You are on page 1of 18

Food Additives & Contaminants: Part A

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tfac20

Authentication of recycled and virgin polyethylene


terephthalate based on UPLC-Q-TOF-MS using non-
volatile organic compounds and chemometrics

Tian-Ying Hao, Qin-Bao Lin, Xue-Feng Wu, Si-Liang Wu, Huai-Ning Zhong, Ben
Dong, Zhi-Feng Chen, Zhi-Kang Ye, Zhi-Wei Wang & Xiaowen Xu

To cite this article: Tian-Ying Hao, Qin-Bao Lin, Xue-Feng Wu, Si-Liang Wu, Huai-Ning Zhong,
Ben Dong, Zhi-Feng Chen, Zhi-Kang Ye, Zhi-Wei Wang & Xiaowen Xu (2023) Authentication
of recycled and virgin polyethylene terephthalate based on UPLC-Q-TOF-MS using non-
volatile organic compounds and chemometrics, Food Additives & Contaminants: Part A, 40:8,
1114-1130, DOI: 10.1080/19440049.2023.2227732

To link to this article: https://doi.org/10.1080/19440049.2023.2227732

View supplementary material

Published online: 06 Jul 2023.

Submit your article to this journal

Article views: 182

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tfac20
FOOD ADDITIVES & CONTAMINANTS: PART A
2023, VOL. 40, NO. 8, 1114–1130
https://doi.org/10.1080/19440049.2023.2227732

Authentication of recycled and virgin polyethylene terephthalate based on


UPLC-Q-TOF-MS using non-volatile organic compounds and chemometrics
Tian-Ying Haoa, Qin-Bao Lina, Xue-Feng Wub, Si-Liang Wub, Huai-Ning Zhongb, Ben Dongb,
Zhi-Feng Chena, Zhi-Kang Yea, Zhi-Wei Wanga and Xiaowen Xua
a
Key Laboratory of Product Packaging and Logistics, Packaging Engineering Institute, College of Packaging Engineering, Jinan
University, Zhuhai, China; bNational Reference Laboratory for Food Contact Material (Guangdong), Guangzhou Customs District
Technology Center, Guangzhou, Guangdong, China

ABSTRACT ARTICLE HISTORY


Plastic packaging waste, such as polyethylene terephthalate (PET) has increased significantly Received 14 April 2023
in recent decades, arousing a considerable and serious public concern regarding the envir- Accepted 9 June 2023
onment, economy, and policy. Plastic recycling is a useful tool to mitigate this issue. Here, a
KEYWORDS
feasible study was performed to investigate the potential of a novel method for identifying
UPLC-Q-TOF-MS; chemomet-
virgin and recycled PET. Ultra-performance liquid chromatography-quadrupole time-of-flight rics; recycled polyethylene
mass spectrometry (UPLC-Q-TOF-MS) was combined with various chemometrics, as a simple terephthalate; non-volatile
and reliable method that achieved a high discrimination rate for 105 batches of virgin PET organic compounds
(v-PET) and recycled PET (r-PET) based on 202 non-volatile organic compounds (NVOCs).
Making use of orthogonal partial least-squares discrimination analysis (OPLS-DA) together
with non-parametric tests, 26 marker compounds (i.e. 12 intentionally added substances
(IAS) and 14 non-intentionally added substances (NIAS) as well as 31 marker compounds
(i.e. 11 IAS and 20 NIAS) obtained from positive and combination of positive and negative
ionization modes of UPLC-Q-TOF-MS, respectively, were successfully identified. Moreover,
100% accuracy was obtained using a decision tree (DT). Cross-discrimination based on mis-
classified samples using various chemometrics allowed the prediction accuracy to be
improved and to identify a large sample set, thus greatly enhancing the application scope
of this method. The possible origins of these detected compounds can be the plastic itself,
as well as contamination from food, medicine, pesticides, industry-related substances, and
degradation and polymerization products. As many of these compounds are toxic, especially
those pesticide related, this indicates an urgent requirement for closed loop recycling.
Overall, this analytical method provides a quick, accurate, and robust way to distinguish vir-
gin from recycled PET and thus addresses the issue of potential virgin PET adulteration
thereby detecting fraud in the area of PET recycling.

Introduction environment, causing a serious plastic pollution


Polyethylene terephthalate (PET) is a polyester that can have a negative impact on ecology and
compound possessing outstanding physiochem- even human health (Palm et al. 2022). Thus,
ical properties, allowing its usage in various recycling is considered as one of the most effi-
applications (Awaja and Pavel 2005). Since the cient ways to reduce the quantity of PET waste,
1980s, PET has been broadly used in preparing which is in line with the circular economy prior-
textile fibers, and more recently widely employed ities proposed by many countries (EFSA Panel on
in food and beverage packaging (Ritchie and Food Contact Materials 2011; US FDA 2021).
Roser 2018). However, PET bottles are disposable Since in China virgin plastic is much more
plastic materials that can be discarded into waste expensive than plastic waste, plastic packaging
rapidly with a high volume ending up in the materials have the possibility to be adulterated

CONTACT Xiaowen Xu xiaowenxu@jnu.edu.cn Key Laboratory of Product Packaging and Logistics, Packaging Engineering Institute, College of
Packaging Engineering, Jinan University, Qianshan Road 206, Zhuhai, 519070, China; Qin-Bao Lin tlinqb@jnu.edu.cn Key Laboratory of Product
Packaging and Logistics, Packaging Engineering Institute, College of Packaging Engineering, Jinan University, Qianshan Road 206, Zhuhai, 519070, China;
Huai-Ning Zhong marco.zhong@iqtc-fcm.co National Reference Laboratory for Food Contact Material (Guangdong), Guangzhou Customs District
Technology Center, Guangzhou, Guangdong, 510623, China
Supplemental data for this article is available online at https://doi.org/10.1080/19440049.2023.2227732
ß 2023 Taylor & Francis Group, LLC
FOOD ADDITIVES & CONTAMINANTS: PART A 1115

with low-quality recycled plastic during the pro- (VOCs) and semi-volatile organic compounds
ducing process. Moreover, the recycled materials (SVOCs) using GC-MS (Pe~ nalver et al. 2022),
may have impurities from their recycling process, Raman spectroscopy (Pe~ nalver et al. 2023), MIR
previous life, and/or collecting system, greatly spectroscopy (Kassouf et al. 2014) and
affecting the quality and safety of the final prod- DSI/GC  GC-Q-TOF-MS (Hao et al. 2022), pre-
uct (EFSA Panel on Food Contact Materials, sumably because of the complex components pre-
Flavourings and Processing Aids 2014; Alvarado sent in r-PET or the limitation of the employed
Chacon et al. 2020; Brouwer et al. 2020; Thoden analytical tools. Discrimination between r-PET
van Velzen et al. 2020; Lambre et al. 2021). and v-PET based on NVOC of r-PET has scarcely
Although it is highly unlikely for such a situation been studied as it was assumed that they would
to occur in Europe due to restrictions under not significantly migrate.
framework directive EU Regulation (EU) Ultra-performance liquid chromatography
2022/1616, China currently lacks stringent regula- coupled to quadrupole time-of-flight mass spec-
tions for recycling PET intended for food use. trometry (UPLC-Q-TOF-MS), using soft ioniza-
While there are a few recycled PET cleaning lines tion, can on the one hand provide parent ion
in China that meet the decontamination require- information. On the other hand, combined with
ments set by the FDA and/or EFSA, they have high resolution mass spectrometry it can offer
not yet been widely adopted. Therefore, the use substructure information, making it suitable for
of recycled PET (r-PET) in food packaging is a qualitative analysis of unknown compounds
vital task to facilitate the recycling of plastics in (Dong et al. 2021). Meanwhile, UPLC-Q-TOF-
China. To achieve a safe r-PET that can be used MS has recently been proven to be an efficient
as food contact materials, the issue of chemical analytical technique for the detection of NIAS
safety following recycling has attracted consider- and intentionally added substances (IAS) in food
able attention. contact materials, such as polyethylene and poly-
Contaminants in r-PET primarily originate propylene (Hong et al. 2020; Wang et al. 2020;
from degradation products of polymer residues, Wrona and Nerın 2020). However, the use of
additives, and input contaminants (Welle 2011; UPLC-Q-TOF-MS for effective and rapid dis-
Barthelemy et al. 2014), where most of them are crimination of virgin and recycled PET based on
non-intentionally added substances (NIAS), espe- NVOC has not been explored yet.
cially non-volatile organic compounds (NVOCs) Herein, we aimed to effectively and rapidly
(Welle 2011; Barthelemy et al. 2014). In light of address the issue of potential v-PET adulteration
the NVOCs detected in r-PET, Europe has imple- based on NVOC tentatively identified by UPLC-
mented effective measures, such as the use of Q-TOF-MS, combined with various chemomet-
‘challenge tests’, to demonstrate the decontamin- rics. To create the most efficient discrimination
ation efficiency before applying for authorization approach, OPLS-DA combined with non-para-
of decontamination processes. Moreover, strict metric tests were employed to exploit the signifi-
legal restrictions are also imposed on recycling cant difference between r-PET and v-PET and to
non-food contact materials or substances in PET. screen the marker and high-level compounds. An
In contrast, in China, there is currently no com- inventory consisting of various models, such as
prehensive standard or database available for the principal components analysis (PCA), linear dis-
detection of NVOCs in PET. This lack of resour- criminant analysis (LDA), artificial neural net-
ces further complicates the accurate and efficient work (ANN), decision tree (DT), random forest
identification of these compounds. Thereby, a (RF), and support vector machine (SVM) based
method to effectively and rapidly monitor and on all compounds, markers and high-level
discriminate r-PET from virgin PET (v-PET) is markers obtained from positive, negative and the
required. Although previous studies have spent a combination of both ionization modes were
great effort to address PET adulteration using established, and the discrimination effects were
various analytical techniques, most current stud- comprehensively discussed and compared.
ies have focused on volatile organic compounds Moreover, cross-discrimination experiments
1116 T.-Y. HAO ET AL.

based on misclassified samples were further used recycled PET cleaning line from Sorema
to improve the prediction accuracy and to inves- Company in Italy and the extrusion production
tigate the possibility of identification of a large line from EREMA Company in Austria were pri-
sample set. Finally, throughout the home-made marily utilized. It is important to note that
databases, the detected characteristics were identi- Sorema’s recycled PET cleaning line meets the
fied to the extent feasible, and their possible cleaning requirements set by the FDA and EFSA
sources discussed. for food-grade applications. Additionally,
EREMA’s extrusion production line performs
Materials and methods decontamination and regeneration in accordance
with the food safety certification issued by the
Instruments, reagents, and plastic samples FDA. More details of PET samples employed in
Ultra-performance Liquid Chromatography- this study can be found in Table S1 of our previ-
Quadrupole Time-of-flight Mass Spectrometry ous study (Hao et al. 2022).
(UPLC-Q-TOF-MS) (1290 UPLC system coupled
to 6546 Q-TOF-MS system, Agilent Technologies, Sample preparation
Santa Clara, CA, USA). Agilent autosampler
A 0.2 g sample was completely dissolved in 10 mL
(Agilent Technologies, Santa Clara, CA, USA).
hexafluoro-isopropanol, followed by precipitation
Electronic balance (ME204, Mettler Toledo,
by the introduction of methanol with vigorous
Zurich, Switzerland). Ultrasonic instrument (AS
stirring. The mixture was then sonicated for
7240BT, Auto Science Instrument, Tianjin,
another 10 min before centrifuging for 5 min at
China). Ultra-pure water machine (Milli-Q,
9000 r/min. The supernatant was collected. The
Millipore Corporation, Billerica, MA, USA).
residues were rinsed three times with 5 mL
Desktop refrigerated centrifuge (3K15, Sigma,
methanol and the rinsed solvent was also col-
Osterode am Harz, Germany). Gas blowing evap-
lected. The collected solvents were removed, and
orator (Evatros T.C.S. EC-1648N, Goojung
another 1 mL methanol was added. The mixture
Engineering, Seoul, Korea). Compact centrifuge
was then centrifuged for 5 min at 13,000 r/min.
(D-37520 1-14, Sigma, Osterode am Harz,
Finally, the supernatant was transferred to a
Germany).
1.5 mL glass vial for further use.
Hexafluoro-isopropanol (99.5%, Macklin
Biochemical, Shanghai, China). Methanol (HPLC
grade, Fisher Chemical, Shanghai, China). Formic UPLC-Q-TOF-MS analysis
acid (HPLC grade, Fisher Chemical, Shanghai, The non-volatile organic compounds (NOVCs)
China). from recycled and virgin PET were detected by a
Seventy different samples of recycled PET (r- 1290 UPLC system coupled to 6546 Q-TOF-MS
PET) and 35 different samples of virgin PET (v- system fitted with SB-C18 column (ZORBAX SB-
PET) were marked as R01–R70 and V01–V35, C18, 2.1  100 mm, 1.8 mm. Agilent J&W
respectively. All v-PET (i.e. brand new PET bot- Scientific, Folsom, CA, USA). The column tem-
tles, particles, and powders) were provided by perature was set to 40  C. Samples were detected
local manufacturers or online purchasing plat- in a data-dependent acquisition (DDA) mode of
forms. In contrast, all r-PET were collected from UPLC-Q-TOF-MS. The solvents used as mobile
recyclers in already formed shapes (i.e. pellet) phase were 0.1% formic acid/water (phase A) and
from various locations in China. The samples methanol (phase B). This mobile phase system
used in this study were indicated as being was run in a gradient elution as follows: 0–
intended for use in food-grade applications. The 1.5 min, 5% B. 1.5–25 min, 5–98% B. 32–
materials for recycling had been subjected to a 32.1 min, 98–5% B. 32.1–35 min, 5% B. The injec-
cleaning process including hot water cleaning, tion volume was 3 lL. Flow rate was 0.3 mL/min.
ethanol rinsing, mechanical crushing, and drying. The parameters of mass detection were set as
In terms of the decontamination process, the follows: sensitivity mode. ionization mode was
FOOD ADDITIVES & CONTAMINANTS: PART A 1117

ESIþ and ESI. The capillary voltage was 4000 V The detected NOVCs and corresponding peak
for positive ion mode and 3500 V for negative area were set as variables and response value of
ion mode. sampling cone voltage was 500 V. variables, respectively. Three matrices based on
source temperature was 250  C. The desolvation all compounds, i.e. 105 (samples)  150 (varia-
temperature was 300  C. The sheath flow rate was bles) for positive ion mode, 105 (samples)  83
11 L/min. The flow rate of desolvation gas was (variables) for negative ion mode, 105
5 L/min. MS1 scan range was 50–1100 m/z and (samples)  202 (variables) for combination of
MS2 scan range was 20–800 m/z. MS2 analysis both, were established. Blank values for peak area
was performed with collision energies of 10, 20, in the data set were replaced with half the value
and 40 eV. of peak area for compounds involved to empha-
All experimental conditions were optimized size the importance of each datum point (Mais
according to our previous study (Dong et al. et al. 2018). Subsequently, the matrix was
2021). imported to the corresponding statistical soft-
ware, i.e. SIMCA, SPSS, and MATLAB, for mod-
eling using orthogonal partial least-squares
Data processing and identification
discrimination analysis (OPLS-DA), principal
Software Agilent MassHunter Qualitative components analysis (PCA), linear discriminant
Analysis 10.0 was used for data acquisition and analysis (LDA), artificial neural network (ANN),
processing. By using a lab-made database for decision tree (DT), random forest (RF) and sup-
food contact materials with the software integra- port vector machine (SVM). After screening of
tor Agile2, a non-targeted screening method was the all-compounds-datasets with OPLS-DA and
performed. The detected characteristics were non-parametric tests (Kruskal–Wallis test and
identified to the extent feasible. To compare median test), a further six new matrices were
feasibility and capability of discrimination mod- built. Matrices containing 105 (samples)  26
els, the datasets based on positive ion mode, (variables) marker compounds and 105
negative ion mode, and combination of both ion- (samples)  8 (variables) high-level marker com-
ization modes were processed, respectively. pounds were built for positive ion mode.
CAS number and possible origins of all identi- Meanwhile, matrices consisting of 105
fied compounds were obtained from a lab-made (samples)  16 (variables) marker compounds
database for food contact materials together with and 105 (samples)  3 (variables) high-level
PubChem (https://pubchem.ncbi.nlm.nih.gov/, marker compounds were established for negative
accessed on 10/10/2022) linking to extensive ion mode. Finally, matrices possessing 105
databases, including ‘EPA Chemical and Products (samples)  31 (variables) marker compounds
Database (CPDat)’, ‘Hazardous Substances Data and 105 (samples)  12 (variables) high-level
Bank (HSDB)’, ‘Joint FAO/WHO Expert marker compounds were created for combination
Committee on Food Additives (JECFA)’, ‘EU of both modes. These matrices were also applied
Pesticides Journal Pre-Proof Database’, and in corresponding discrimination modeling. Both
‘European Medicines Agency (EMA)’ (Chen et al. logarithm and pareto transformation were inves-
2022). tigated before OPLS-DA and PCA modeling,
thereby highlighting the importance of each vari-
able and eliminating discrepancies due to the
Statistical analysis
order of magnitude differences of peak areas of
The data of a three-dimensional matrix contain- different compounds. The normalization of data-
ing the names of detected compounds, the type sets before ANN and SVM modeling was also
of samples, and the peak areas of compounds performed, which can summarize the statistical
were exported to SIMCA 14.1 (Umetrics, Umeå, distribution and match the models. Meanwhile,
Sweden), SPSS 20.0 (SPSS, Chicago, IL, USA) and two-thirds of each dataset was marked as calibra-
MATLAB 2016a (MathWorks, Natick, MA, tion while the rest was set to validation, when
USA), respectively, for further processing. LDA, ANN, DT, RF, and SVM models were
1118 T.-Y. HAO ET AL.

established. Since manually sampling may lead to modeling. Other parts of the samples were set to
unstable results, random, systematic, and strati- default settings from the neural net fitting toolbox.
fied sampling were used in combination to inves- The decision tree (DT) model was performed
tigate the training and prediction set. by the decision tree toolbox of MATLAB software.
OPLS-DA has been known as a supervised The leaf nodes optimization and pruning were
projection method achieved by software SIMCA performed for better discrimination accuracy.
14.1. Accordingly, the data and scale can easily Random forest (RF) is a classifier containing
be mean-centered to unit variance. Meanwhile, multiple decision trees. The RF model was per-
the differences between two groups can be maxi- formed by a random forest toolbox named
mized to analyze the origins of the differences. MexStandalone-v0.02 designed for MATLAB soft-
From OPLS-DA, we can acquire a crucial param- ware. The RF model can be carried out on a few
eter named variance importance in project (VIP). lines of startup codes coupled with toolbox, and
Generally, variables with VIP > 1 are referred to the optimization was also performed.
as the important points of discrimination. The support vector machine (SVM) model was
Moreover, non-parametric tests (Kruskal–Wallis carried out on a LIBSVM toolbox version 3.25
test and median test) were calculated using SPSS developed by Professor Chih-Jen Lin and
25.0, which can determine the variables with stat- Professor Chih-Chung Chang, from National
istical significance. Variables with p < 0.05 were Taiwan University (https://www.csie.ntu.edu.tw/
considered as significant differences. Therefore, cjlin/libsvm/). The best error parameters, which
the variables with VIP > 1 and p < 0.05 were determined the discrimination accuracy of the
selected as marker compounds, while the varia- model, e.g. C value and G value, should be given
bles with VIP > 2 and p < 0.05 were marked as using a cross-validation and parameter optimiza-
high-level marker compounds. tion process. The variation range of both C and
Principal components analysis (PCA) is com- G values was set between 5 and 5 and the par-
monly used to reduce the dimensions of dataset ameter of cross-validation was set to 3. The step
into a few principal components (PCs). These size of C and G values were both set to 1, to get
PCs can maintain the most important informa- the most accurate model. Other parameters were
tion of data and the differences among samples set as the default values of LIBSVM toolbox. The
can be observed. In this study, PCA was per- LIBSVM toolbox will select the highest accuracy
formed by SIMCA 14.1 software. of cross-validation and the corresponding C and
Linear discriminant analysis (LDA) is a super- G values as the optimized parameters.
vised machine learning method. The dataset can
be divided into calibration and validation sets. Results and discussion
The testing accuracy of calibration and validation
Overview of NVOCs tentatively identified in v-PET
sets indicated the stability and predictability of
and r-PET samples
the model. The LDA model was processed with
SPSS 25.0 software. NVOCs are common compounds widely present
Artificial neural network (ANN) is also a super- in many consumer products, which can reflect
vised machine learning method. The backpropaga- their history. Thus, NVOCs are proposed as a
tion (BP) algorithm and the multi-layer model means of addressing potential v-PET adulteration
were constructed in this model. The ANN model- given the fact that the r-PET and v-PET samples
ing was performed by the neural net fitting tool- provided will have significantly different histories.
box of MATLAB software. The number of hidden Although the identification of unknown NVOCs
layers was determined according to the conven- from r-PET used in food contact materials is a
tional formula and Levenberg-Marquardt algo- very challenging task due to the number of pos-
rithm was chose as training algorithm. Some sible compounds involved, the use of UPLC-
important parameters, such as mean square error QTOF-MS is able to identify non-targeted com-
(MSE) and regression (R), were optimized during pounds. To test this hypothesis, we first employed
FOOD ADDITIVES & CONTAMINANTS: PART A 1119

two representative samples (e.g. R02 from r-PET mode (Figure 2(a)) and combination of both ion-
and V12 from v-PET) in UPLC-QTOF-MS using ization modes (Figure 2(c)), demonstrating the
both positive ion mode and negative ion mode to outstanding effectiveness of the discrimination. In
investigate the capability of identifying NVOCs. As contrast, all compounds were mixed together
shown in Figure 1, significant differences between when a negative ion mode was applied (Figure
r-PET and v-PET in both positive and negative ion 2(b)), presumably because of the small number of
modes were clear, indicating that UPLC-QTOF-MS detected NVOCs with negative ion mode, result-
can represent an alternative to discriminating PET ing in insufficient discrimination. Compared to
based on NVOCs. Thereby, NVOCs contained in OPLS-DA model based on negative ion mode
70 batches of r-PET samples and 35 batches of v- having a fitting parameter of 0.430 for R2Y (cum)
PET samples were screened by non-target UPLC- and 0.325 for Q2 (cum), OPLS-DA model based
QTOF-MS. A total of 202 types of NVOCs were on positive ion mode and combination of both
tentatively identified, of which 150 and 83 NVOCs ionization modes showed higher fitting parame-
were identified in positive and negative ion mode, ters of 0.938 for R2Y (cum) and 0.836 for Q2
respectively, indicative of the high complexity of (cum), as well as 0.969 for R2Y (cum) and 0.876
NVOCs present in PET samples. The details of
for Q2 (cum), respectively, indicative of the excel-
detected compounds are listed in Table S1. The
lent fitting and prediction performance.
resulting NVOCs data was imported into various
The excellent prediction was further verified
chemometrics for further analysis.
by permutation testing using 200 times calcula-
tion. The results are shown in Figures 2(d–f).
Analysis of various chemometrics The intercept of R2 in OPLS-DA model based
Analysis of OPLS-DA model and non-parametric on positive ion mode and combination of ion-
tests ization modes are both larger than 0.4 with a
Initially, OPLS-DA with optimized parameters value of 0.578 and 0.695, respectively, com-
was employed to perform PET discrimination. pared to that of negative ion mode with a
Figure 2 shows the OPLS-DA score plot based on value of 0.216, further verifying that OPLS-DA
all compounds in positive ion mode, negative ion model based on all compounds in positive ion
mode, and combination of both ionization mode and combination of ionization modes for
modes. A significant difference between r-PET PET discrimination were reliable and non-
and v-PET was clearly observed in positive ion overfitting.

Figure 1. Comparison of UPLC-QTOF-MS chromatograms between sample R02 and V12 in positive ion mode (a) and negative ion
mode (b).
1120 T.-Y. HAO ET AL.

Figure 2. OPLS-DA score plot based on all compounds detected from (a) positive ion mode, (b) negative ion mode, and (c) com-
bination of both ionization modes. Permutation at 200 times based on all compounds detected from (d) positive ion mode, (e)
negative ion mode, and (f) combination of both ionization modes.

Subsequently, the importance of variables for respectively. In contrast, the substances situated
PET adulteration assessment was also assessed on between VIP > 2 in OPLS-DA and p < 0.05 in
VIP in OPLS-DA modeling. Figure 3(a–c) show non-parametric tests were marked as ‘high-level
the VIP prediction plot of NVOCs from OPLS-DA marker compounds’ with 8, 3, and 12 compounds
modeling based on all compounds in positive ion for positive, negative, and combination of both ion-
mode, negative ion mode, and combination of both ization modes, respectively (Figure 3(a–c)).
ionization modes. The higher VIP values suggested Thereby, the marker compounds with VIP >1 in
the higher discrimination performance. three modes were plotted and the results were illus-
Accordingly, the compounds with VIP >1 in trated in loading scatter plots (Figure 3(d–f)). It
OPLS-DA and p < 0.05 in non-parametric tests should be noted that the substance closely located
were recorded as ‘marker compounds’ resulting in in the PET group suggests the great importance of
26, 16, and 31 compounds for positive, negative, this substance can be obtained in this group and
and combination of both ionization modes, vice versa. For example, 11 substances derived from
FOOD ADDITIVES & CONTAMINANTS: PART A 1121

Figure 3. VIP values of all compounds in (a) positive ion mode, (b) negative ion mode, and (c) combination of both ionization
modes. Loading scatter plot of OPLS-DA based on marker compounds in (d) positive ion mode, (e) negative ion mode, and (f)
combination of both ionization modes.

positive ion mode were the closest position to r- the marker compounds derived from a combin-
PET group, and thus can be categorized as repre- ation of both ionization modes showed a result
sentative NVOCs in the r-PET group (Figure 3(d)). similar to that of positive ion mode, where two
In contrast, the others displayed a much closer pos- clear separation regions corresponding to r-PET
ition to the v-PET group, thereby can be marked and v-PET group can readily be seen, also demon-
as representative NVOCs in v-PET group. strating that the non-volatile profiles had the poten-
Meanwhile, the substances were averagely distrib- tial to address PET adulteration.
uted in both r-PET and v-PET group, indicative of
the outstanding identification effect of the model. Analysis based on PCA model
However, only two substances showed greater Next, two-dimension score plots of PCA models
importance in v-PET group when a negative ion based on all compounds, markers, and high-level
mode was employed compared to the rest having a markers obtained from positive, negative, and
greater importance in r-PET group (Figure 3(e)), combination of both ionization modes were per-
suggesting that the negative ion mode was unstable formed to investigate the discrimination effective-
and non-reliable for PET identification. Moreover, ness (Figure 4). However, no clear boundary
1122 T.-Y. HAO ET AL.

Figure 4. PCA score plots of recycled and virgin PET samples based on different compound datasets and ion modes.
FOOD ADDITIVES & CONTAMINANTS: PART A 1123

between r-PET and v-PET was observed for all selected and assigned to marker and high-level
the tests, leading to the mixture of the detected marker groups, respectively. Twenty-six substan-
substances, showing the poor discrimination ces contained in the marker group obtained from
effects of the model. Although a tendency to positive ion mode can be divided into 12 IAS
show the difference between r-PET and v-PET and 14 NIAS (Table S1) including twelve plastic-
was visible in the marker group using positive related IAS, one plastic-related NIAS, one food-
ion mode (Figure 4(d)), some substances over- related NIAS, three medicine-related NIAS, two
lapped with each other resulting in the distortion pesticide-related NIAS, two industry-related
of models, thus making it difficult to identify. NIAS, and five NIAS coming from degradation
The first two PCs of PCA score plots based on and polymerization processes (Figure 5(a)).
all compounds (Figure 4(a–c)), markers (Figure Specifically, the plastic-related group mainly con-
4(d–f)), and high-level markers (Figure 4(g–i)) in tained 12 IAS, indicating that they originated
different ionization modes explained 16.0, 36.7, from primary compositions during plastic manu-
18.3, 29.5, 64.3, 37.7, 46.5, 83.3, and 49.1% of the facturing and recycling progress. For example,
original variations, respectively. Although the phthalic anhydride, dibutyl phthalate, diisononyl
explanation ratio increases after selecting the cyclohexane-1,2-dicarboxylate, stearic acid, and tri-
marker and high-level markers, it is hard to dif- octyl trimellitate are common plasticizers used in
ferentiate between r-PET and v-PET. The results plastic production. N-methylpyrrolidone is a solvent
of PCA revealed that PCA model coupled to used to produce resins, which is also the final prod-
UPLC-Q-TOF-MS is insufficient to discriminate uct of c-butyrolactone (or 4-butyrolactone) gener-
between r-PET and v-PET. ally applied to make the plasticizer. Di-propylene
glycol is an important solvent used facilitate the
Analysis of other chemometric algorithm models degradation of PET waste during alcoholysis. 1,3-
To get more insight into the discrimination effects, Diphenylguanidine is a coupling agent used in
we also imported the data for all compounds, polymerization. N,N0 -Ethylenebis(stearamide), and
markers, and high-level markers received from three Irgafos 168 phosphate are widely used as lubricants
ionization modes into other chemometrics and the and antioxidants, respectively. 1,3:2,4-Bis(3,4-dime-
results are listed in Table 1. The data obtained from thylbenzylidene)sorbitol is a nucleating agent used
positive ion mode showed the best prediction accur- in food contact materials to produce disposable
acy varying from 86.7 to 100% compared to the use film wrapping and bottle caps as well as repeat-use
of combination of both ionization modes with pre- food containers. Overall, the screening of plastic-
diction accuracy of 77.1–100%, where 100% predic- related IAS contained in marker group suggested
tion accuracy was achieved using DT model in both that they were significantly different and can be
ion modes, demonstrating excellent identification also served as important characteristic compounds
performance, which is in line with the results of for addressing PET adulteration.
OPLS-DA model. In addition, over 95% prediction Compared to IAS, the detected NIAS con-
accuracy was achieved when markers were chosen tained in marker groups were more complicated,
using LDA, ANN, RF, and SVM model, indicative of whose possible origins can be divided into six
the possibility and capacity in addressing PET adul- categories including plastic-related, food-related,
teration. In contrast, poor prediction accuracy was medicine-related, pesticide-related, and industry-
observed whichever chemometrics was used when related as well as degradation and polymeriza-
the dataset was obtained from a negative ion mode, tion-related groups. Specifically, medicine-related
showing that it is unable to detect PET adulteration. groups contained three substances consisting of
oxibendazole, Wulignan A1, and cucurbitacin,
which are anti-parasitic, anti-influenza, and anti-
Possible origins of marker compounds
neoplastic pharmaceuticals, respectively, presum-
Analysis of marker compounds in positive ion mode ably becoming contaminants during the first life
A total of 150 NVOCs were detected and identi- cycle of PET containers. The pesticide-related
fied, of which 26 and eight substances were group consisted of the active constituents of
1124 T.-Y. HAO ET AL.

Table 1. The discrimination accuracy between rPET and vPET measured by UPLC-Q-TOF-MS under positive, negative, and combin-
ation of both ionization modes coupled to various chemometrics.
Sampling method
Dataset Accuracy Random sampling Stratified sampling Systematic sampling Average
Positive ion mode
LDA
All compounds Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 68.6 77.1 82.9 76.2 ± 5.9
Marker compounds (VIP > 1) Calibration (%) 97.1 95.7 95.7 96.2 ± 0.7
Validation (%) 85.7 91.4 91.4 89.5 ± 2.7
High-level marker compounds (VIP > 2) Calibration (%) 94.3 94.3 95.7 94.8 ± 0.7
Validation (%) 88.6 97.1 91.4 92.4 ± 3.5
ANN
All compounds Calibration (%) 97.1 95.7 98.6 97.1 ± 1.2
Validation (%) 88.6 94.3 82.9 88.6 ± 4.7
Marker compound (VIP > 1) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 80.0 91.4 88.6 86.7 ± 4.9
High-level marker compounds (VIP > 2) Calibration (%) 97.1 94.3 98.6 96.7 ± 1.8
Validation (%) 77.1 88.6 77.1 80.9 ± 5.4
DT
All compounds Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
Marker compound (VIP > 1) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
High-level marker compounds (VIP > 2) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
RF
All compounds Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 97.1 100.0 100.0 99.0 ± 1.4
Marker compound (VIP > 1) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
High-level marker compounds (VIP > 2) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
SVM
All compounds Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 88.6 97.1 82.9 89.5 ± 5.8
Marker compound (VIP > 1) Calibration (%) 94.3 92.9 100.0 95.7 ± 3.1
Validation (%) 94.3 91.4 74.3 86.7 ± 8.8
High-level marker compounds (VIP > 2) Calibration (%) 94.3 91.4 94.3 93.3 ± 1.4
Validation (%) 88.6 91.4 71.4 83.8 ± 8.8
Negative ion mode
LDA
All compounds Calibration (%) 92.9 91.4 92.9 92.4 ± 0.7
Validation (%) 68.6 62.9 80.0 70.5 ± 7.1
Marker compounds (VIP > 1) Calibration (%) 87.1 82.9 82.9 84.3 ± 2.0
Validation (%) 60.0 80.0 80.0 73.3 ± 9.4
High-level marker compounds (VIP > 2) Calibration (%) 78.6 74.3 75.7 76.2 ± 1.8
Validation (%) 68.6 68.6 74.3 70.5 ± 2.7
ANN
All compounds Calibration (%) 90.0 92.9 90.0 91.0 ± 1.4
Validation (%) 77.1 65.7 80.0 74.3 ± 6.2
Marker compound (VIP > 1) Calibration (%) 81.4 81.4 81.4 81.4 ± 0.0
Validation (%) 68.6 68.6 74.3 70.5 ± 2.7
High-level marker compounds (VIP > 2) Calibration (%) 81.4 80.0 81.4 80.9 ± 0.7
Validation (%) 68.6 77.1 74.3 73.3 ± 3.5
DT
All compounds Calibration (%) 97.1 95.7 92.9 95.2 ± 1.7
Validation (%) 62.9 80.0 74.3 72.4 ± 7.1
Marker compound (VIP > 1) Calibration (%) 94.3 84.3 90.0 89.5 ± 4.1
Validation (%) 65.7 77.1 74.3 72.4 ± 4.9
High-level marker compounds (VIP > 2) Calibration (%) 90.0 82.9 87.1 86.7 ± 2.9
Validation (%) 65.7 77.1 77.1 73.3 ± 5.4
RF
All compounds Calibration (%) 94.3 92.9 92.9 93.4 ± 0.7
Validation (%) 77.1 80.0 82.9 80.0 ± 2.4
Marker compound (VIP > 1) Calibration (%) 90.0 84.3 85.7 86.7 ± 2.4
Validation (%) 71.4 77.4 80.0 76.3 ± 3.6
High-level marker compounds (VIP > 2) Calibration (%) 87.1 82.9 84.3 84.8 ± 1.7
Validation (%) 68.6 77.1 77.1 74.3 ± 4.0
SVM
All compounds Calibration (%) 92.9 92.9 92.9 92.9 ± 0.0
Validation (%) 82.9 74.3 85.7 81.0 ± 4.9
(continued)
FOOD ADDITIVES & CONTAMINANTS: PART A 1125

Table 1. Continued.
Sampling method
Dataset Accuracy Random sampling Stratified sampling Systematic sampling Average
Marker compound (VIP > 1) Calibration (%) 85.7 81.4 80.0 82.4 ± 2.4
Validation (%) 68.6 77.1 77.1 74.3 ± 4.0
High-level marker compounds (VIP > 2) Calibration (%) 82.9 80.0 80.0 81.0 ± 1.4
Validation (%) 71.4 77.1 74.3 74.3 ± 2.3
Combination of both ionization modes
LDA
All compounds Calibration (%) 100.0 100.0 98.6 99.5 ± 0.7
Validation (%) 60.0 65.7 68.6 64.8 ± 3.6
Marker compounds (VIP > 1) Calibration (%) 98.6 100.0 97.1 98.6 ± 1.2
Validation (%) 85.7 97.1 91.4 91.4 ± 4.7
High-level marker compounds (VIP > 2) Calibration (%) 94.3 97.1 95.7 95.7 ± 1.1
Validation (%) 91.4 91.4 91.4 91.4 ± 0.0
ANN
All compounds Calibration (%) 95.7 92.9 95.7 94.8 ± 1.3
Validation (%) 88.6 77.1 80.0 81.9 ± 4.9
Marker compound (VIP > 1) Calibration (%) 90.0 92.9 88.6 90.5 ± 1.8
Validation (%) 85.7 80.0 74.3 80.0 ± 4.7
High-level marker compounds (VIP > 2) Calibration (%) 92.9 94.3 90.0 82.4 ± 1.8
Validation (%) 82.9 88.6 77.1 82.9 ± 4.7
DT
All compounds Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
Marker compound (VIP > 1) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
High-level marker compounds (VIP > 2) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
RF
All compounds Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
Marker compound (VIP > 1) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 100.0 100.0 100.0 100.0 ± 0.0
High-level marker compounds (VIP > 2) Calibration (%) 100.0 100.0 100.0 100.0 ± 0.0
Validation (%) 97.1 100.0 88.6 95.2 ± 4.8
SVM
All compounds Calibration (%) 92.9 94.3 92.9 93.4 ± 0.7
Validation (%) 80.0 85.7 82.9 82.9 ± 2.3
Marker compound (VIP > 1) Calibration (%) 91.4 91.4 92.6 91.8 ± 0.6
Validation (%) 80.0 85.7 65.7 77.1 ± 8.4
High-level marker compounds (VIP > 2) Calibration (%) 92.9 92.9 92.9 92.9 ± 0.0
Validation (%) 85.7 85.7 85.7 85.7 ± 0.0

pesticides (e.g. dodemorph) and intermediates of TPA2-PEG2, cyclic TPA2-PEG4, HOOC-TPA3-


pesticides (e.g. triphenylphosphine oxide), which PEG3-OH, HO-TPA3-PEG4-OH, and cyclic
may originate as adventitious contaminants of TPA4-PEG4 are monomers or oligomers used in
the PET used or during the recycling process. the preparation of polymers or the degradation
These compounds arouse a serious safety concern products, possibly resulting from high tempera-
and necessary procedures should be taken (e.g. ture and pressure during the recycling decontam-
closed loop recycling) due to the high toxicity ination process. The plastic-related group
listed in several databases (e.g. HSDB, NIOSH, possessed only one NIAS (i.e. 9,10-dihydroxystea-
and Chinese national standards). Although the ric acid) coming from the oxidation of oleic acid
food-related group contained only one NIAS (i.e. in PET.
eicosenamide), a flavor component of tea, the
safety concern should not be ignored due to its Analysis of marker compounds from combination of
various applications in medicine, pesticide, and both ionization modes
plastics that may cause hazardous concern to Next, we further analyzed the possible origins of
consumers health. In addition, isobornyl acrylate marker compounds obtained from combination
and 4-t-butylcyclohexyl acrylate were included in of both ionization modes. As shown in Table S1
the industry-related group, which were intro- and Figure 5(b), 202 types of NVOCs were
duced from plastics coatings. The detected cyclic detected and identified, where 31 substances
1126 T.-Y. HAO ET AL.

Figure 5. Distribution of possible origins of IAS and NIAS for marker compounds detected from (a) positive ion mode and (b) com-
bination of both ionization modes. Venn diagrams of (c) marker compounds and (d) high-level marker compounds detected from
positive ion mode and combination of both ionization mode.
FOOD ADDITIVES & CONTAMINANTS: PART A 1127

contained in marker groups were selected com- 10 mgdm2) and Asian countries (e.g.
prising 11 plastic-related IAS, one plastic-related 30 mgdm2) (Kim and Lee 2012).
NIAS, three food-related NIAS, six medicine- These findings provide additional evidence that
related NIAS, two pesticide-related NIAS, one r-PET collected in China may contain impurities
industry-related NIAS, and seven degradation- derived from the recycling process, previous
related NIAS. Compared to the positive ion usage, or the collection system. Therefore, there
mode having five unique marker compounds is an urgent need to implement closed-loop food-
detected, 10 marker compounds involving two grade plastic recycling and effective plastic waste
IAS and eight NIAS were uniquely detected from sorting practices in China, particularly consider-
combination of both ionization modes while the ing the high toxicity of most marker compounds.
rest (21 marker compounds) overlapped and
were identically found in positive ion mode
Analysis of misclassified samples and cross-
(Figure 5(c,d)), indicative of stability and inde-
discrimination
pendence of discrimination models. Although the
combination of both ionization modes would To improve the prediction accuracy, we further
provide additional compound information as analyzed the misclassified samples. As shown in
compared to the single ion mode used, such add- Table S2, misclassified samples largely appeared
itional information has no significant impact on in models based on all compounds and models
the final discrimination. based on compounds obtained from negative ion
For example, two IAS (i.e. 1,3-benzenedicar- mode, presumably because of contamination and
boxylic acid and bis(2-hydroxyethyl) terephthal- poor sample quality, leading to discrimination
ate) are the intermediates of plastic serving as instability. In contrast, a significant difference in
plasticizers, and thus were included in the plas- total peak area between two samples within the
tic-related group. Regarding the unique NIAS, inner groups based on markers detected by posi-
the food-related group contained 4-((2-hydroxye- tive and combination of both ionization modes
thoxy)carbonyl)benzoic acid and linoleic acid that was clearly observed, resulting in a smaller num-
are often used as food additive, flavoring agent, ber of misclassified samples. However, careful
and edible essence. However, linoleic acid has a observation between two samples in inter-groups
wide application in various fields in addition to showed small differences in total peak area that
being a flavoring agent, which may pose a poten- were close to the average peak area, which may
tial threat to human health and should arouse a increase the instability and lower the prediction
safety concern. Isoscopoletin, tectochrysin, and accuracy of models. Nonetheless, a smaller num-
pectolinarigenin were identified as constituents of ber of misclassified samples further proved that
medicines particularly employed in the treatment substances obtained from positive mode, and
of heart disease, tumor, or inflammation. These combination of both ionization modes would
medical-related compounds possibly resulted have a better discrimination between r-PET and
from medicine containers in the first life cycle of v-PET groups compared to that of the negative
PET or contaminants during the recycling pro- ion mode.
cess, indicative of the essential need for effective Subsequently, we also performed a cross-dis-
sorting of PET waste. HO-TPA3-PEG5-OH, crimination based on misclassified samples using
HOOC-TPA4-PEG3-COOH, and HOOC-TPA4- various chemometrics, such as LDA, ANN, DT,
PEG4-OH were monomers or degradation prod- RF, and SVM to study the possibility of applying
ucts that can be assigned to polymerization and the discrimination approach to a large sample
degradation-related group. The safety hazards of set. A value was first given to each misclassified
these NIAS are unknown but require attention as sample, where ‘1’ and ‘1’ were ascribed to r-
they might migrate to fatty and non-fatty simu- PET and v-PET, respectively. Accordingly, the
lants, although their migration levels (e.g. 0.05– final scores can be readily calculated based on the
2.46 lgdm2) were within the range of overall summary of chemometrics score and the sample
migration limits proposed by EU (e.g. categories can be easily assigned. For example,
1128 T.-Y. HAO ET AL.

final positive scores were marked as v-PET while a single model, which may enable identifying a
final negative scores were referred to as r-PET. large sample set.
The details of cross-discrimination results are Overall, UPLC-Q-TOF-MS combined with
shown in Table S3. It was clearly observed that various chemometrics is a promising tool for
the misclassified samples can be readily validated PET wastes separation that can further promote
by cross-discrimination with an accuracy of 90.9 discrimination performance in this area; however,
and 96.3% for models based on marker com- further technological improvement, such as
pounds obtained from positive mode and the enhancement of prediction accuracy of cross-dis-
combination of both ionization modes, respect- crimination is required before employing this
ively, thus indeed improving the prediction approach at an industrial level, considering that
accuracy of various chemometrics and avoiding all of this study was conducted under laboratory
the false results of single model used, which may conditions.
also enable identifying a large sample set.
Author contributions
Conclusions
Tian-Ying Hao: conceptualization, methodology, software,
In conclusion, this work demonstrates the results resources, writing-original draft. Xiaowen Xu: validation,
of a feasible study on the discrimination of 105 resources, supervision, writing-review and editing and
batches of r-PET and v-PET broadly employed in English polishing, funding acquisition. Qin-Bao Lin: valid-
ation, resources, supervision, writing-review and editing,
the plastic packaging industry based on 202 types
funding acquisition. Xue-Feng Wu: methodology, resources.
of NVOCs tentatively identified by UPLC-Q- Si-Liang Wu: resources. Huai-Ning Zhong: validation,
TOF-MS combined with various chemometrics. resources. Ben Dong: methodology, software, resources.
With the help of OPLS-DA and non-parametric Zhi-Feng Chen: methodology, software. Zhi-Kang Ye: meth-
tests, the marker compounds (i.e. VIP > 1 and odology. Zhi-Wei Wang: resources, funding acquisition.
p < 0.05) between r-PET and v-PET possessing
great importance for discrimination were non- Disclosure statement
targeted screened and can be recorded as charac-
No potential conflict of interest was reported by the
teristic IAS and NIAS. The possible origins of
author(s).
these compounds could be plastic, food, medi-
cine, pesticide, industry-related and degradation
and polymerization products. As many of them Funding
have high toxicity, especially those pesticide- This work was supported by Guangdong Basic and Applied
related, this indicates the urgent requirement for Basic Research Foundation (Grant No. 2022A1515110015),
closed loop recycling and sorting of PET waste. Zhuhai Basic and Applied Basic Research Foundation
Moreover, based on the marker group, the accur- (Grant No. 2220004002982), Science and Technology
Planning Project of Guangzhou (Grant No. 2023A04J0081),
acy of both validation and calibration could reach
National Key R&D Program of China (Grant No.
100% in DT model whichever ionization modes 2022YFF0607202), and National Natural Science
were used. Meanwhile, the prediction accuracy of Foundation of China (32061160474).
ANN, RF, and SVM models can also reach over
95% when the positive ion mode of UPLC-Q-
ORCID
TOF-MS combined with marker compound data-
set was employed, suggesting the positive ion Xiaowen Xu http://orcid.org/0000-0003-1404-6619
mode would be the optimal approach to gain the
best discrimination efficiency. Meanwhile, a References
cross-discrimination based on misclassified sam- Alvarado Chacon F, Brouwer MT, Thoden van Velzen EU.
ples using various chemometrics showed that 2020. Effect of recycled content and r-PET quality on the
over 90% accuracy can be achieved, thus indeed properties of PET bottles, part I: optical and mechanical
improving the prediction accuracy of various che- properties. Packag Technol Sci. 33(9):347–357. doi: 10.
mometrics and avoiding the false results of using 1002/pts.2490.
FOOD ADDITIVES & CONTAMINANTS: PART A 1129

Awaja F, Pavel D. 2005. Recycling of PET. Eur Polym J. Qual. 11(13):4139–4149. doi: 10.19812/j.cnki.jfsq11-5956/
41(7):1453–1477. doi: 10.1016/j.eurpolymj.2005.02.005. ts.2020.13.001.
Barthelemy E, Spyropoulos D, Milana M-R, Pfaff K, Kassouf A, Maalouly J, Rutledge DN, Chebib H, Ducruet V.
Gontard N, Lampi E, Castle L. 2014. Safety evaluation of 2014. Rapid discrimination of plastic packaging materials
mechanical recycling processes used to produce poly- using MIR spectroscopy coupled with independent com-
ethylene terephthalate (PET) intended for food contact ponents analysis (ICA). Waste Manag. 34(11):2131–2138.
applications. Food Addit Contam Part A Chem Anal doi: 10.1016/j.wasman.2014.06.015.
Control Expo Risk Assess. 31(3):490–497. doi: 10.1080/ Kim D-J, Lee KT. 2012. Analysis of specific migration of
19440049.2013.871755. monomers and oligomers from polyethylene terephthalate
Brouwer MT, Alvarado Chacon F, Thoden van Velzen EU. bottles and trays according to the testing methods as pre-
2020. Effect of recycled content and r-PET quality on the scribed in the legislation of the EU and Asian countries.
properties of PET bottles, part III: modelling of repetitive Polym Test. 31(8):1001–1007. doi: 10.1016/j.polymertest-
recycling. Packag Technol Sci. 33(9):373–383. doi: 10. ing.2012.07.012.
1002/pts.2489. Lambre C, Barat Baviera JM, Bolognesi C, Chesson A,
Chen Z-F, Lin Q-B, Dong B, Zhong H-N, Wang Z-W. Cocconcelli PS, Crebelli R, Gott DM, Grob K, Mengelers
2022. Comparison of the ability of UV-Vis and UPLC-Q- M, Mortensen A, et al. 2021. Safety assessment of the
TOF-MS combined with chemometrics to discriminate process PET STAR RECYCLING, based on Starlinger
recycled and virgin polyethylene. J Hazard Mater. 423(Pt iV þ technology, used to recycle post-consumer PET into
B):127165. doi: 10.1016/j.jhazmat.2021.127165. food contact materials. EFSA J. 19(8):e06791. doi: 10.
Dong B, Wu X, Hao T, Li D, Zeng Y, Wu S, Kou X, Li Z, 2903/J.EFSA.2021.6791.
Chen S, Lin Q. 2021. Untargeted screening of nonvolatile Mais E, Alolga RN, Wang S-L, Linus LO, Yin X, Qi L-W.
organic compounds in recycled PET flakes for food con- 2018. A comparative UPLC-Q/TOF-MS-based metabolo-
tact applications by ultra-performance liquid chromatog- mics approach for distinguishing Zingiber officinale
raphy–quadrupole time-of-flight mass spectrometry. J Roscoe of two geographical origins. Food Chem. 240:
Instrum Anal. 40(11):1545–1552. doi: 10.19969/j.fxcsxb. 239–244. doi: 10.1016/j.foodchem.2017.07.106.
21080403. Palm E, Hasselbalch J, Holmberg K, Nielsen TD. 2022.
EFSA Panel on Food Contact Materials, Flavourings and Narrating plastics governance: policy narratives in the
Processing Aids. 2014. Scientific opinion on the safety European plastics strategy. Environ Polit. 31(3):365–385.
assessment of the process SOREPET, based on Buhler B doi: 10.1080/09644016.2021.1915020.
technology, used to recycle post-consumer PET into food nalver R, Marın C, Arroyo-Manzanares N, Campillo N,
Pe~
contact materials. EFSA J. 12(4):3631. doi: 10.2903/j.efsa. Vi~nas P. 2022. Authentication of recycled plastic content
2014.3631. in water bottles using volatile fingerprint and chemomet-
EFSA Panel on Food Contact Materials. 2011. Flavourings rics. Chemosphere. 297:134156. doi: 10.1016/j.chemo-
and processing aids, scientific opinion on the criteria to sphere.2022.134156.
be used for safety evaluation of a mechanical recycling Pe~
nalver R, Zapata F, Arroyo-Manzanares N, L opez-Garcıa
process to produce recycled PET intended to be used for I, Vi~nas P. 2023. Raman spectroscopic strategy for the
manufacture of materials and articles in contact with discrimination of recycled polyethylene terephthalate in
food. EFSA J. 9(7):2184. doi: 10.2903/j.efsa.2011.2184. water bottles. J Raman Spectroscopy. 54(1):107–112. doi:
Hao T, Lin Q, Zhong H, Dong B, Kou X, Chen Z, Ye Z, 10.1002/jrs.6457.
Wang Z. 2022. Study on the Ability of HS- Ritchie H, Roser M. 2018. Plastic pollution.
SPME/GC  GC-Q-TOF MS Combined with OurWorldInData.org [accessed 2022 Dec 29]. https://our-
Chemometrics to Discriminate Recycled Polyethylene worldindata.org/plastic-pollution.
Terephthalate. J Instrum Anal. 41(10):1–12. doi: 10. Thoden van Velzen EU, Brouwer MT, St€arker C, Welle F.
19969/j.fxcsxb.22030701. 2020. Effect of recycled content and r-PET quality on the
Hao T-Y, Xu X, Lin Q-B, Wu S-L, Wu X-F, Hu J-L, Zhong properties of PET bottles, part II: migration. Packag
H-N, Dong B, Chen Z-F, Ye Z-K, et al. 2022. Rapid dis- Technol Sci. 33(9):359–371. doi: 10.1002/pts.2528.
crimination of recycled and virgin poly(ethylene tereph- US FDA. 2021. Use of recycled plastics in food packaging
thalate) based on non-targeted screening of semi-volatile (chemistry considerations): guidance for industry. U.S.
organic compounds using a novel method of Department of Health and Human Services Food and
DSI/GC  GC-Q-TOF-MS coupled with various chemo- Drug Administration Center for Food Safety and Applied
metrics. Food Packag Shelf Life. 34:100978. doi: 10.1016/ Nutrition; [accessed 2023 Mar 21]. https://www.fda.gov/
j.fpsl.2022.100978. media/150792/download.
Hong L, Liang W, Sheng C, Dan L, Huaining Z, Zhaofang Wang Y, Gao X, Liu B, Lin Q, Xia Y. 2020. Identification of
R, Zicheng L. 2020. Research progress on high-resolution chemicals in a polyvinyl chloride/polyethylene multilayer
mass spectrometry analysis of high concern compounds film by ultra-high-performance liquid chromatography/-
in food contact plastic materials and products. J Food Saf quadrupole time-of-flight mass spectrometry and their
1130 T.-Y. HAO ET AL.

migration into solution. J Chromatogr A. 1625:461274. JECFA Joint FAO/WHO Expert Committee on Food
doi: 10.1016/j.chroma.2020.461274. Additives
Welle F. 2011. Twenty years of PET bottle to bottle recy- LDA linear discriminant analysis
cling—an overview. Resour Conserv Recycl. 55(11):865– MSE mean square error
875. doi: 10.1016/j.resconrec.2011.04.009. NIAS non-intentionally added substance
Wrona M, Nerın C. 2020. Analytical approaches for analysis NVOCs non-volatile organic compounds
of safety of modern food packaging: a review. Molecules. OPLS-DA orthogonal partial least squares discrimination
25(3):752. doi: 10.3390/molecules25030752. analysis
PCA principal components analysis
Abbreviations PCs principal components
PET polyethylene terephthalate
ANN artificial neural network R regression
BP backpropagation RF random forest
CAS Chemical Abstracts Service r-PET recycled polyethylene terephthalate
CPDat EPA Chemical and Products Database SVM support vector machine
DDA data-dependent acquisition SVOCs semi-volatile organic compounds
DT decision tree UPLC-Q-TOF-MS
EFSA European Food Safety Authority ultra-performance liquid chromatography-quadru-
EMA European Medicines Agency pole time-of-flight mass spectrometry
EPA Environmental Protection Agency VIP variance importance in project
FAO Food and Agriculture Organization of the United VOCs volatile organic compounds
Nations v-PET virgin polyethylene terephthalate
HSDB Hazardous Compounds Data Bank WHO World Health Organization
IAS intentional food additives

You might also like