Detection of Carbon Nanotubes in Bovine Raw Milk Through Fourier Transform-Raman Spectroscopy

The potential use of carbon-based methodologies for drug delivery and reproductive biology in cows rises concerns about residues in milk and food safety. This study aimed to assess the potential of FT-Raman spectroscopy and PLS-DA to detect functionalized multi-walled carbon nanotubes (MWCNT) in bovine raw milk. Oxidized MWCNT were diluted in milk at different concentrations from 25.00 to 0.01 µg/mL. Raman spectroscopy measurements and discriminant analysis using partial least squares (PLS-DA) were performed to identify low concentrations of MWCNT in milk samples. PLS-DA model was characterized by the analysis of the variable importance in projection (VIP) scores. All the training samples were correctly classified by the model, resulting in no false-positive or false-negative classifications. While for test samples, only one false-negative result was observed, for 0.01 µg/mL MWCNT dilution. The association between Raman spectroscopy and PLS-DA was able to identify MWCNT diluted in milk samples up to 0.1 µg/mL. The PLS-DA model was built and validated using a set of test samples and spectrally interpreted based on the highest VIP scores. This allowed the identification of the vibrational modes associated with the D and G bands of MWCNT, as well as the milk bands, which were the most important variables in this analysis.


INTRODUCTION
Carbon nanotubes (CNT) are carbon-based nanostructures composed by one or more concentrically rolled graphene sheets, with diameter ranging at the nanometer scale and length achieving few micrometers (Ijima et al., 1991).Because of its relatively high surface area and the possibility to add functional groups over its surface, CNT can be solubilized in water and easily interact with other biological molecules.So, functionalized CNTcan be employed in many fields of science as biosensor, tissue engineering, drug delivery, cancer, and gene therapy (Liu et al., 2008;Varkouhi et al., 2011;Shimizu et al., 2012).Therefore, CNT represent an interesting alternative to improve diagnostic systems (Pourasl et al., 2014;Lee et al., 2018), regenerative medicine (Newman et al., 2013;Lekshmi et al., 2020), and conventional vectors for drug delivery in human and animals (Liu et al., 2008;Ladeira et al., 2010).CNT have been employed as transfection agent for small interfering RNA into mice and human cell lines for in vitro gene silencing (Ladeira et al., 2010;Apartsin et al., 2014).In the reproductive biology field, nanomaterials and CNT have been employed in chronic disease treatment (Chaudhury et al., 2013), support for assisted reproduction techniques (Barkalina et al., 2014) and bovine embryogenesis (Munk et al., 2016).
However, the application of CNT-based methodologies in farm animals, especially in dairy cows, rises concerns about food safety, as milk represents an excretion route for drugs residues in lactating cows.Currently, one of the major problems faced by worldwide dairy industry is the detection of drug residues in raw milk, due to the indiscriminate use of antibiotics and other drugs in dairy herds without veterinary assistance and no respect to the withholding period (Bando et al., 2009;Redding et al., 2014;Tempini et al., 2018).This behavior is associated with the increasing microbial resistance to antibiotics, which impacts directly in veterinary and human health (Chen et al., 2019).
Studies evaluating the in vivo pharmacokinetics of functionalized CNT (Ali-Boucetta and Kostarelos, 2013;Rodriguez-Yañes et al., 2013;Jacobsen et al., 2017) demonstrated that nanotube injected intravenously in mice mainly accumulated in liver and spleen, and lesser amounts in lung and kidneys, followed by a high excretion through urine.By intraperitoneal route, functionalized CNT accumulated in stomach, kidney, bone, blood, spleen, and liver.Oral administration of functionalized CNT in rats results in accumulation in stomach, upper and lower intestines, with higher excretion through feces (Ali-Boucetta and Kostarelos, 2013;Rodriguez-Yañes et al., 2013;Jacobsen et al., 2017).Studies with in vitro cell culture using human and mice cell lines, demonstrated some cytotoxic effect when CNT were used in higher concentrations (5 to 10 mg/ mL) (Pantarotto et al., 2004;Bianco et al., 2005), while smaller ones (20 to 100 µg/mL) had no significant cell damage or toxicity (Singh et al., 2006;Bottini et al., 2006;Firme III and Bandaru, 2010).Up to date, no study has been conducted to assess the accumulation of functionalized CNT in udder and its elimination through milk.
Fourier transformer Raman spectroscopy (FT-Raman) was used in in vitro studies to assess the nanotube toxicity in mice and zebrafish embryos (Dal Basco et al., 2015;Girardi et al., 2016).The spectral signature of carbon nanotubes can be directly observed inside the tissues and toxicological effects have been investigated regarding the nanotubes biodistribution.In the field of milk safety, recent studies employed FT-Raman allied to chemometrics tools to analyze milk quality and the presence of adulterants like whey and starch (Almeida et al., 2012;Mazurek et al., 2015;Rodrigues Júnior et al., 2016).Raman spectroscopy is based on the detection of electromagnetic energy that is inelastically scattered by molecular vibrations after excitation by a laser.It is a fast analytical tool, which access the vibrational characteristics of molecules and discriminate, in complexes samples, more than one organic and inorganic components at the same time, without previous sample preparation or sample destruction (Almeida et al., 2011;El-Abassy et al., 2011).The association between Raman spectroscopy and the supervised method PLS-DA (partial least square for discriminant analysis) showed a high potential of use in dairy industry for quality assessment of the main nutrients (fats, proteins and carbohydrates) and detect the presence of adulterants in dairy products (Oliveira et al., 2016;Rodrigues Júnior et al., 2016;Genis et al., 2021).
The growing application of nanotechnologies in disease therapy and food technology rises a worldwide concern about biocompatibility, toxicity, food safety and residual contamination with nanoparticles.In the literature, studies developed analytical methodologies for detection of carbon nanotubes in plant tissues.Das et al. (2018a) developed a digestion method coupled with programmed thermal analysis to quantify multiwalled carbon nanotubes (MWCNT) in lettuce tissues.The methodology presented a detection limit equal to 64.9 µg of CNT-C/g of plant tissues.The same research group (Das, et al., 2018b) proposed the use of the digestion method coupled with Raman spectroscopy to detect MWCNT in lettuce.The authors were able to detect functionalized MWCNT in lettuce leaves grown with 5, 10 and 20 mg/L of the nanomaterial.
Based on ethical guidelines for the use of animals in science, lack of research focused on unrevealing the pharmacokinetics and toxicity of CNT in livestock animals, reservations about food safety due to the increasing use of nanomaterials in veterinary medicine and its future application in dairy cows, to our knowledge, this is the first exploratory study that aimed to develop an alternative fast and non-expensive screening methodology using FT-Raman combined with chemometric tool for detection of MWCNT in bovine raw milk.Accordingly, the goal of this study was to assess the potential of FT-Raman spectroscopy and PLS-DA method to detect functionalized MWCNT in bovine raw milk.The developed model was validated with an independent test set and the estimation of appropriate figures of merit.In addition, we developed a quantitative model employing PLS (Partial Least Square regression) to show the potentiality of Raman spectroscopy in detecting and quantifying MWCNT in raw milk samples.

Multi-walled carbon nanotubes
The MWCNT were produced by chemical vapor deposition (CVD) and their dimensions ranged from 10 to 25 nm of diameter and 5 to 30 µm of length.Pristine MWCNT were functionalized by oxidation using reflux with nitric and sulfuric acids, and purified according to Pacheco et al. (2015), using ethyl alcohol.Then, nanotubes were diluted in deionized water at 1 mg/mL and submitted to sonication in cold bath for 60 min, using ultrasonic processor with tapered horn tip.Nanotubes were dried in air at 80°C overnight.All the additional information about MWCNT characterization for experimental use is presented in supplementary information (http: / / osf .io/s5eq6).

Raman spectra acquisition on raw milk and MWCNT dispersions in milk samples
To avoid factors that interfere in milk composition (Linn, 1988), raw milk samples were harvested from 3 crossbred (3/4 Holstein x Zebu) cows at 65 to 92 d in milk (DIM), with an average yield at 29.04 Kg.Cows were maintained at a dairy farm in Minas Gerais province, under same health and nutrition management.

Nunes et al.: DETECTION OF CARBON NANOTUBES…
The selected animals had no clinical sings of udder inflammation and were not under recombinant bovine somatotropin treatment.Milk samples were harvested within a period of 21 d, one sample per week.Oxidized MWCNT (1.00 mg/mL) were dispersed in all milk samples, at 9 different concentrations, namely 25.00, 20.00, 15.00, 10.00, 7.50, 5.00, 1.0, 0.10, and 0.01 µg/ mL.These concentrations were chosen accordingly to in vitro studies assessing transfection efficiency of nanotubes over mammal's cell lines (Ladeira et al., 2010) and concentrations associated with minimal or no toxicity effect (Firme III and Bandaru, 2010).Raw milk and MWCNT dispersed in milk samples were aliquoted in 2 mL glass vials and sonicated in ultrasonic bath for 10 min.Before analysis, samples were homogenized for one minute by inversion and 15 s by gentle vortex.The macroscopic aspect of milk solutions was observed just before and after Raman spectra acquisition to evaluate phase segregation.Raman spectra of the samples were collected on a Vertex 70 spectrometer coupled to a FT-Raman module with a liquid nitrogen germanium detector and 1064 nm Nd: YAG laser (Bruker, USA).Laser excitation light, with 1000 mW power, was introduced and focused on the sample and the backscattered radiation was collected.The spectrum was collected after 1000 scans and with a resolution of 4 cm −1 over the range of 500 to 3100 cm −1 .Raman spectrum of isopropanol was acquired as analytical standard.The software OPUS TM (Bruker, USA) was used for FT-Raman data acquisition and spectral preprocessing.

Data analysis
To identify the presence of low concentrations of MWCNTs dispersed in milk samples, discriminant analysis using partial least squares (PLS-DA) was performed.The discriminant analysis method was selected for this study due to the nature of the samples and the high variance of control samples, therefore discriminant methods for classification are more indicated (Nunes, et al., 2020).
PLS-DA is based on PLS regression, which correlates independent spectral variables (X-matrix) with a vector of dependent dummy variables (y-vector).The PLS-DA model was built using MATLAB version 8.4 (MathWorks, Natick, USA) and PLS Toolbox version 7.0 (Eigenvector Technologies, Manson, USA).A dummy vector y was created with 0 and 1 representing bovine raw milk samples (control milk) and milk dispersed MWCNT samples, respectively.Since PLS-DA provides predicted y values that are not exactly 0 or 1, a Bayesian threshold was calculated.
The Raman spectra were preprocessed using the standard normal variate method (SNV) for the correction of baseline deviations.Savitzky-Golay smoothing, with a window of 9 points, and mean centering were also performed.The spectral data matrix was split into 28 samples (12 bovine raw milk and 16 MWCNT dispersed in milk) for the training set and 17 samples (7 bovine raw milk and 10 MWCNT dispersed in milk) for the test set.This was done using the Kennard and Stone algorithm (Kennard & Stone, 1969) The concentration of MWCNTs in milk ranged from 0.01 to 25.00 µg/mL.The training set was used to build the model, while the test set assessed its performance.The number of latent variables (LVs) was selected using Venetian blinds cross-validation with 5 splits, based on the smallest cross-validation classification error.The performance of PLS-DA model was evaluated using the parameters of sensitivity, specificity, and efficiency.Sensitivity is the ability of a model to detect truly positive samples as positive (MWCNT-positive).Specificity is the ability of a method to detect truly negative samples as negative (MWCNT-negative), and efficiency is a global figure of merit that is estimated as the difference between the sum of results (100%) and the sum of false-positive and false-negative rates.Table 1 presents the equations used to calculate the figures of merit.
The PLS-DA model was characterized by analyzing the variable importance in projection (VIP) scores.The VIP scores indicate the significance of specific variables in the discriminant ability of the model.When the VIP value is greater than one, the variable is considered important in distinguishing a positive sample from a negative one.
For constructing the PLS model, 27 samples were used, 3 samples of milk without the addition of MW-CNT, and 24 samples with MWCNT dispersed in the concentration range of 0.01 to 25.00 µg/mL.Samples were manually divided into calibration and validation sets.For the calibration samples, milk and MWCNT diluted in milk samples, corresponding to the entire spectral range of 0.00 to 25.00 µg/mL, were selected.The pre-processing was the same used in the PLS-DA model.The optimum number of latent variables was chosen by the root mean square error of cross-validation (RMSECV) obtained from the calibration set by internal validation (Random subsets method).The performance of the models was evaluated by the root mean square error of calibration (RMSEC), root mean square error of prediction (RMSEP) and residual prediction deviation (RPD).The characterization of the model was done through the analysis of the VIP scores.

RESULTS AND DISCUSSION
Immediately after removing the milk samples from spectrometer, no phase segregation such as the super-Nunes et al.: DETECTION OF CARBON NANOTUBES… natant milk fat ring was observed in any of the samples evaluated, indicating macroscopic stability during measurement.Raman spectra for raw milk and milk containing MWCNT dilutions from 25.00, to 0.01 µg/mL are presented in Figure 1.The Raman spectrum for raw milk, has scattering bands at 2931,2856,1655,1450,1261,1122,1084 and 1003 cm −1 .For milk samples with MWCNT, the Raman spectra have scattering bands at 2931, 2856, 1450, 1122, 1084 and 1003 cm −1 , and they are similar to those observed in milk spectrum.Scattering bands at 1605 and 1286 cm −1 presented high intensities in the lower nanotube dilution in milk, while signals decreased as the nanotube dilution increases.These 2 bands are associated, respectively, with the G and D bands of MWCNT, and their presence in the spectrum is clearly related to the presence of standardized MWCNT dispersed in the milk.Conversely, scattering band at 1655 cm −1 presented a very small intensity in 25.00 µg/mL dilution, while its intensity increases as nanotube dilution also increases.This occur because the 1655 cm −1 is a well-established scattering band of milk spectrum associated with proteins and fatty acids (Almeida et al., 2011;El-Abassy et al., 2011;Rodrigues Júnior et al., 2016).When the dilution of MWCNT in milk is low, this band signal is suppressed by the high intensity signal of the G band.As the nanotube dilution in milk increases, the 1655 cm −1 band signal becomes more evident in the Raman spectrum.
The Raman spectrum of milk observed in this study agrees with studies that also employed cow's liquid milk (El-Abassy et al., 2011) and whole milk powder (Almeida et al., 2011;Rodrigues Júnior et al., 2016).The bands at 2931 and 2856 cm −1 are related to the asymmetric (ν ass ) and symmetric (ν s ) CH 2 stretching modes, respectively.These (C-H) stretch is related to the milk's fatty acids content, and the prominent intensity of the symmetric CH 2 stretch band at 2853 cm −1 are characteristic of milk with higher fat content.The band at 1655 cm −1 correspond to association between ν(C = O) stretching of amide I mode from milk pro-teins and ν(C = C) cis double bond stretching from the unsaturated fatty acids.The band observed at 1450 cm −1 are correlated to the carbohydrate content of milk and represents the δ(CH 2 ) scissoring CH 2 deformation.The bands observed in milk under the 1280 cm −1 are related to carbohydrates.The band at 1261 cm −1 correspond to the γ(CH 2 ) twisting mode, the 1122 to 1082 cm −1 correspond to associations between ν(C-O) and ν(C-C) stretching modes and γ (C-O-H) twisting modes.The band observed at 1003 cm −1 are related to the ring-breathing mode of the phenylalanine amino acid present in milk.The weak and unidentified bands under 1000 cm −1 are related to δ(C-O-C) and δ(C-C-O) scissoring modes, and the presence of glucose and lactose (Almeida et al., 2011;El-Abassy et al., 2011;Rodrigues Júnior et al., 2016).
The MWCNT has vibrational characteristics that can also be assessed by Raman scattering spectroscopy.Beyond milk scattering bands, the 2 most important bands in the spectrum of milk samples with MWCNT are identified in 1605 and 1286 cm −1 wavelengths.The first represents the G band, resulted from the ν(C = C) tangential stretching modes of the MWCNT (Jorio et al., 2004;Bokobza and Zhang, 2012).The second one represents the D band, resulted from the aromatic ring breathing modes of the MWCNT, induced by disorders on the hexagonal lattice.This D scattering band is also induced by surface modifications due to the addition of carboxylic and phenolic functional groups over MW-CNT during oxidation process (Sato et al., 2005;Sekar et al., 2015).
A PLS-DA model was built to identify the presence of MWCNTs in milk samples, especially at low concentrations, since at higher concentrations visual detection is possible.As previously described in Section 2.3 of material and methods, the model was constructed using 28 samples with 5 LV, accounting for 97.6% and 95.1% of the total variance in the X-matrix and yvector, respectively.The number of latent variables was selected according to the smallest misclassification in cross-validation, as shown in Figure S3 (Supplementary Information).The model with 5 latent variables showed a sensitivity of 100% and specificity of 93.8% and a classification error rate of 3.1% during cross-validation.
Figure 2 shows the PLS-DA predictions for the training and test sets.The Bayesian threshold was calculated as 0.476; milk samples below this value were classified as MWCNT-negative, while those above were classified as MWCNT-positive.
All the training samples were correctly classified by the model, resulting in no false-positive or false-negative classifications (Figure 2).In the test set, only one falsenegative result was observed.The misclassified sample had an MWCNT concentration of 0.01 µg/mL, indicating that it is not possible to carry out identification using ordinary Raman spectroscopy for concentrations equal to or less than this value.The PLS-DA model can correctly classify milk samples with up to 0.1 µg/mL MWCNTs.However, this exploratory research did not assess the capability of the PLS-DA model to classify milk with dilutions of MWCNT between 0.09 to 0.02 µg/mL, which requires more investigation.
The figures of merit were calculated to demonstrate the ability of the PLS-DA model in detecting milk samples with MWCNTs.The sensitivity was 100% and 90% for the training and test sets, respectively.The specificity was 100% for the training and test sets, and the efficiency rate was estimated to be 100% for the training set and 90% for the test set.These results are summarized in Table 1.
For the spectral characterization of the developed model, it is interesting to observe the VIP scores shown in Figure 3. Variables with VIP scores higher than 1.0 are considered to contribute significantly to the detection of milk samples containing MWCNTs in the model.Analysis of these scores reveal that the D band (1290 cm −1 ) is the highest contributor.The remaining notable VIP score bands (1356, 1449, 1657, and 1757 cm −1 ) are associated with vibrational modes characteristic of milk (Almeida et al., 2011;El-Abassy et al., 2011;Rodrigues Júnior et al., 2016;Goméz-Mascaraque et al., 2020, Genis et al., 2021).Out of these scores, the second most intense band at 1449 cm −1 may be associated with the carbohydrate and fat content of milk, which represents δ(CH 2 ) scissoring CH 2 deformation (Almeida et al., 2011;Goméz-Mascaraque et al., 2020).The third most intense band at ~1650 cm −1 is assigned to the amide I mode from milk proteins (Almeida et al., 2011;Rodrigues Júnior et al., 2016).The Raman band at 1757 cm −1 corresponds to the vibrational mode associated with the presence of fatty acids (C = O stretching ester), and the band at 1356 cm −1 (C-O stretching and C-O-H deformation) is attributed to the presence of carbohydrates in milk samples (Almeida et al., 2011;El-Abassy et al., 2011;Goméz-Mascaraque et al., 2020).Finally, the fourth highest score at 1595 cm −1 represents the G band of MWCNT.To verify a possible interference between MWCNT and milk components, a PCA model was built with only milk samples (without the addition of MWCNT).The contribution of milk Raman bands in VIP scores is due to changes in milk composition from sample to sample, as can be seen in the scores and loadings of the PCA model (Figure S4 -Supplementary Information).
After building the classification model to detect the presence of MWCNT in milk samples, a regression model was built to predict the amount of MWCNT.The PLS model was built with 3 latent variables (LV), and the number of LV was selected according to the smallest RMSECV (1.92 µg/mL), as shown in Figure 4a.The results for the PLS model, including root- mean-square errors of calibration (RMSEC) and prediction (RMSEP) and RPD, are shown in Table 2.The accuracy of the model can be evaluated by the RMSEC and the RMSEP, which for the model built are 1.53 and 1.99 µg/mL, respectively.The prevision relative errors are lower than 18%, and for samples with low concentrations, the relative error was greater.Another parameter evaluated in the model was the RPD (Williams, 2001).Models with RPD above 2.4 present good predictive ability.The PLS model presented an RPD of 5.7 and 4.5 for the calibration and validation sets, respectively.Figure 4b

CONCLUSION
This exploratory study proposes a simple and rapid method based on Raman spectroscopy and chemometrics tools to detect and quantify concentrations of MW-CNT in cow milk equal to greater than 0.1 µg/mL.The chemometric models were built and validated using a set of test samples and spectrally interpreted based on the highest VIP scores.This allowed the identification of the vibrational modes associated with the D and G bands of MWCNT and characteristic milk bands.This methodology requires more research and has the potential to be implemented by dairy industry as a screening method for detecting carbon nanotubes in milk intended for processing, as a food safety practice while the use of nanomaterials in veterinary medicine is growing.

Figure 1 .
Figure 1.Raman spectra of raw milk (top spectrum) and MWCNT dispersions in milk, from 0.01 µg/mL to 25.00 µg/mL (bottom spectrum).The concentration values in µg/mL are shown for each spectrum.The doted lines represent the main Raman features observed for milk (blue) and MWCNT (red).
shows a plot of the MWCNT content predicted by the PLS model.The calibration set shows a correlation of 0.968, while the validation samples showed a correlation of 0.975.As in the PLS-DA model, the informative vectors were analyzed to characterize the model and the VIP scores are shown in Figure 4c.The identification of the spectral regions that contribute more to predicting the MWCNT content was possible.As it was expected, the D band at 1286 cm −1 is the main contribution to the quantification model.The other MWCNT band at 1605 cm −1 also contributes to the model, just like the 1450 and 1665 cm −1 milk bands.The results obtained in this work show the potential of Raman spectroscopy for detecting and quantifying MWCNT in raw milk samples.

Figure 2 .
Figure 2. Results of the training and test sets of the PLS-DA model, showing positive MWCNT milk samples (*) and milk samples (▼).Black arrow indicates the only one false-negative sample of the test set, with 0.01 µg/mL MWCNT.Sensitivity, specificity and efficiency for training and test sets were 100% / 90%; 100% / 100%; and 100% / 90%, respectively.

Figure 3 .
Figure 3. Result of VIP scores distribution for spectral characterization of the PLS-DA model developed using spectra from different MWCNT dilutions in milk samples.The VIP scores (black full line) over the threshold (red dash line) contribute significantly to the model.

Figure 4 .
Figure 4. PLS model: (a) Choose of variable latent number based in lower RMSECV.(b) Plot of reference values versus predicted values in the PLS model for MWCNT concentration.(c) VIP scores showing the Raman bands contribution for the PLS model.

Table 1 .
Nunes et al.: DETECTION OF CARBON NANOTUBES… Figures of merit for PLS-DA