Suspect screening for chemical residues in aquacultured shrimp and fish using liquid chromatography-high resolution mass spectrometry: comparison of data evaluation approaches

Abstract

High-resolution mass spectrometry (HRMS) has become an important tool for monitoring chemical residues in food, but the time and effort required to evaluate the large amount of data generated by HRMS can be a limiting factor in the widespread application of this tool. Suspect screening, i.e., searching HRMS data against large compound databases or mass lists, represents a practical compromise between using HRMS data to only look for target compounds and performing full non-target analysis. Several different approaches for suspect screening using HRMS data were tested using data from shrimp and eel spiked with veterinary drugs and pesticides as well as from imported aquaculture samples. Most of the analytes (>70%) in the spiked samples were detected and identified by searching against compound databases. To query larger databases and on-line resources such as mzCloud, it was necessary to use software capable of differential analysis and selective filtering, such as Compound Discoverer. Using selective filtering, the number of compounds detected in fish sample extracts could be reduced from tens of thousands to a few hundred by subtracting method blanks and comparing to matrix blank extracts. This smaller list of potential compounds could be further evaluated and compared to available databases and libraries. Analysis of imported aquaculture samples resulted in detection of unexpected contaminants including the dewormer levamisole, the insecticide buprofezin, and potentially the plant alkaloid ricinine.

Introduction

High-resolution mass spectrometry (HRMS) has become an important tool for monitoring chemical residues in food. Our laboratory has developed a method using HRMS to screen for veterinary drug residues in aquacultured products [1]. Antibiotics and other veterinary drugs are used in aquaculture to maintain animal health and improve productivity. Other chemicals including pesticides could also contaminate aquaculture products. This HRMS method has been applied to the analysis of veterinary drug residues in incurred and imported farmed fish samples [2] and was expanded to include other chemical contaminants such as pesticides and human drugs [3]. The method was validated [4] for the most likely ~100 analytes expected to be found in aquaculture samples and has since been used to survey imported shrimp [5].

In addition to monitoring for a specific set of validated target compounds, data obtained from analysis of fish extracts with HRMS can be further evaluated for additional chemical contaminants. This is possible because data are collected in a non-targeted manner (full-scan MS1 with subsequent MS2 data acquisition). There are many alternative approaches to collecting HRMS data depending on the intended purpose [6-9]. It has been shown that full-scan MS1 combined with variable data independent analysis (vDIA) MS2 acquisition can generate data that can be used not only to reliably detect and identify low-level chemical residues of known concern but also to search for additional chemical contaminants [6, 8].

As there are many variations of data acquisition methods, there are also many options for data evaluation. Data collected in the non-target manner described above can initially be used to detect and identify only target analytes. This has been described as non-targeted data acquisition for target analysis (nDATA) [8] and was successfully used for a large multi-laboratory study of pesticide residues in fruits and vegetables [10]. This approach was also used in our laboratory for the special assignment to monitor for veterinary drugs in shrimp [4, 5].

The evaluation of HRMS data for additional chemical contaminants has been generally described as either suspect screening analysis (SSA) or non-target analysis [11]. Non-target analysis (NTA) with HRMS data detects and identifies compounds without any prior knowledge of what may be in the sample. With NTA, compounds are detected, and chemical formulas are generated using exact mass data of precursor ions (MS1). These chemical formulas can be searched against large chemical databases such as those included in ChemSpider. Exact mass data from product ion spectra (MS2) can be then used to provide tentative identification. Because it is not practical to identify all compounds found in a complex sample, the initial results are filtered to focus on compounds that may be unique to the sample of interest [12]. For example, determining the identity of a toxin in a specific food sample known to have caused harm can be successfully done by comparison to a control sample of the same type of food. Approaches to standardize the terminology used to describe NTA, particularly for analysis of residues and contaminants in food and environmental samples, have been described recently [13, 14]. Although processing techniques for NTA are continuously improving, the amount of time needed to routinely investigate complex samples in a high-throughput laboratory may still be prohibitive.

With suspect screening analysis using HRMS data, compounds are also detected using exact mass [11, 15]. The exact mass of precursor ions found in the sample is compared to those from a specific list or compound database; product ion and retention time information can also be compared if known. For example, data may be searched against a curated list of pesticides or drugs. This can be particularly effective for monitoring chemical residues for food safety analysis as possible contaminants are generally known and not unlimited. Suspect screening represents a practical compromise between using HRMS data to only look for target compounds (nDATA) and performing full NTA. This was the conclusion of Jansen et al. [9] who found that suspect screening was useful in retrospectively analyzing data from feathers for additional veterinary drug residues. Our past work also shows the value of suspect screening as several new compounds and metabolites were detected and identified in aquacultured fish samples when searching against a fairly limited compound database generated in-house by analyzing known standards [2]. It is expected that NTA and suspect screening with HRMS will become more routine as regulatory agencies, particularly in the EU, transition to risk-based analysis where a flexible analytical scope and sampling protocol including unexpected analytes are encouraged to proactively protect the food supply [9, 16].

The purpose of this work is to evaluate different data analysis workflows for suspect screening analysis in fish sample extracts analyzed by HRMS to detect and possibly identify residues and chemical contaminants beyond the set of validated analytes. The advantages and limitations of these approaches will be discussed, and the identification of unexpected compounds found will be highlighted.

Materials and methods

Sample information

The samples, blanks, and controls used for this study are described here. The solution used to reconstitute samples (10% acetonitrile in water) was used as a solvent blank. Method (reagent) blanks consisted of 2-g of water processed using the extraction procedure described below. For matrix controls, 2-g portions of shrimp or fish that had been shown to be negative for residues of the targeted compounds were processed using the extraction procedure described below. The matrix control was matched to the samples being investigated (i.e., shrimp matrix control used for shrimp samples). Matrix spikes were a 2-g portion of tissue fortified with the mix of compounds described previously [4] and are shown in Table S1. Although this mix consisted of 99 compounds (89 detected by positive ion mode, 10 by negative ion mode), the results for this study were limited to data collected in positive ion for extracts taken through the entire procedure described below. Imported samples collected for shrimp assignments [5] or as part of regulatory compliance program [17] were also analyzed. The shrimp or fish samples used for matrix blanks and spikes were either purchased from local market or collected as part of the compliance program.

Sample preparation

The samples were prepared and extracted as described previously [1-4, 6]. Briefly, 2-g of samples was extracted with 8-mL of acetonitrile solution containing 2% acetic acid and 0.2% p-toluene sulfonic acid for 30 min using a multi-tube vortex mixer. The samples were then centrifuged (7 min, 17,000 RCF, 4 °C). A 3-mL portion of supernatant was taken through a Waters Oasis HLB Prime SPE. The eluent from the SPE was taken to near dryness using nitrogen evaporation (55 °C) and reconstituted in 400 μL of 10% acetonitrile in water. A final centrifugation step (as above) was then performed. For the initial comparison of data evaluation approaches, matrix spikes of shrimp or eel were prepared (in triplicate) by fortifying with standard compound mixture at the target testing levels (Table S1). Imported samples (2 g) were extracted in duplicate.

Data acquisition

The instrument used was a Thermo Fisher Scientific Q-Exactive High Field (QE-HF) Orbitrap high-resolution mass spectrometer (HRMS) with a heated electrospray ionization (HESI) source coupled to a Thermo Fisher Scientific Vanquish LC system. Thermo Fisher Scientific XCalibur software (V.4.1) was used to analyze samples. LC separation was performed using a Supelco Ascentis Express C18 (7.5 cm × 2.1 mm, 2.7 μm) fused-core reversed-phase column. The mobile phase consisted of 0.1% formic acid (A) and acetonitrile (B) at a flow rate of 0.3 mL/min. The LC gradient program was initialized at 5% B and held for 1.5 min then ramped to 50% B from 1.5 to 8.5 min, followed by a ramp to 99% B from 8.5 to 9 min, and then was held at 99% B from 9 to 12 min. The mobile phase was returned to 5% B from 12 to 12.5 min and the column was re-equilibrated for an additional 2 min. The total LC runtime was 14.5 min; MS data were collected for 12.5 min. The column temperature compartment was kept at 30 °C, and the autosampler tray temperature was maintained at 10 °C. The LC injection volume was 10 μL.

General HRMS heated electrospray source parameters are as follows: spray voltage, 4kV (positive ion); S-Lens RF level, 50; capillary temperature, 350 °C; auxiliary gas temperature, 325 °C; gas flow rate (N₂, arbitrary units): sheath, 50; auxiliary, 10; sweep, 0. Initially, a full MS1 scan (m/z 150–1000; resolution 60,000) was acquired followed by 11 vDIA MS2 scans with variable isolation widths as shown in Table S2. The resolution for the MS2 experiments was 30,000 and normalized collision energies of 10, 30, and 50 were used [4, 6].

Each batch of samples included the following sequence of injections: extracted matrix standards with target compounds to determine system suitability, solvent blanks, method blank, at least two matrix controls, at least two additional matrix spikes, solvent blank, and duplicate extracts of any sample.

Data analysis

Targeted compound analysis using TraceFinder (TF)

The Thermo Fisher Scientific TraceFinder Environmental Food Safety (EFS) Software (Version 4.1) Quan workflow was used for a defined list of compounds (Table S1) with 5 ppm mass extraction window using semi-quantitative analysis to estimate concentration by comparing to a one-point extracted matrix standard by response factor calibration. The analytes are identified with minimum of one fragment or product ion (within 10 ppm), isotope pattern (value ≥ 70), and retention time match (0.5 min).

Suspect screening analysis using TF

The Thermo TF Software Targeted Screening Method workflow was utilized with the following parameters: detect peaks with 3 ppm MS1 mass tolerance, 50,000 threshold override, signal/noise threshold of 100; confirm peaks with retention time within 75 s (ignore if not defined), ≥ 1 fragment (product) ion with intensity threshold ≥ 500 and 10 ppm mass tolerance (ignore if not defined), and isotopic pattern fit threshold ≥ 70%, 5 ppm, 20% deviation. Library search option was not used. When processing the batch, most often a method (reagent) blank was used for blank subtraction, but matrix blanks or solvent blanks could be used instead. Subtraction is performed with TF by selecting peaks in sample and blank that are closest to each other and within 5 ppm mass accuracy. The area of the peak in blank was then multiplied by specified amount (1000× amplification) before subtracting from area of peak in sample.

Initial suspect screening method searched against a database developed in-house by analyzing authentic standards (N ~ 500) using the LC-HRMS method. The next level of suspect screening with TraceFinder used a larger vendor Thermo EFS HRMS database (N ~1500) supplied with instrumentation software. Processing of the raw data using the TF targeted screening workflow took less than 1 min per sample, even with the larger database. Sorting through the results manually can take several hours or longer because depending on the number of compounds in the database, there will be a significant number of potential hits. These compounds were initially sorted to prioritize those without any red flags for discrepancies in retention time, fragment (product) ions, or isotope pattern. This prioritized list was visually inspected to select analytes that had reasonable (Gaussian) peak shape, similar area counts between replicates, and area counts ≥ 10⁶, and were not in all samples as pervasive background.

Suspect screening analysis using Compound Discoverer (CD)

Initial data processing with the Thermo Fisher Scientific Compound Discoverer (Version 3.3) was done using a workflow based on Food Research w Stats Unknown ID w Online and Local Database Search template. A workflow diagram is shown in Figure S1 and important parameters for the workflow nodes are outlined in Table S3. This is similar to how data were processed previously [6], but some changes were made due to updated software and to achieve more focused results. Background compounds are identified with CD by setting the maximum allowed ratio of the sample vs. blank (5:1 ratio) to be considered background. Data for the compounds considered to be background are still generated but can be easily filtered out during later processing steps. To evaluate how well analytes from a matrix spike extract were detected and identified, a method (reagent) blank was used to mark background compounds and then matrix spikes were compared to matrix blanks. When data from imported samples were evaluated, a matrix blank was used to mark background compound, and then, samples were compared to matrix spikes. This helped eliminate detection of compounds that were due to low-level carryover from spikes and filtered out what was unique to that sample. Processing raw files with CD generally took less than one hour for each full scan/vDIA file in a batch if ChemSpider searching was not included.

Post-process filtering of results files was used. For example, results could be filtered to show compounds with a match to mass lists and/or mzCloud after ratioing peaks to matrix blank after background subtraction. After post-processing filtering, remaining compounds could be visually inspected to identify presumptive positives using similar principles as described above for TF such as similar results between replicates and reasonable chromatographic peak shapes. This final evaluation process using CD can also take several hours for a given batch but would not be possible without application of post-processing filters.

Follow-up to confirm suspect positives

Regardless if presumptive positives are found using TF or CD, additional steps may be needed to help tentatively identity the compound. For example, additional HRMS analysis of extracts using a more targeted data analysis such as parallel reaction monitoring with precursor isolation will produce cleaner MS2 spectra for a stronger mzCloud match. Data can be further evaluated manually using programs such as Thermo QualBrowser or FreeStyle. Reference standards can be purchased and analyzed retrospectively using method conditions to compare retention times and product ions spectra found in fish extracts.

Results and discussion

Testing data analysis approaches with matrix spikes

An LC-HRMS method was previously developed, optimized, and validated to screen for veterinary drug residues in aquaculture samples using non-targeted data acquisition and targeted data analysis (nDATA) [1, 6]. The sample extraction, chromatographic, and data acquisition parameters for this method were not modified. The purpose of this study was to investigate different approaches for further data analysis. Data from triplicate extracts of shrimp or eel that had been fortified with target compounds (Table S1) were evaluated using TraceFinder and Compound Discoverer software provided with the Thermo instrumentation. Other HRMS platforms and vendors have similar data analysis programs. For example, suspect screening of plastic-related chemicals in fish was reported using a LC-quadrupole time-of-flight instrument with targeted and non-targeted data analysis workflows within Agilent software [18]. Another paper describes optimizing data analysis strategies using milk samples spiked with model compounds analyzed with a Bruker UHPLC-TOF-MS and an open source software package designed for metabolomics (TracMass 2) [19]. Although the results presented in this study were generated with Thermo software, the strategies regarding type of data collected, database selection, and filtering parameters should be applicable to other platforms and software packages as well.

Suspect screening using limited compound databases (Trace Finder)

This study verified previous work that demonstrated spiked compounds in fish extracts analyzed by full-scan MS/vDIA HRMS could successfully be detected and identified using TF quantitative workflow for a specific list of compounds [1-4, 6]. In this study, of the 89 positive ion analytes spiked into 2-g of shrimp or eel samples, 93–97% were identified when comparing to an extracted one-point standard and using established US FDA criteria for exact mass data [20]. Specifically, to be considered identified, analytes must match retention time (±0.2 min) and the exact mass of precursor ion (≤ 5 ppm) and one product ion (≤ 10 ppm) of an authentic standard.

As a first test of suspect screening analysis, the TF workflow for “targeted screening” was used to generate lists of compounds that matched the exact mass of precursor ions for chemicals in a compound database. Data for the exact mass of product ions, isotopic patterns, and retention times were used to prioritize the lists. With this approach, a majority (≥70%) of the 89 spiked compounds were detected when compared to either the in-house or vendor compound database (Table 1) for both shrimp and eel extracts. The number of target compounds found was consistent (within 1–2 analytes) between the three replicates for each matrix. Using the in-house database, detected compounds were identified with exact mass of precursor and product ions as well as retention time match. Compounds were tentatively identified against the vendor database by comparing exact mass of at least one of 4–5 product ions (no retention time information).

Table 1.

Suspect screening results for matrix spikes using Trace Finder

In-house compound database			Vendor compound database
	Shrimp matrix spike	Eel matrix spike		Shrimp matrix spike	Eel matrix spike
# Compounds match MH+ m/z	254	274	# Compounds match MH+ m/z	921	971
# Compounds match MH+ m/z, product ion, isotope, and RT	96	86	# Compounds match MH+ m/z, product ion, isotope (no RT)	420	395
# Target compounds identified (of 89)	76 (85%)	62 (70%)	# Target compounds identified (of 89)	66 (74%)	63 (71%)

Average of triplicate extractions, relative standard deviation for # of compounds was < 3%

Some of the false negatives that occurred with the TF target screening workflow were the same as analytes missed when quantitative TF workflow was used (dyes, avermectins). There were approximately 20 compounds (metabolites, dyes, sodium adducts) from the spiking mix that are not included in the vendor database. When retention time matching was required for identification with the in-house database, false negatives could result when peaks shifted over time from the specified value. However, not including retention time criteria in the search greatly increased the number of false positives.

Reviewing the lists of potential analytes detected by HRMS to differentiate between false positives and compounds that should be investigated further is an important and often challenging step in data evaluation. Even with the smaller in-house database, a large number (150–200) of compounds matched the exact mass of a precursor in the database, so using product ion information was critical. Once the data were filtered to match product ion, isotopic pattern, and retention time, a relatively small number of compounds (N= 20–24) were found in addition to the target analytes that had been spiked into the samples. Many of these additional compounds were related to the those in the spiking mix including epimers of tetracyclines or degradation products of β-lactam or macrolides.

Without the advantage of pertinent retention times, 850–900 false positives were found with the larger vendor compound database, approximately 40% of these (420 for shrimp and 395 for eel) also matched product ions and isotope patterns in addition to matching the exact mass of the precursor (Table 1). Evaluating this large list of compounds found in the larger database with TF was not practical, and it is possible that some of these compounds may be true positives. The number of product ions that matched these compounds varied between only one to all of those listed in the database. The quality of isotopic match also varied compound to compound and would need to be evaluated for each one. The area counts (i.e., significantly more area counts in spikes or samples compared to matrix blank and consistency of area counts between replicate samples) as well as quality of chromatographic peak shape are critical to evaluate when searching against larger databases. Using a method (reagent) blank to subtract background compounds removed some of the false positives (before background subtraction, there were an additional 100 false positives), but there were still more than could be evaluated in a timely manner. This software does not effectively provide a way to ratio sample types (e.g., compare matrix spikes to matrix controls). For these reasons, searching larger databases with TF was not an efficient approach for initial data analysis.

Next level suspect screening (Compound Discoverer)

Compound Discoverer (CD) is a software platform designed specifically for NTA and suspect screening of small molecules for clinical, environmental, and food safety applications. CD can efficiently filter data to reduce the number of compounds found in a sample. When analyzing HRMS data with CD, the initial data processing will determine how many compounds are found in a dataset and then post-process filtering can be used to limit the number of potential hits. This is illustrated using the 2-g shrimp and eel matrix spike extracts as a test case. To maximize the number of low-level residues that are detected in the matrix spikes, it is better to initially cast a wide net. Some restrictions (e.g., minimum peak intensity and group peak ratings) were needed to allow the data files to be processed in a reasonable timeframe. Aligning retention times and using chromatographic signal/noise thresholds can minimize the detection of broad background signals. This process can still lead to a large number (~20,000) of detects in complex fish and eel extracts. If the number of detects can then be reduced to 50–100, the list can then be prioritized by area counts and matches to mass lists or mzCloud to expedite manual review. Changing the variables of the results filtering can determine how many compounds are detected and potentially identified. This post-processing filtering can be redone at any time for additional retrospective screening. This type of data reduction has been shown to be critical for NTA [12] and can be very helpful for suspect screening as well.

The effects of systematic results filtering for data from matrix spike shrimp and eel extracts evaluated using CD are shown in Table 2. Initially, using the workflow described in “Materials and methods” (Fig S1), approximately 14,000–20,000 compounds are listed for shrimp and eel, respectively. Using a method blank to mark and remove background components, this number can be reduced by 40–50%. Eliminating early eluting compounds (≤1.2 min) can further reduce the list of compounds slightly (another ~10%) in the initial stage but was not determined to be a critical element in subsequent steps of data analysis. Ratioing the spike matrix to control matrix (Log2 fold ≥ 2 or 4× change) effectively decreases the lists for both shrimp and eel extracts to ~ 250 compounds that can then be further evaluated. The first level of suspect screening using CD would be to compare to known mass lists. The mass lists used for CD were the same set of compounds used in TF (from both the in-house and vendor compound databases), but only MS1 data were compared as the product ion information is not transferred when importing mass lists. Retention time matching was not used as relevant information was not available for the vendor database. Because mass lists do not include product ions, this CD screening was only used for analyte detection. Product ion data would be needed to meet identification criteria [20]. Nevertheless, Table 2 shows that 83 (shrimp) or 73 (eel) compounds were detected when comparing MS1 data to the two mass lists simultaneously. Most of these compounds, 62 (shrimp) or 56 (eel), corresponded to the target compounds that had been added to the samples. Others that matched the mass lists were, as with TF, related to the target analytes (epimers, degradation products, multiple peak detections). The ~ 180 compounds that were more (4×) abundant in the spike compared to the matrix blank but not annotated were not identified using the current workflow. Changing the filter to require a larger difference (16×) between matrix spike and matrix blank decreased this significantly, but an additional 15–20 target analytes were also not detected.

Table 2.

Suspect screening results for matrix spikes using Compound Discoverer

Result filters applied						Shrimp matrix spike		Eel matrix spike
Subtract method Blk	RT≥1.2	≥4X Matrix Blk	≥ 16X Matrix Blk	Match mass list	Partial match mzCloud	#Compounds found	#Target compounds (of 89)	#Compounds found	#Target compounds (of 89)
						13,955	Not evaluated	20,785	Not evaluated
X						6519	≤ 64	11,779	≤ 54
X	X					5017	≤ 64	10,161	≤ 55
X		X				269	66 (74%)	251	68 (76%)
X			X			120	51 (57%)	119	42 (47%)
X	X	X				227	66 (74%)	244	68 (76%)
X	X	X		X		81	62 (70%)	73	56 (63%)
X		X		X		83	62 (70%)	73	56 (63%)
X	X	X			X	40	30 (34%)	25	20 (22%)
X		X			X	40	30 (34%)	25	20 (22%)
X		X		X	X	86	64 (72%)	76	57 (64%)

Triplicate extractions of matrix spikes or matrix blanks were grouped for comparison

A larger suspect screening list that can by searched with CD is mzCloud, a curated on-line database with product ion spectra which is routinely updated and currently has over 21,000 compounds [21]. In theory, searching mzCloud with CD could expand the search beyond mass lists. However, in this study, the number of fortified compounds found in matrix spikes by mzCloud was less than that found by comparing to mass lists using CD. For example, only 30 target compounds (34%) in shrimp or 20 (22%) in eel were identified in spiked extracts by mzCloud using vDIA MS2 data. The target compounds that were identified tended to be those with larger area counts because the spiking levels were higher and/or they had higher ionization efficiency. An example is fenbendazole identified with mzCloud in an extract from shrimp matrix spike (50 ng/g) with a match score of 90 (Fig. 1). A higher match score indicates a better match of experimental data to available library product ion spectra in mzCloud. The lower number of target compounds identified using mzCloud illustrates the limitations of acquiring MS2 data using vDIA. Because individual precursor ions are not isolated prior to collisional dissociation, the resulting MS2 spectra represent product ions from a range of precursors and are less amenable to mzCloud search. This is consistent with previous reports [6, 9]. When compounds are found, the match score can be lower due to interference from non-related ions in the product ion spectra. However, the mzCloud search did find one or two spiked compounds that did not match the mass lists. Filtering using a combination of matches from the mass lists and mzCloud proved to be the most effective with 72% and 64% of spiked compounds detected in shrimp and eel, respectively. In addition, the real-life examples provided below show that mzCloud can still generate candidates with matching spectra for unexpected compounds in complex food matrices. This can provide justification to conduct further investigation such as re-analysis with more targeted data acquisition or purchasing of authentic standards.

Fig. 1.

(a) Extracted ion chromatograms from Compound Discoverer for MH⁺ of fenbendazole (m/z 300.0803) in extracts from three replicates of eel spiked with target compounds. The spiking level for fenbendazole was 50 ng/g. (b) mzCloud match for fenbendazole in spiked eel sample (top) compared to mzCloud library (bottom)

The CD program is, of course, also used extensively for true NTA [12]. For this application, additional nodes in the workflow are important including predicted composition to generate molecular formulas with subsequent searching of the large databases within ChemSpider. For this study, the predicted composition was included in the CD workflow. Predicted composition can specify minimum and maximum predicted elements and degree of saturation as well as relative intensity tolerance for molecular isotopic patterns. A partial match to predicted composition could be used to filter out compounds; however, subtle isotope patterns may not match criteria at very low analyte levels in complex samples. It was helpful to flag obvious mismatches in annotation (e.g., presence or absence of halogens). Initially, CD workflows included searches against relevant ChemSpider databases, and some target analytes, including fenbendazole in fortified shrimp, were found with this tool. However, in addition to fenbendazole, 30 other compounds were found in that sample with the same predicted composition and elemental formula. The data processing time with ChemSpider searching took much longer and sorting through the resulting data for compounds found by ChemSpider was beyond the scope of the study, so faster CD workflows without ChemSpider searching were more often used for suspect screening.

Application of suspect screening to aquaculture samples

Several additional chemical contaminants in aquacultured samples were found using suspect screening. Initially, an analyte, the fungicide boscalid, that was not included in the validated spiking mixture was fortified into shrimp at 10 and 100 ng/g to evaluate the different data evaluation approaches. As shown in Table 3, this compound was found at both fortification levels using all data evaluation approaches, including in both databases (TF) and mass lists (CD) and with a mzCloud quality match score of 84 or 61 for boscalid fortified in shrimp at 100 or 10 ng/g, respectively.

Table 3.

Summary of findings for suspect screening of aquaculture samples

Initial Finding		TraceFinder (TF)		Compound Discoverer (CD)				Follow-up with Authentic Standard			Certainty
Compound	Sample	Find with In-house DB1 (Number of fragments)	Find with Vendor DB2 (Number of fragments)	Match In-house Mass List1 (Δ ppm)	Match Vendor Mass List2 (Δ ppm)	Ratio Sample/ Matrix Spike3	Match mzCloud (Match Score)	Match Ret Time (Δ min)	Match MS2 # Productions	Estimated Amount (ng/g)	Confidence Level
Boscalid 100 ng/g	Fortified Shrimp	YES (1)	YES (4)	YES (0.69)	YES (0.69)	1521	YES (84)	YES (0)	YES (3-4)	> 90% recovery 4	Confirmed
Boscalid 10 ng/g	Fortified Shrimp	YES (1)	YES (4)	YES (0.54)	YES (0.54)	188	YES (60.6)
Buprofezin	River Barb	NI	YES (4)	NI	YES (−0.31)	28	YES (41.7)	YES (0.03)	YES (4)	1-2 ng/g5	Confirmed
Levamisole	Eel	YES (2)	YES (4)6	YES (1.12)	YES (1.12)6	46	NO7	YES (0.05)	YES (1-2)	3 ng/g5	Confirmed
Antipyrene	Shrimp	NI	YES (3-4)	NI	YES (0.33)	21	Yes 8 (42.7)	YES (0.04)	Weak ions not observed	< 2 ng/g4	Uncertain
Ricinine	Eel	NI	YES (5)	NI	YES (−0.11)	30	YES (35.1)	YES (0.11)	Partial (1)	< 4 ng/g4	Uncertain
Ricinine	River Barb	NI	YES (5)	NI	YES (0.37)	16	NO	(0.02)	Partial (1)	< 1 ng/g4	Uncertain
Citrinine	Tilapia9	NI	YES (1)	NI	YES (0.33)	50	YES (35.4)	NO (4.2)	NO	NA	Ruled out
Decarboxyl Ofloxacin	Eel	YES (0)	NI	YES (−1.36)	NI	14	YES (42.2)	NO (3.1)	1 common (later RT)	NA	Ruled out

¹Compound database generated in-house. Majority of entries include experimental retention times and product ions which are used for TF (data shown in blue), but not CD (data shown in gold)

²Thermo Environmental Food Safety compound database. Product ions are used for TF (data shown in blue), but not CD (data shown in gold)

³Area ratio of compound in sample compared to corresponding matrix (i.e., shrimp, eel) spiked with validated compounds mix (Table 1S)

⁴Amount estimated by comparing to solvent standard of authentic compound

⁵Amount estimated by comparing to matrix spike using authentic compound (spiked shrimp for buprofezin, eel for levamisole)

⁶Found in Thermo EFS database/mass list as “tetramisole”

⁷mzCloud did match eel sample fortified with 10 ng/g of levamisole (41.8 quality match)

⁸Found in mzCloud as “phenazone”

⁹Citrinine was flagged in many samples, this tilapia sample was investigated as an example

The other compounds detected by suspect screening listed in Table 3 were found in imported aquaculture samples. One example of an unexpected chemical contaminant detected and identified by suspect screening analysis of HRMS data was the insecticide buprofezin in imported river barb fish. This was not found when using the in-house database/mass list because it was not included. When searching against the vendor database, the compound was identified using TF with 4 matching product ions. It also matched the vendor mass list using CD analysis. The MS2 spectra collected for the fish extract using vDIA were sufficient to match buprofezin with mzCloud quality match score of 42 (Fig. 2). Based on this information, an analytical standard of the insecticide was later purchased and analyzed using the LC-HRMS method. The retention time and product ions spectra matched the compound found in the fish sample. The insecticide was also fortified into shrimp (at 100 and 10 ng/g), taken through the extraction method, and analyzed. The recoveries for this compound through the method were > 70% as compared to solvent standards. The amount of buprofezin in the imported river barb sample, compared to fortified shrimp data, was estimated to only be 1–2 ng/g. However, the signal (ionization efficiency) for the analyte was quite high relative to most compounds in the method allowing for intense product ion spectra for successful mzCloud searching. Buprofezin is used for rice, cotton, and other crops [22]. This compound is included in the US FDA’s pesticide methods and was found 35 times in a variety of matrixes in 2020 [23]. The fact that this compound is identified at low levels in imported fish serves as an illustration of how suspect screening using HRMS data can detect and identify additional compounds, even at low levels.

Fig. 2.

(a) Extracted ion chromatograms from Compound Discoverer for MH⁺ of buprofezin (m/z 306.1634) in extracts from two replicates of imported river barb fish. (b) mzCloud match for buprofezin in river barb samples (top) compared to mzCloud library (bottom)

In another example, the dewormer levamisole (a levorotatory enantiomer) was found in duplicate extracts of imported eel samples by TF and CD when comparing to both in-house and vendor databases/mass lists. Using TF suspect screening with the in-house database, levamisole was detected and identified with two product ions and matching retention time. When the vendor database was used with TF, “tetramisole” (the racemic mixture) was detected and identified with 4 product ions (retention time not available). The compound was detected by CD when compared to mass lists, but not identified by mzCloud. When eel tissue was fortified with 10 ng/g levamisole, the compound was identified by mzCloud at the same retention time as the levamisole/tetramisole peak from the imported eel (Fig. 3). When compared to the matrix spike, the estimated concentration in the imported eel was ~ 3 ng/g. Identification of levamisole was later confirmed using more targeted parallel reaction monitoring HRMS acquisition as well as using a triple quadrupole LC-MS/MS method with different chromatography (data not shown). There is precedence for using this anthelmintic in eel [24, 25].

Fig. 3.

(a) Extracted ion chromatogram (EIC) from extract of imported eel matching MH⁺ of levamisole (m/z 205.0794) with in-house database including 2 product ions. (b) EIC from extract of imported eel matching tetramisole (m/z 205.0794) with vendor database including 4 product ions. (c) EICs from Compound Discoverer for MH⁺ of levamisole in duplicate extracts of imported eel (blue trace) and eel matrix spiked with 10 ng/g levamisole (green trace)

In other instances, efforts to identify compounds detected using suspect screening were inconclusive. For example, antipyrine, an analgesic also known as phenazone, was detected in a shrimp sample with CD matching the vendor mass list (antipyrine), predicted composition, and mzCloud (phenazone with match score 43). A reference standard was purchased and analyzed. Antipyrene eluted at a retention time (~ 5 min) which corresponded to the peak found by suspect screening of the shrimp sample. However, antipyrine is a small stable molecule with precursor ion at m/z 189.1022 that did not form significant product ions with these LC-HRMS conditions (product ions were included in TF vendor database, but were not readily observed in mzCloud entries), so it was not possible to unequivocally confirm the identity of this analyte (Fig S2).

Ricinine is another potential analyte that was detected in extracts of eel and river barb samples using suspect screening. It was detected in the eel sample with TF using vendor database (matched 5 product ions). Using the CD software, ricinine matched to the vendor mass list and mzCloud, but the mzCloud quality match score was low (35) (Figure S3). Other target compounds, including ofloxacin, 2-amino mebendazole, and ciprofloxacin, were identified in this eel sample with much higher quality mzCloud matching. A mixture of standards used for toxin analysis that included low levels of ricinine was available for analysis, and it was determined that the retention time for ricinine was within 0.1 min of the compound found in eel. A low abundance product ion at m/z 138.0550 was listed in mzCloud and was found experimentally in both the reference standard and eel sample (Fig. 4). Ricinine was also detected in a river barb sample at lower area counts. In the river barb sample, it was detected by CD when compared to vendor mass list but did not match mzCloud (either no match indicated or matched isomeric compound, 5-nitroindoline). Ricinine is of particular interest because it is an alkaloid by-product of castor cakes. These cakes are sometimes used as an agricultural fertilizer so it is plausible that ricinine could be found in fish [26-28]. Ricinine may serve as a biomarker for ricin, a toxic protein present in castor oil and its products. Further work to further investigate and monitor for residues of ricinine, for example, more targeted MS analysis utilizing alternative chromatography, will be required to determine if residues of this compounds are present or not.

Fig. 4.

Extracted ion chromatograms from Thermo FreeStyle for ricinine in (a) solvent standard, 100 ng/mL and (b) extract from imported eel sample. EICs (5 ppm window) for MH⁺ in full MS1 scan and MH⁺ and product ion in MS2 scan (vDIA from precursors m/z 149–201)

A few other compounds that were flagged using suspect screening could be ruled out after subsequent analysis of reference standards. This includes citrinine which was detected in many different types of fish samples including tilapia by matching vendor mass list with CD and mzCloud with a low match quality score (35) at a retention time of 4.4 min. However, a purchased standard of citrinine analyzed with this method eluted at 8.6 min with higher abundance of diagnostic product ions. Decarboxy ofloxacin, a proposed degradant of ofloxacin, was detected with a mzCloud quality match score of 42 in an eel sample contaminated with high levels of ofloxacin. The retention time of decarboxy ofloxacin found in the eel sample with CD suspect screening was the same as ofloxacin, but the analytical standard purchased and analyzed separately eluted 1.7 min earlier. This may indicate that the decarboxy ofloxacin found by CD resulted from the degradation of ofloxacin after it eluted from the LC column and entered the heated electrospray source. There are common ions in the mzCloud product ion spectra for ofloxacin and decarboxy ofloxacin. These examples show why it is important to follow up with an authentic reference standard rather than just relying on software results. Comparison to an authentic standard to confirm the identity of a suspect finding is required in HRMS guidance [20] and best practice documents [13, 14, 29].

Shrimp assignment

The optimized data evaluation approach was further applied to 100 samples collected as part of special assignments to enhance the safety of imported shrimp [5] and analyzed using the full scan/vDIA LC-HRMS method. Other than the potential antipyrine residue described above, suspect screening analysis using both TraceFinder and CD with in-house and Thermo compound databases (mass lists) did not detect any other contaminants at a significant abundance and confidence level (matching product ions, retention times, predicted composition). Although it is impossible to definitively prove that no other contaminants of concern are present in these samples, collecting and evaluating the HRMS data provided further assurance that the shrimp were free of harmful residues.

Limitations and further work

It was shown in a previous study [6] that full scan/vDIA data acquisition allows for confident identification of low-level target analytes of regulatory concern in complex matrices while generating product ion data for other compounds in the sample. In the same study, data-dependent MS2 (ddMS2) generated cleaner product ion spectra, but up to 30% of the target compounds spiked in fish at residue levels did not meet the criteria to initiate MS2 scans. However, the quality of the product ion spectra from vDIA scans is not ideal for mzCloud searching as compared to data acquisition methods with initial precursor isolation such as ddMS2. Improvements in data acquisition techniques, such as Acquire X, which generate automated on-the-fly matrix-related exclusion lists through multiple injections of a sample to provide better coverage using ddMS2 [30], may be worth further investigation. Also, statistical programs such as CD will produce better results with more replicates for each condition or study variable. However, generating large numbers of analytical replicates may not be practical for routine laboratory applications. For the imported fish samples analyzed in this study, two analytical replicates were generated and analyzed (two portions of sample were extracted). These duplicates were used to provide minimum number of files for analysis with CD.

Conclusions

It was determined previously that the nDATA approach, i.e., acquiring data using non-targeted data acquisition but searching for targeted veterinary drugs [4, 6] or pesticides [8, 10], is an effective monitoring strategy for food safety. Searching against a database for additional analytes using relevant retentions times with a software program such as TraceFinder can be an effective next step. To search larger vendor-supplied databases/mass lists or on-line resources such as mzCloud, it is preferable to use a program that is capable of some differential analysis and selective filtering, such as Compound Discoverer, to efficiently rule out many potential hits. Effectively filtering the number of compounds detected by HRMS to minimize the time spent manually evaluating results and avoiding false positives is an important and challenging step in data evaluation. Data analysis approaches were evaluated using data from extracts of fortified shrimp and eel. They were then applied to imported aquaculture samples resulted in detection of new and unexpected contaminants including the dewormer levamisole, the insecticide buprofezin, and potentially the alkaloid ricinine. Using a combination of these data evaluation approaches, suspect screening can value to data already collected for target compound analysis by expanding the scope of chemical contaminants that can be detected and identified.