectrum28. Gebrekidan et al. measured difference spectra of pig tissue (bone, fat, gland and mucosal).
After a sophisticated data processing including normalization, first baseline correction, reconstruction and second baseline correction to receive fluorescence-free pure Raman spectra, a classification by PCA was performed29. By measuring a plate of clear polystyrene, Maiwald et al. showed that SERDS was able to filter out ambient light passing through polystyrene.
They also conducted SERDS in an orchard measuring the wax on the skin of an apple and the chlorophyll in a leaf using a handheld device with a high numerical aperture30,31. Schmälzlin et al. obtained SERDS images from different samples, e.g. cross-section of a pig ear, skin and a dissolving brown sugar cube using a custom-built multi-focus probe head and an integral field spectrograph.
This system was able to simultaneously detect 400 spectra delivered by the probe head of 20 × 20 pixels32.
Since photobleaching and intensity variations due to e.g. laser power and filter characteristics often result in varying background intensities, most difference spectra are not completely background-free.
This makes additional background correction steps necessary. Also reconstruction steps are usually implemented to transform the difficult to interpret difference spectra into accustomed Raman spectra. There are several reconstruction approaches, such as deconvolution, linear data manipulation, integration, kernel function, or non-negative least squares fitting33–34.
These reconstruction methods always harbor the risk of introducing artefacts into the reconstructed Raman spectrum, due to the correlation between the fixed wavelength shift and varying Raman band widths35.
As a case study pollen samples of eight different plant genera were investigated. Pollen are a valuable case study, since their Raman spectra experience intensity differences in the fluorescence backgrounds (see supplementary information in Ref.36). In palynology, pollen are taxonomically evaluated under a microscope considering their morphology.
This is time-consuming and requires a highly trained expert to differentiate several hundreds of different pollen. There are several ideas for automatization and technical improvement of this gold standard37–39. Other spectroscopic approaches like infrared absorption40–42, laser-induced breakdown43 or Raman scattering have been applied35,43–55. Raman spectroscopy was implemented to build a spectral database of pollen including a chemometrical classification by their growth habit36.
The approach in this work to differentiate several pollen genera uses difference spectra, that were obtained by novel processing of SERDS data, and lends itself as a case study to classify biological samples.
The new streamlined method to handle and classify SERDS data of biological samples is based on their single difference spectra without a reconstruction step to retrieve the familiar profile of Raman spectra or baseline correction procedures. Furthermore, spectra were reconstructed from the differences and Raman spectra were processed by state-of-art baseline correction.
For classification using difference spectra, reconstructed spectra and baseline corrected spectra as input, a PCA followed by a linear discriminant analysis (PCA-LDA) was chosen. The classification results were compared with respect to sensitivities, specificities, accuracies and precisions.
In this contribution, the methodology was described to use difference spectra based on SERDS for classification of pollen data. Its main advantage is the analysis of single difference spectra by PCA and subsequent LDA with few PCs as input without complex data pre-processing.
An optimization procedure compensated the photobleaching effects and minimized the remaining background in the difference spectra, whereas the normalization was necessary to obtain the same intensity range for all measured pollen difference spectra. This resulted in a classification based on the spectral features of the difference spectra and not based on the overall intensity of the Raman spectrum of a pollen sample.
A further improvement would be rapid, serial acquisitions of Raman spectra at both wavelengths, which suppresses photobleaching effects and avoids the optimization procedure58. Since no reconstruction of the familiar profiles of Raman spectra is necessary, possible artefacts are not introduced into the spectra.
The down side of this direct classification using SERDS spectra is the difficult spectral analysis of difference spectra and especially the resulting PC and LD loadings. The increase in noise level due to the subtraction of two spectra compared to a single Raman spectrum is compensated by PCA, which separated the spectral variations in the first PCs from the noise in the higher PCs.
The possibility to classify different pollen samples using normalized and optimized difference spectra by a linear PCA-LDA model was successfully demonstrated. In Supplementary Table S1 the average values for sensitivity, specificity, accuracy and precision for all classifications are shown that achieved good to very good results using difference spectra as input.
Since birch, hazel and alder all belong to the same family, in case of birch and alder even to the same subfamily, the pollen could even be separated on a genus level. For comparison, the reconstructed spectra and the raw spectra at 784 and 786 nm excitation after baseline correction were subjected to PCA-LDA classification.
The classification rates only show small variations with a tendency of worse results for reconstructed spectra. This demonstrates the validity of our new approach based on difference spectra. The full potential of SERDS will become evident for Raman spectra that are affected by high autofluorescence background, ambient light or etaloning effects. Since the pollen detection, the laser focusing, the wavelength shifting and the data recording was fully automated, this streamlined method could be a robust and versatile system for the automated differentiation of different pollen into their genera.
resounce link: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7343813/
More information: Megan Makela et al, Benzene Derivatives Analysis Using Aluminum Nitride Waveguide Raman Sensors, Analytical Chemistry (2020). DOI: 10.1021/acs.analchem.0c00809