Introduction

There has been an increasing demand for analytical techniques that can detect and identify unexpected or unknown contaminants in food matrices without any prior knowledge on their occurrence. As a result, non-targeted strategies, as opposed to the traditional targeted analysis in which methods are designed based on the availability of analytical standards, need to be developed. The interest for novel non-targeted tools is not limited to food analysis but is also emerging in numerous fields such as environmental analysis or forensics [1,2,3].

There are generally four components in a non-targeted workflow: sample preparation, instrument analysis, post-analysis data treatment, and data interpretation. The non-targeted analysis of trace contaminants in food, where the low concentration of the analytes poses an added challenge, can be achieved using state-of-the-art instruments, e.g., high-performance liquid chromatography coupled to hybrid quadrupole time-of-flight mass spectrometry (HPLC-QTOF-MS) [4, 5]. This approach results in relatively large datasets and requires advanced data treatment in order to extract and identify the relatively small signals of the contaminants [6]. The non-targeted identification of compounds based on liquid chromatography or gas chromatography coupled to mass spectrometry (LC/GC-MS) data is a multistep process. First, the whole chromatogram needs to be examined in order to create a compound list of all peaks that represent real molecules [7]. During this process, detected ions are assigned a monoisotopic peak with a corresponding m/z value, and an isotopic distribution analysis is performed to identify isotopic clusters that might be characteristic of certain elemental compositions. Next, the presence of adducts that may be associated with the eluting compounds is assessed based on a list of potential adducts specified beforehand [8]. At this point, if the mass accuracy is sufficient and there is a minimal isotopic distribution error, the correct molecular formula can be generated for the compounds of interest [9, 10]. Finally, if a compound database is available, mass spectral similarities between an experimental mass spectrum and each mass spectrum in a reference library are assessed [11]. Often, several compounds from a specific database can be associated with a single molecular formula. In that case, acquisition methods using data-dependent or data-independent MS/MS can provide some information (exact mass, isotopic patterns) for both the parent ions and their fragments that can be used to further confirm the correct identification of the target compounds [12].

The correct detection and identification of contaminants in food is critical to ensure the safety of the consumers. Contaminated food samples wrongly assigned as a non-detect result (false negative) could lead to an unacceptable exposure of the consumers to potential toxicants. On the other hand, food products wrongly identified as contaminated (false positive) could potentially result into unnecessary food recalls and economic losses. Since the unbiased identification of chemical residues in food products is a matter of concern for public health, regulations such as the 2002/657/EC European Commission Decision have established a system of identification points (IPs) to confirm the identification of organic residues and contaminants using mass spectrometric techniques [13]. The main advantage of this system is to provide a standardized approach to confirm the identity of chemical contaminants [14]. So far, most non-targeted studies have focused on improving the accuracy of compound identification, but little attention has been paid to reducing the false discovery rate [15].

Analytical steps such as sample preparation or instrument analysis are known to influence the list of molecular features obtained in a non-targeted metabolomics [16]. In the case of complex matrices, such as food samples, matrix effects are also known to affect the automatic identification of compounds such as pesticides [12]. More specifically, interferences from the matrix (e.g., signal suppression, co-elution with isobaric compounds) can induce a loss of mass accuracy and lead to erroneous identification of unknowns [7]. This underlines the need for a thorough study and optimization of the data treatment steps to ensure the correct identification of compounds in food matrices. Some researchers have reported the influence of the “exact mass tolerance” parameter used to compare the measured m/z (experimental) with that in the database (theoretical) on the non-targeted identification of contaminants in food matrices (i.e., fruits and vegetables) [7, 17, 18]. In metabolomics, the need for identifying the parameters that have the main impact on the number and quality of reported metabolites has been recognized for a while [19]. Recently, Tian et al. [20] demonstrated the influence of post-acquisition data processing for the non-targeted screening of trace leachable residues from reusable plastic bottles. However, to the best of our knowledge, there has been no systematic assessment of the influence of any other parameters of post-analysis data treatment other than the exact mass tolerance on the identification of trace contaminants in food.

Among the infinite variety of food matrices that can be used as a model for the study of the influence of non-targeted identification parameters, honey offers the advantage of being able to contain an extended range of contaminants, both agricultural and environmental, at trace levels (in the μg kg−1 to mg kg−1 range) [21,22,23]. Honeybees fly up to 4 km in all directions from their hive and thus have access to an area of 50 km2, which allows them to be exposed to a wide variety of contaminants in a large area [24]. In addition, a method for the determination of seven veterinary drugs in honey based on a simple dilute-and-shoot procedure followed by direct injection into HPLC-QTOF-MS/MS was recently developed [25]. Although this was initially a targeted analysis method, the use of a data-independent MS/MS acquisition such as the All Ions mode showed promising applications in non-targeted analysis of contaminants in honey.

The aim of the study was to investigate the influence of the post-analysis data treatment parameters on the non-targeted identification of trace contaminants (suspected-target) in honey, using seven veterinary drugs as model compounds. More specifically, an assessment was made to study the impact of the following: match mass tolerance, the mass extraction window, the isotope abundance score, the peak filter absolute height, the average of spectra, the exclusion of TOF spectra, and the post-processing peak filters. To date, there has no comprehensive study of the impact of these parameters in the non-targeted identification of contaminants in honey. Optimized data treatment conditions were then applied to the non-targeted screening of veterinary drugs, pesticides, and other contaminants in honey collected in Canada.

Experimental

Chemicals and Reagents

Analytical standards (tylosin A [≥ 94.2%], lincomycin [≥ 95.0%], furazolidone [≥ 99%], sulfamethoxazole [≥ 99%], sulfadimethoxine [≥ 98.5%], sulfamethazine [≥ 99%], and 5-hydroxymethyl-2-furaldehyde [≥ 99%]) were purchased from Sigma-Aldrich (St. Louis, MO, USA). Tylosin B (≥ 99%) was purchased from Toku-E (Bellingham, WA, USA). Deuterated internal standards (D3-diphenylhydramine [≥ 99.4%] and D3-6-acetylmorphine [≥ 99.5%]) were purchased from Cerilliant (Round Rock, TX, USA). HPLC grade solvents (water, methanol, acetonitrile, acetone, and 2-propanol), as well as LC/MS grade formic acid, were all obtained from Fisher Chemical (Pittsburgh, PA, USA).

Honey Samples

Fifty-five honey samples were purchased from different stores and farmers’ markets in the Montreal and Calgary regions (Canada) in May 2016 and May 2017. They were all labeled as “unpasteurized” and were of various prices and quality (i.e., different colors and different floral and geographical origins, some labeled as organic as defined in Canada) [26]. Subsamples of each honey were transferred from their original container to 40-mL amber glass vials and kept in the freezer at − 18 °C until analysis. In an earlier study [25], these samples were analyzed and 15 of them showed no detectable residues of any of the seven tested antimicrobials, so they were considered as blank matrices for the present study.

Sample Preparation

The sample preparation follows a method described in an earlier paper [25]. In short, approximately 0.2 g of honey was weighed in a glass conic tube and 2 mL of acetonitrile:water (1:1) mixture was added. Samples were vortexed until the honey was completely dissolved, and then filtered through a 0.22-μm PTFE filter from Chrom4 (Thüringen, Germany). Before injection into the HPLC system, the extract was further diluted with water to a final concentration corresponding to 1% of honey (w/v), and 50 μL of a 0.4 μg mL−1 mixture of the two deuterated internal standards was added. These internal standards were not added for quantification purposes but serve as a reference for retention time and sensitivity [27].

For the optimization of the identification parameters, the 15 matrix blanks were spiked before dilution with 50 μL of a mixture of the seven antimicrobials at a concentration corresponding to 0.2 μg g−1 of honey. This concentration corresponds to the maximum regulatory limit (MRL) for tylosin A + B in honey in Canada [28].

Instrument Analysis

Samples were analyzed using a 1290 series LC system from Agilent Technologies (Santa Clara, CA, USA) equipped with an InfinityLab Poroshell 120 Phenyl Hexyl (3.0 mm × 100 mm, 2.7 μm) column fitted with an InfinityLab Poroshell 120 EC-C18 (3.0 mm × 5 mm, 2.7 μm) guard column, both from Agilent Technologies. This column, which has improved selectivity for aromatic compounds, was used in a previous study and was found to successfully separate the seven veterinary drugs used in the present study [25, 29]. The mobile phase composition was prepared as reported in the literature on non-targeted analysis using reverse-phase liquid chromatography. In that case, the mobile phase commonly consists of a mixture of water and methanol or acetonitrile, often with the addition of a small quantity of mobile phase modifiers (e.g., ammonium formate, ammonium acetate, formic acid or acetic acid) to improve the separation and detection of compounds [30, 31]. This mobile phase is generally used starting with a low proportion of organic phase (i.e., 5–10%), held in isocratic mode for a few minutes. A gradient is then generally applied until 100% of organic phase is obtained. Finally, the mobile phase composition went back to the initial conditions to re-equilibrate the system for a few minutes [31]. In the present study, the mobile phase consisted of water (solvent A) and methanol (solvent B), both with 0.1% formic acid, at a flow rate of 0.2 mL min−1. The mobile phase gradient profile was as follows: 1 min, 5% B; from 1 to 15 min, gradient to 100% B; from 15 to 20 min, 100% B; from 20 to 20.10 min, gradient to 5% B; and from 20.10 to 25 min, 5% B. The injection volume was 20 μL, and the column temperature was set to 20 °C.

The LC system was coupled to a 6545 series Q-TOF from Agilent Technologies equipped with a Dual AJS ESI ion source operating in positive ionization mode. The drying gas temperature was 325 °C with a flow of 5 L min−1, the sheath gas temperature was 275 °C with a flow of 12 L min−1, the pressure on the nebulizer was 20 psi, the capillary voltage was 4000 V, the fragmentor voltage was 175 V, the skimmer voltage was 65 V, and the nozzle voltage was 2000 V. All Ions MS/MS data was collected [scans between m/z 70 and 1700 at a scan rate of 3 spectra/s for four different collision energies (0 V, 10 V, 20 V, and 40 V)]. This scan speed corresponded to about 14 spectra per peak for each collision energy. This data-independent MS/MS acquisition mode has been reported to be more appropriate for large-scale screening than the classic product ion scans because it provided excellent fragmentation information for confirmatory purposes for an unlimited number of compounds [32]. The first 3 min of elution was diverted to waste. Samples were kept at 4 °C in the multisampler compartment.

Data Treatment

The datasets were processed with MassHunter Profinder B.08.00 software (Agilent Technologies), using Batch Targeted Feature Extraction. This algorithm extracts features from the acquired data, combining the “Find Compounds by Formula” algorithm with a database containing molecular formulas, mass, and/or retention time information [33]. Batch Targeted Feature Extraction was run under different conditions to determine which parameters influence the rates of detection and identification of veterinary drugs present in honey at trace level. The parameters were selected based on what has been reported by others, and other additional parameters representative of the different steps of the feature extraction algorithm were also added [7, 8, 18, 34, 35]. The seven parameters tested included the following: match tolerance mass, expansion values for chromatogram extraction, isotope abundance score, peak filter absolute height, spectra to include average scans, exclude TOF spectra if above, and post-processing filter absolute height. A detailed explanation of the selected parameters and the values tested for each of them is presented in the Supplementary Information. For each experiment, all parameters were fixed at an initial value and only one of them was altered at a time. Compound identification in the non-targeted mode was completed using a reduced library dataset that contained only the seven antimicrobials, which was created from previously existing libraries using the software Mass Hunter PCDL Manager for Metabolomics B.07.00 from Agilent Technologies. The library for the first part of this study was reduced to seven compounds with the purpose of speeding up the experiment, but the size of the library does not influence the identification results from the Batch Targeted Feature Extraction algorithm. Retention time was not considered when searching in the database, as it is specific to the HPLC conditions used to acquire the data. The positive ions and adducts considered were H+, Na+, and K+. Finally, only compounds with a minimum total score of 70% were considered, which is what is commonly reported for this type of analysis [7, 11].

Experimental Design

According to the literature, veterinary drug residues may be detected at a low frequency in actual honey samples. In 2013–2014, the Canadian Food Inspection Agency carried out tests for veterinary drug residues on 280 samples of domestic honey and 108 samples of imported honey in Canada, and the percentage of samples with detected residues was 26.07% and 12.04% for domestic and imported honeys, respectively [36]. As a result, we studied the influence of the number of samples above the LOD in a batch, on the correct non-targeted identification of trace compounds. Different groups of datasets were produced using different ratios of matrix blanks (spiked and unspiked samples) to mimic a range of contaminant frequencies of the veterinary drug residues in honey samples (Table 1). Each batch of data files contained 15 honey sample extracts and 5 procedural blanks, so a total of 20 data files per batch were used. Groups A, B, and C corresponded to 75%, 25%, and 5% of the samples in the batch containing detectable amounts of veterinary drug residues, respectively. The spiked samples in the three batches of groups B and C were different and were selected to represent a range of color, floral origin, packaging, and organic/non-organic farming matrices.

Table 1 Composition of the Groups of Samples Used to Optimize the Data Treatment Parameters
Full size table

Two-way analyses of variance (ANOVA) were performed using IBM SPSS Statistics 23 (NY, USA) to determine whether the different parameters tested produced statistically significant differences.

Application of the Optimized Data Treatment to the Non-targeted Analysis of Actual Honey Samples

As a proof of concept, the optimized data treatment was used to screen 55 honey samples from the Canadian market for veterinary drugs, pesticides, and other contaminants. Since no commercial library dataset was found containing only compounds related to honey, a subset from the METLIN Metabolomics Database was created based on 42 chemicals reported in the production of honey or previously detected in honey. Hydroxymethylfurfural (HMF) was not part of the METLIN Metabolomics Database, so it was manually added using the experimental exact mass and spectral data with LC-MS recorded from an analytical standard. Details on the compounds included in the library are provided as Supplementary Information (Table S1). Five procedural blanks were run together with the honey samples and were included in the data treatment. In this study, only compounds whose signal was greater than the average + 3σ of the blank signals were considered.

Confirmation of the Identity of Compounds in Honey Samples

According to the level system proposed by Schymanski et al. [37], the level of confidence of the identification for the compounds obtained using the Batch Targeted Feature Extraction algorithm is 4, since it is based on the exact mass and the isotopic signature. In order to increase the identification confidence level of these compounds, MS/MS data should be compared with that of a compound database to provide a probable structure. Ultimately, the structure should be confirmed with a reference standard to achieve level 1 of confidence, which represents the highest level. In the present study, the identity of the compounds identified by the Batch Targeted Feature Extraction was further confirmed using the parent ion and the most abundant daughter ion found with the All Ions MS/MS data with those reported in the database or in the literature. For each compound, both the presence of these ions in the MS/MS spectra and their intensity ratio were studied.

When the MS/MS spectra of the sample matched those reported elsewhere, a final confirmation was performed with analytical standards to achieve a confidence level of 1. This identification was done following the IP system established by the 2002/657/EC European Commission Decision [13]. For the confirmation of the substances listed in group B of Annex I of Directive 96/23/EC, which include veterinary drugs and other contaminants (i.e., organochlorine compounds including PCBs, organophosphorus compounds, chemical elements, mycotoxins, dyes, and others), a minimum of three IPs is required [38]. In addition to having three IPs, the 2002/657/EC European Commission Decision establishes that a minimum of one ion ratio should be measured, that all measured ion ratios need to meet maximum permitted tolerances for relative ion intensities, and that a maximum of three separate techniques can be combined to achieve the minimum number of IPs [13]. For LC or GC with mass spectrometric detection, IPs are assigned depending on the type of ion measured (i.e., parent or fragment ions) and the resolving power of the mass analyzer. However, some difficulties appear when trying to classify mass analyzers as low- or high-resolution because modern low-resolution mass analyzers can now reach resolutions comparable to the traditional HRMS instruments [39]. Alternatively, Hernández et al. [40] proposed a criterion for assignment of IPs based on mass measurement accuracy instead of resolution, which was used in the present study. For the calculation of the IPs, two ions were selected for each compound: MS1 (parent ion) and MS2 (most abundant fragment). The abundance of each ion was read at the collision energy (0 V, 10 V, 20 V, or 40 V) that produced the highest peak. In addition to the IPs, the retention time difference between the compound in honey and the analytical standard was calculated, and only differences below 0.3 min were considered a match [17].

Results and Discussion

Optimization of the Identification Workflow

The aim of this study was to describe and optimize the influence of the post-analysis data treatment parameters on the non-targeted identification of trace contaminants in honey using seven veterinary drugs as model compounds. The number of erroneous identifications (i.e., false positives and false negatives) for each parameter for test groups A, B, and C is presented in Tables 2, 3, and 4, respectively.

Table 2 Number of False Positives (+) and False Negatives (−) for Each Compound in Samples from Group A
Full size table
Table 3 Number of False Positives (+) and False Negatives (−) with Scores Above 70% for Each Compound in Samples from Group B
Full size table
Table 4 Number of False Positives (+) and False Negatives (−) with Scores Above 70% for Each Compound in Samples from Group C
Full size table

The maximum mass error between the measured ion (experimental) and the database (theoretical) to consider a match did not influence the correct identification of compounds (ANOVA, p = 0.084). In general, an increase in mass tolerance makes the search in the compound database less restrictive, and this could lead to an increase in the number of false positives. For this reason, when screening for pesticide residues in food, Mezcua et al. [18] recommended the use of a mass tolerance of ± 10 mDa for screening purposes and to lower the tolerance to ± 1 mDa and/or 5 ppm when confirming the identity of compounds. Similar mass tolerances have been reported for the non-targeted analysis of trace contaminants using a library database. Indeed, a mass tolerance of ± 5 ppm was reported in environmental monitoring studies, including the screening of environmental contaminants in honeybees and pollen, and the occurrence of xenobiotics in the blood of sea turtles [11, 41]. Herrera-Lopez et al. [12] reported a mass tolerance of ± 2.5 ppm and ± 5 ppm for MS and MS/MS data, respectively, for the screening of organic contaminants in water samples, and Sjerps et al. [34] used a mass tolerance of ± 5 ppm. Regarding food matrices, Malato et al. [7] performed a screening for pesticides in fruits and vegetables with three mass tolerances (0.6 mDa, 1 mDa, and 5 ppm) and determined that the best option according to their results was to use a mass tolerance ≤ 5 ppm when confirming the identity. Malato et al. [7] discussed the difference between the use of relative (ppm) or absolute (mDa) mass tolerance values and concluded that a relative mass tolerance value should be applied in order to avoid dependence on the search criteria with each compound mass (especially when dealing with a wide range of compounds). Another factor to consider when determining which mass match tolerance to use is the mass accuracy of the instrument. Using a match mass tolerance value far lower than the error of the acquired mass could result in an increase of false negatives, as the measured mass would not match those reported in the library. The average mass accuracies obtained in the present study were − 1.8 ppm ± 0.3 ppm, − 1.8 ppm ± 0.3 ppm, − 1.4 ppm ± 0.4 ppm, − 2.4 ppm ± 0.5 ppm, − 1.9 ppm ± 0.5 ppm, − 2.2 ppm ± 0.3 ppm, and − 2.5 ppm ± 0.6 ppm for the [M+H]+ fragments of tylosin A, tylosin B, lincomycin, furazolidone, sulfadimethoxine, sulfamethazine, and sulfamethoxazole, respectively. Therefore, a match mass tolerance of ± 5 ppm was selected as it led to acceptable results in this study, in line with what has been reported by others.

Next, the influence of the expansion value for chromatogram extraction was tested. The rate of erroneous identifications of the compounds increased significantly with the expansion value (ANOVA, p < 0.001). Indeed, up to 11 identifications were incorrect when the expansion window was increased to ± 500 ppm (Tables 2, 3, and 4). An increase of false positives was expected when increasing the extraction window, because this criterion would become less restrictive and more features may be matched with a specific formula. However, this was not observed in the present study. Instead, the number of false negatives increased, notably for tylosin B and sulfadimethoxine when applying an expansion value of ± 500 ppm. To further explore this, the signal-to-noise (S/N) ratio was assessed for each compound in their respective chromatograms extracted using ± 10 ppm and ± 500 ppm (Figure 1). The S/N ratio significantly decreased with an increasing mass extraction window for tylosin A, lincomycin, furazolidone, sulfamethoxazole, sulfamethazine, and sulfadimethoxine (ANOVA, p < 0.001, in all cases). Dasenaki et al. [42] have demonstrated that unsatisfactory fragmentation or insufficient sensitivity of fragment ions can hinder the identification of veterinary residues in fish and milk using HPLC-QTOF-MS. Using the present method, the S/N ratios were systematically above 10 for the seven trace contaminants (spike level 0.2 μg g−1; see Figure 1) and molecular features corresponding to the compounds were detected. The chemical noise or the presence of interferences could be at the origin of the false identifications, for example for tylosin B and sulfadimethoxine. There was no significant difference in the number of erroneous identifications between ± 10 and ± 50 ppm (ANOVA, p = 0.844), so ± 10 ppm was selected as an expansion value for chromatogram extraction. This value has been reported by other authors [35, 41].

Figure 1
figure 1

Comparison of the mean S/N ratio obtained from chromatograms extracted with a mass extraction window of ± 10 ppm and ± 500 ppm (m/z values for each compound are described in Table 2). Asterisk indicates a significant difference between the S/N ratios at ± 10 ppm and ± 500 ppm

Full size image

The next parameter, the isotope abundance score, had no significant impact on the identification of trace contaminants in honey (ANOVA, p = 0.334), as the same rate of correct identification was obtained for 60% and 100% of isotope abundance score (Tables 2, 3, and 4). This parameter is rarely reported in similar studies. Sjerps et al. [34] used an isotopic pattern similarity of above 90% for the analysis of different organic contaminants in water, but no information on its optimization had been reported. Instead, the literature generally focuses on the total score, which is a combination of mass match score, isotope abundance score, isotope spacing score, and retention time score. In this study, only the identifications with a total score ≥ 70% were considered, as suggested by other authors in similar studies [7, 11]. Alternatively, Knolhoff et al. [8] used an overall quality score of ≥ 50% for the non-targeted analysis of organic compounds in orange juice and milk. However, lowering the total score makes the identification less restrictive, and this could potentially lead to an increase in the rate of erroneous identifications. Since altering the isotope abundance score did not affect the identification, the default value of 60% given by the software was chosen for the final method.

Three absolute height filters were tested, and there was no statistical difference among the false identification rates (ANOVA, p = 0.556). An increase in false negatives could be expected when increasing the absolute height filter, especially for trace contaminants as their peaks may be relatively small. However, increasing this threshold from 100 to 1000 counts did not affect the rate of correct identification of the veterinary drug residues in the present study (Tables 2, 3, and 4). This study was performed at a residue concentration of 0.2 mg kg−1, and this resulted in peaks whose height was in the order of 105 to 106 counts, explaining the lack of impact of the absolute height filter. This threshold could, however, become of great importance when screening for compounds in concentrations in the low microgram per kilogram to nanogram per kilogram range. Knolhoff et al. [8] used a minimum ion peak height of 500 counts for the non-targeted analysis of organic compounds in orange juice and milk, while Mezcua et al. [18] and Malato et al. [7] used a peak filter of 100 for the analysis of pesticides in fruits and vegetables. In the present study, an intermediate value of 200 counts was finally selected.

Regarding the two parameters affecting the spectra to be included in the identification (spectra to include average scans and exclude TOF spectra if above), the different averages of scans produced significantly different results, while the possible exclusion of the TOF spectra did not affect them (ANOVA, spectra to include average scans [p < 0.001] and exclude TOF spectra [p = 0.556]). When examining closely the multiple comparisons between pairs of values, it was observed that taking the spectra at the apex of the peak or at above 30% of the peak height lead to the exact same results (ANOVA, p = 1.00), which correspond to the lowest rate of erroneous identifications (Tables 2, 3, and 4). For the rest of the study, spectra at the apex of the peaks were used in order to avoid possible interferences from co-eluting compounds on trace contaminants. Consequently, the option of excluding part of the TOF spectra was deselected.

Finally, the different post-processing peak filters were tested and did not significantly affect the correct identification of the veterinary drug residues (ANOVA, p = 0.556). As in the case of peak filters, selecting a high value for the post-processing peak filters could lead to the omission of relevant peaks and thus increase the rate of false negatives for trace contaminants. The range of post-processing peak filters used by other authors in similar studies is in the same order of those tested in the present study, going from a compound filter of only 400 counts for the analysis of pesticides in fruits and vegetables by Malato et al. [7] to up to 5000 counts for the analysis of organic compounds in orange juice and milk by Knolhoff et al. [8]. In some studies, this parameter has been expressed as a relative abundance instead of an absolute height. Indeed, Mezcua et al. [18] used a compound filter of 0.01% of the most abundant peak. However, the use of a relative abundance filter for the identification of trace contaminants, especially in a dilute-and-shoot approach such as the one used in the present study where the final extract is a very complex matrix, could lead to a significant elimination of peaks corresponding to trace-level compounds. An intermediate value of 1000 counts was selected in the present study.

The number of spiked samples in the sample batch was found to influence the rate of correct identification of the seven veterinary drugs for all tested parameters (ANOVA, match mass tolerance [p < 0.001], expansion values for chromatogram extraction [p < 0.001], isotope abundance score [p < 0.001], peak filter absolute height [p < 0.001], spectra to include average scans [p < 0.001], exclude TOF spectra [p < 0.001], post-processing filters [p < 0.001]). As shown in Table 5, the lowest amount of erroneous identifications (0.5% of false negatives and no false positives) was obtained for group C, which corresponded to a frequency of true positives of 5% (one spiked sample per batch of 20 samples). In their two most recent National Chemical Residue Monitoring Program (NCRMP) (2013–2014 and 2012–2013) reports, the Canadian Food Inspection Agency (CFIA) reported the occurrence of tylosin A, tylosin B, sulfadimethoxine, and sulfamethoxazole at rates of detection of 6.10–10.05%, 14.71–19.74%, 4.35%, and 3.26%, respectively [36, 43]. Thus, the frequency of true positives of group C would correspond to the frequency of detection of these compounds in a real-case scenario. Therefore, based on these results, the optimized identification workflow developed in the present study should be applied to the screening of veterinary drugs in honey in order to expect the lowest rate of erroneous results.

Table 5 Summary of the Rate (%) of False Positives and False Negatives Compared to the Total Number of Identifications for All the Compounds and All the Samples for Each Test Group
Full size table

Screening of Veterinary Drugs and Other Contaminants in Actual Honey Samples

Forty-three compounds were included in the library of honey-related compounds to screen 55 honey samples collected in Canada. Eight compounds were tentatively identified in at least one sample with a score above 70% and with an intensity above the mean + 3σ of the blank signals (Table 6). All Ions MS/MS data were compared to MS/MS fragmentation information from the literature or from database to investigate the identity of these eight compounds. Among the compounds with spectral information in the database (i.e., carvone, metolcarb, thiacloprid, and tylosin A), only tylosin A presented MS/MS spectra similar to those in the library. With regard to the remaining four compounds (i.e., HMF, nitenpyram, piperonyl butoxide, and rolitetracycline), only HMF was found to match the spectra reported in the literature [44,45,46,47].

Table 6 List of Compounds from the Library of Honey-Related Products Identified by the Non-targeted Workflow in at Least One Honey Sample Collected in Canada
Full size table

The identity of tylosin A and HMF was further confirmed using analytical standards according to the IP system established by the 2002/657/EC European Commission Decision [13]. Details on the identification of both compounds are presented in Table 7. For tylosin A, a total of four IPs was earned from MS1 and MS2 and the error of the measured MS2/MS1 was 1%, confirming the identity of this compound with the maximum level of confidence. Regarding HMF, a total of four IPs was earned from MS1 and MS2 and the error of the measure MS2/MS1 was 10% ± 8%, confirming the identity of this compound with the maximum level of confidence. Additionally, in both cases, the difference in retention time between the compound in honey and the analytical standard was below 0.3 min. Visual comparisons of the retention time and MS/MS spectra between the honey samples and standards for tylosin A and HMF are presented in Figures 2 and 3.

Table 7 Identification of Tylosin A and HMF Following the IP System Established by the 2002/657/EC European Commission Decision
Full size table
Figure 2
figure 2

Extracted chromatograms and MS spectra (collision energy = 40 V) of tylosin A obtained using All Ions MS/MS in a 20 ng mL−1 standard solution in methanol (a) and in the honey sample where it was identified (b)

Full size image
Figure 3
figure 3

Extracted chromatograms and MS spectra (collision energy = 0 V) of HMF obtained using All Ions MS/MS in a 10 μg mL−1 standard solution in water (a) and in one of the honey samples where it was identified (b)

Full size image

Tylosin A and HMF are both commonly detected in honey. Tylosin A is a macrolide antibiotic approved for example by the US Food and Drug Administration for emergency use in the control of American foulbrood (AFB) of honeybees, the most virulent disease known to affect these animals [48]. The CFIA, in their 2013–2014 National Chemical Residue Monitoring Program report, detected tylosin in up to 10.05% of the tested domestic honeys and 6.10% of honeys imported to Canada [36]. HMF is a heat-induced contaminant formed as a product of the Maillard reaction occurring in many food commodities, such as bread or baked goods [49]. In honey, this compound can be formed when honey is submitted to heat treatments, such as pasteurization, or a long storage time [50]. For this reason, HMF is commonly used as a quality indicator in honey [51]. In addition, HMF can be formed in food commodities even at low temperatures in acidic conditions, such as the case of honey [52].

Conclusion

The impact of the match mass tolerance, the mass extraction window, the isotope abundance score, the peak filter absolute height, the average of spectra included, the exclusion of TOF spectra, and the post-processing peak filters on the correct identification of seven veterinary drug residues (model trace residues) in honey was assessed. Among these seven parameters, the expansion window for chromatogram extraction and the average scans included in the spectra influenced significantly the identification results. Based on the present comprehensive study of the influence of each parameter and a review of values reported in the literature, the following optimized conditions were selected: match mass tolerance of ±5 ppm, an expansion value for chromatogram extraction of ± 10 ppm, an isotope abundance score of 60%, a peak filter absolute height of 200 counts, the spectra being recorded only at the apex of the peak and a post-processing peak filter absolute height of 1000 counts.

This optimized method was used to screen 55 honey samples from the Canadian market using a library of 43 honey-related compounds, including veterinary drugs, pesticides, and other contaminants. Eight compounds were tentatively identified in at least one sample with a score above 70% and with intensity significantly above the blanks. Among these compounds, tylosin A and hydroxymethylfurfural were further identified using analytical standards. The results of this study show that post-acquisition data treatment parameters can affect the identification of trace contaminants in food such as honey. Since the presence of chemical residues in food products is a matter of concern for public health, a systematic assessment of post-acquisition data treatment parameters should be conducted before applying non-targeted workflows to the identification of trace contaminants in food matrices.