Analysis of decomposition in 23 seafood products by liquid chromatography with high‐resolution mass spectrometry with sensory‐driven modeling

Abstract Samples of 23 seafood products were obtained internationally in processing plants and subjected to controlled decomposition to produce seven discrete quality increments. A sensory expert evaluated each sample for decomposition, using a scale of 1–100. Samples were then extracted and analyzed by liquid chromatography with high‐resolution mass spectrometry (LC‐HRMS). Untargeted data processing was performed, and a sensory‐driven Random Forest model in the R programming language for each product was created. Five samples of each quality increment were analyzed in duplicate on separate days. Scores analogous to those obtained through sensory analysis were calculated by this approach, and these were compared to the original sensory findings. Correlation values (r) were calculated from these plots and ranged from 0.971 to 0.999. The finding of decomposition state of each sample was consistent with sensory for 548 of 550 test samples (99.6%). Of the two misidentified samples, one was a false negative, and one false positive (0.2% each). One additional sample from each of the 1st, 4th, and 7th increments of each product was extracted and analyzed on a third separate day to evaluate reproducibility. The range of these triplicate calculated scores was 15 or less for all samples tested, 10 or less for 63 of the 69 triplicate tests (91%), and five or less for 41 (59%). From the models, the most predictive compounds of interest were selected, and many of these were identified using MS2 data with standard or database comparison, allowing identification of compounds indicative of decomposition in these products which have not previously been explored for this purpose.


| INTRODUC TI ON
The United States Food and Drug Administration (USFDA) routinely samples and tests seafood products for decomposition, or spoilage, to ensure consumer safety and sanitary production conditions (USFDA, 2010). Samples are analyzed primarily by sensory analysis according to guidelines developed in conjunction with the Canadian Food Inspection Agency (CFIA) and the National Oceanic and Atmospheric Administration (NOAA) (FAO, 1999). However, a major drawback to sensory is the extensive training required (ASTM, 1981;Labbe et al., 2004;USFDA, 2013). Uniformity of testing results is also a potential concern (Naes, 1990;Rossi, 2001;Wilkinson & Yuksel, 1997). Alternative testing procedures currently in use are limited to analysis of histamine (AOAC, 1987) or indole (AOAC, 1982). Histamine content is typically only useful for particular products (Hungerford, 2010), and indole is indicative only of warmer temperature decomposition in shrimp (Chang et al., 1983).
Since neither of these measures decomposition in a way that can be directly compared with sensory results, there is a need for an alternative technique.
Following an initial proof of concept limited to six species of salmon (Self et al., 2019), there has been Congressional interest in sensory alternatives for regulatory use (U. S. Congress, 2018). This has enabled the current study, which improves and greatly expands on that previous work.

| Experimental overview
This study involves the use of sensory-driven computer models based on liquid chromatography with high-resolution mass spectrometry (LC-HRMS) data for 23 internationally sourced seafood products. Samples were first analyzed by sensory and then immediately stored in −20°C (±2°C) conditions until further use. These were subsequently thawed, homogenized, and extracted using a modified "Quick, Easy, Cheap, Effective, Rugged, and Safe" (QuEChERS) technique (Anastassiades et al., 2003). Extracts were analyzed with LC-HRMS, with untargeted data processing.
Sample responses were used in conjunction with sensory data to create statistical models. This enables the calculation of a "decomposition score" which is analogous to that generated by sensory analysis. These model-produced scores were then compared with the original sensory data.

| Sample collection and development
Sampling teams, each led by a USFDA certified National Seafood Sensory Expert (NSSE), were deployed to seafood packing facilities in Kodiak, Alaska (USA); Guayaquil, Ecuador; Georgetown, Guyana; and Huy Toa, Vietnam, to collect samples of 23 seafood products (Table 1). Products were sampled in the freshest possible state and subjected to controlled decomposition on ice onsite. Samples were removed from ice at sensory-controlled timepoints to create discrete quality increments. The general strategy was to create seven such increments, ranging from the freshest available (1) to very advanced decomposition (7). However, eight such increments were collected for swordfish (without CO), and nine for canned tuna, at NSSE discretion. Canned tuna samples were then canned, and all others were vacuum-sealed and stored at −20°C (±2°C) until further use. Five sample portions (approximately 200 g) of each increment were used in the study.

| Sensory analysis
Following official policy implemented in USFDA regulatory laboratories (USFDA, 2010), sensory evaluations were made using a single, highly qualified expert (NSSE) in lieu of a sensory panel. Assessments were made according to established procedures (FAO, 1999), based on both quality and intensity of odor characteristics. In this way, a numerical score was assigned on a 100-point scale, with greater values indicative of lower quality. Scores below 50 are considered nondecomposed, and greater than 50 decomposed. For samples consisting of many small pieces (e.g., scallops, shrimp, squid), assessments were made for each sample in bulk, and care was taken to remove any strongly outlying pieces, although this was rare.

| Extraction
Samples of scallops, shrimp (peeled), and squid were added to an approximately equal mass of dry ice and blended to a powdery consistency in an industrial-grade blender as described in the literature (Bunch et al., 1995) and stored overnight at −20°C to allow CO 2 to sublimate. All other samples were skinned if present, then blended to a uniform consistency using a consumer-grade food processor.
Samples of these composites (10.0 ± 0.5 g) were transferred to 50-mL centrifuge tubes and extracted via a QuEChERS (Anastassiades  et al., 2003) technique, modified as in previous work (Self et al., 2019).

TA B L E 1 List of products sampled
Resulting extracts were mixed with an equal volume of high-purity deionized water and filtered via a 0.45µm syringe filter prior to analysis. Extraction was performed in duplicate on separate days for all samples, using the same composite. Additionally, a third replicate on a third day was prepared for one sample each from the 1st, 4th, and 7th (or highest) sets (low, borderline, and high decomposition states) of each product to establish reproducibility across the range.

| HRMS conditions
Positive mode electrospray ionization (ESI) was used, with a probe temperature of 400°C, at position C,1.0,0. The spray voltage was 3.00 kV, inlet capillary temperature was 380°C, and the S-lens RF level was 65.0. Sheath, auxiliary, and sweep gasses were set to 60, 30, and 10 units of N 2 , respectively. A full MS scan was taken with the 60,000-resolution setting, followed by an all-ions fragmentation (AIF) scan, at 30,000 resolution, using normalized collision energies (NCE) of 10, 40, and 60. Each of these scanned from m/z 80-1200 with an automatic gain control (AGC) target of 3 × 10 6 and maximum injection time of 200 ms. Additional analysis was performed on selected samples and standards to assist with compound identification, which employed the same full scan settings with data-dependent MS 2 (DDMS2) and NCE as described for the AIF scan. These either included a target list of interest or targeted the top five ions in each scan, as needed.

| Data processing and analysis
2.6.1 | Untargeted analysis Sample data were subjected to an untargeted workflow using Compound Discoverer® 3.1, which included retention time alignment, grouping, feature merging, and gap filling. Each product was treated separately, and samples were segregated between decomposed, with sensory scores greater than 50, and nondecomposed, with scores less than 50. Differential analysis was performed using a volcano plot, with a log 2 fold change greater than one, and compounds with the 100 lowest p-values were selected to find those most relevant to decomposition.

| Modeling
Peak area data for the 100 compounds of interest for samples in each product set were exported from Compound Discoverer and combined with the original sensory score data. These were incorporated into a script in the R (3.6.3) programming environment. One third of the samples, rounding up, were randomly selected to be the test set, using a "set.seed" value to ensure reproducibility, while the remaining samples comprised the training set. A data matrix for each product was formed using peak area counts for each compound as dependent variable columns, sensory scores as the independent variable column, and one row for each training sample. These matrices were then subjected to the Random Forest algorithm (Breiman, 2001), using regression with 2000 trees and fitting to the sensory data. This initial model was used to evaluate the predictive power of each compound and optimize the number of compounds used for further modeling.
Compounds were ranked by predictive power, and then, the number of compounds used to model was iterated and optimized with respect to correlation values (r) between sensory and modeled scores. This optimized compound list for each product (

| Accuracy
Regulatory sensory analysis is reported as a qualitative pass/fail finding (USFDA, 2010), but the scoring technique as described in section 2.3.2 above is useful in training and for comparison between analysts. These sensory scores were used to train the models and are also useful in evaluating the modeling method, although this should also be considered a primarily qualitative technique.
Correlation (r) between model-produced and sensory values was calculated for each product test set. These ranged from 0.971 to 0.999 (Table 2). A false-positive finding was defined as any finding for which the sensory score was less than 50 (i.e., nondecomposed) while the calculated value was greater than 50 (decomposed). A sensory score greater than 50 with a calculated score less than 50 was likewise considered a false-negative finding.
One of each of these was found in all test samples (Figure 1

| Reproducibility
As discussed in section 2.4 above, triplicate analysis was performed for one sample from low, borderline, and high decomposition sets for each product. Scores were calculated for each of these samples, whether they were randomly selected for the test set or not.
Reproducibility of the technique was evaluated by examining the overall point score range of these triplicate analyses ( Figure 2).
All ranges for these triplicate measurements (n = 69) were less than 15 sensory score points, with 90% within 10 points and 61% within five points. Ranges were similar for low, middle, and high decomposition states, but trended somewhat lower for the middle state. Overall, it appears that reproducible results were generated by this technique as demonstrated by these data.

| Compound identification
As shown in Table 3, the putative identities of 13 compounds of interest were confirmed using reference standards as described in section 2.6.3 above. These were confirmed by comparing sample peak retention time (RT, ±0.2 min), parent ion mass accuracy (±5 ppm), and that of at least one structurally significant fragment ion (±5 ppm), to those of the standard, as described in SANCO guidelines for pesticide analysis (SANCO, 2013). The identifications of an additional 14 compounds were confirmed by comparing the parent ion mass to calculated values, and fragment ion masses to spectral databases, within the same specifications. These should still be considered putative findings, as closely related or isomeric compounds may be responsible, and in some cases, cis/trans or stereoisomer

| Biogenic amine-related compounds
Biogenic amines, with their relatively high polarity and small masses, present challenges for reverse-phase liquid chromatography. It is therefore not surprising that of the amines commonly associated with seafood decomposition (Self et al., 2011), only tyramine was observed as a compound used in models. The Nacetylated analogs of tyramine and several other compounds were detectable, however, as these are generally less polar than their base counterparts. Since no source of acetylation was added during the extraction, these likely arise from natural metabolic or fermentation processes. With more reliable chromatographic options, these compounds may also represent an interesting novel approach to characterization of biogenic amines in seafood decomposition in future work.

| DNA-related compounds
DNA nucleosides were contributors to models in both free and modified forms. The only free base used in modeling was thymine, whereas methylated forms of adenine, cytosine, and guanine were used. Each of these was confirmed with standard analysis. While the exact mechanism of generation for these compounds was not explored, DNA degradation would seem most likely. There was also a database-driven match for 7-aminomethyl-7-deazaguanine, which is a known product of purine metabolism (Kanehisa & Goto, 2000).
Another potentially related database match was adenosine, although there are other potential sources for this. While DNA degradation has previously been used in a similar way to assess poultry spoilage (Faullimel et al., 2005), this appears to be a novel application of post-degradation analytes in this way.

| Other compounds of interest
Lipids of various forms were major contributors to models and are primarily more prevalent in decomposed products. These vary extensively by product, likely due to natural differences in their initial makeup. Choline, a product of lipid metabolism (Kanehisa & Goto, 2000), was also identified.
In addition to the biogenic amine-related compounds described above, other products of amino acid metabolism were also used in models. These include creatinine, a product of the same arginine, and proline metabolic process responsible for putrescine production (Kanehisa & Goto, 2000). Other examples include 1-(beta-Dribofuranosyl)-1,4-dihydronicotinamide, 1-acetylpiperidine, and tryptophanamide.

| CON CLUS IONS
In this study, a novel technique was explored for the evaluation of seafood products for decomposition. By using sensory data as a training factor, the mass spectral data can be modeled to generate a finding which is more comparable to sensory as compared to alternative chemical analysis techniques. This may provide a valuable compliment to sensory testing in regulatory or industrial settings in the future.
As a qualitative technique, the finding of only a single false negative and positive each out of the 550 test samples F I G U R E 2 Ranges of triplicate data, showing ratio within 5, 10, and 15 sensory score points for low, middle, and high decomposition states for all products By evaluating the predictive power of compounds of interest, it was possible to identify compounds which are indicative of decomposition in these products and have not previously been explored for this purpose. These may provide an interesting avenue for future work.

ACK N OWLED G M ENTS
The authors wish to thank the members of the international sam-

CO N FLI C T O F I NTE R E S T
The authors declare no conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
Due to the very large quantity of data, this will be made available upon request.