Toward the Prediction of the PSE-Like Muscle Defect in Cooked Hams

Animal, muscle, and blood samples

A batch of 120 pigs (from a sow Large White × Landrace and Pietrain sire), from different farms, were slaughtered at the commercial abattoir “la Guerche de Bretagne, France.” Carcasses were blast-chilled for 1 h and 15 min, and blood was collected from each animal and treated as described in the following section. At 24 h post mortem, after pH and temperature stabilization, hams were sorted according to their level of destructuration on a scale from 1 to 4 (Vautier et al., 2008), their pH, temperature, and exudate values. From the 120 animals, 20 were selected, based on the destructuration score—10 “normal hams” (NH) and 10 “PSE-like hams” (DH). The inner part (DHi) and outer parts (DHo) of semimembranosus muscles were sampled in each of the DH, and the inner part of semimembranosus muscles was sampled in each of the NH.

MALDI-TOF mass spectrometry imaging of ham samples

At 24 h post mortem, semimembranosus muscle samples (1 × 1 × 1 cm) were collected, positioned on a cork plate, and cryofixed by immersion at −160°C in isopentane cooled with liquid nitrogen (−196°C). Serial cross-sections (10 μm thick) were cut using a cryostat (Microm HM 560; Thermo Scientific) and were collected on glass slides for histological stains and on indium tin oxide glass slides (Bruker Daltonik, Bremen, Germany) for MALDI-mass spectrometry imaging (MSI). The glass slides were stored under vacuum until use. The muscle sections were subjected to 2 washing steps at 70% and 95% ethanol to deplete lipids and then dried in a desiccator for 30 min. The sinapinic acid matrix was 10 mg/mL in water/acetonitrile at 60:40 (v/v) with 0.2% trifluoroacetic acid, applied using ImagePrep (Bruker Daltonics) according to the manufacturer’s recommendations. The spectral data were acquired on an Autoflex Speed MALDI-TOF/TOF mass spectrometer equipped with a Smartbeam laser, using FlexControl 3.4 and FlexImaging 3.0 software packages (Bruker Daltonics). Ions were detected in positive linear mode over a mass range of m/z 2,000–20,000 with a sampling rate of 0.63 GS/s. The lateral spatial resolution was set to 75 μm, and a total of 500 laser shots were accumulated per pixel at constant laser power, using random movement within each pixel. The deflection was set at an m/z of 1,500 and laser focus at medium. Analyses were performed using a detector gain of 2.69 V, ion source voltage 1 at 19.5 kV, ion source voltage 2 at 18.15 kV, and lens voltage at 7 kV. Spectral data were loaded into SCiLS Lab 2016 software (http://scils.de/; Bremen, Germany). The following workflow was used for data treatment: baseline subtraction using the TopHat algorithm, normalization by the total ion current algorithm, and peak picking using the orthogonal matching pursuit algorithm, with peak alignment and spatial denoising.

Plasma proteins spectral fingerprints by MALDI-TOF mass spectrometry and FTIR spectroscopy

Immediately after slaughter, a whole blood sample was collected from each pig in a heparinized tube (Venosafe, Terumo, Shibuya, Tokyo, Japan). After 5 stirrings, the tubes were centrifuged at 1500g for 10 min. The plasma was recovered and frozen immediately in liquid nitrogen. The samples were stored at −80°C until use. For the MALDI-TOF mass spectrometry protein fingerprint, plasma proteins were pre-purified using Spin Tubes (C18 Agilent Peptide Cleanup, Agilent Technologies, Wilmington, DE) according to the manufacturer’s instructions. Then, 1 μl of pre-purified plasma proteins was manually spotted in triplicate on a polished steel target (MTP 384 Target Plate Polished Steel, Bruker Daltonics GmbH, Bremen, Germany) at a ratio of 1:1 with the matrix. The matrix used was α-cyano-4-hydroxycinnamic acid matrix at 7 mg/mL in water/acetonitrile 50:50 (v/v) with 0.2% trifluoroacetic acid. The mass spectrometer used to acquire the protein fingerprinting was an Autoflex Speed MALDI-TOF/TOF with a Smartbeam laser, using FlexControl (version 3.4) software (Bruker Daltonics GmbH, Bremen, Germany). A total of 4,000 spectra were accumulated randomly per sample. The laser power was constant for all the samples, and the laser focus was set at medium. Ion detection was done in linear mode at a mass range of m/z 1,000–10,000, with a sampling rate of 0.31 GS/s. Acquisition deflection was set at m/z 1,000. The detector gain was set at 2,500 V, the ion source voltage 1 at 19.56 kV, the ion source voltage 2 at 18.11 kV, and the lens voltage at 7 kV. External calibration of spectra was done through the deposition of a protein standard (Protein Calibration Standard I, Bruker Daltonics) before each measurement on the same target. MALDI-TOF spectra were processed using FlexAnalysis software (version 3.4) (Bruker Daltonics GmbH, Bremen, Germany). Baseline subtraction using the TopHat algorithm was set with a 5% minimal baseline width. Smoothing was performed with a 1 Da width and 5 cycles. Individual spectra were normalized using the Total Ion Count (TIC) value. Peak picking was applied using the peak intensities with a signal to noise ratio of 2. Infrared spectra were obtained using a Bruker Tensor II spectrometer, and an ATR accessory with 250 μm × 250 μm diamond crystal (Specac Ltd., Orpington, UK) operated by the OPUS 7.5 software (Bruker, Bremen, Germany). The spectral resolution was set at 4 cm⁻¹, and 100 scans were used for each measurement in the range of 4,000–600 cm⁻¹. All the sample acquisitions were performed 5 times as follows: 1 μL of plasma was spotted on the diamond crystal and left to dry at room temperature for 8 min. Between samples, the ATR crystal was cleaned with deionized water and ethanol, and for each sample measurement, a background spectrum was recorded for correction. The 3 best spectra out of 5 were analyzed. After atmospheric compensation and baseline correction, the spectra were pre-processed by the second derivative operation and cropped to obtain 2 areas of interest: 1,800–900 cm⁻¹ and 3,450–2,700 cm⁻¹. Each area was then normalized using vector normalization, and the mean spectra of the 3 replicates were used.

Chemometrics

Neuronal network analysis was performed using Orange software (Demsar et al., 2013). The number of neurons per hidden layer was set at 100 with an activation function by logistics. The solver parameter for weight optimization was “Adam,” a stochastic gradient-based optimizer. The alpha parameter, the regularization term, was set at 0.0001, and the maximum number of iterations was set at 200. Internal validation was performed using a stratified 5-fold cross-validation. The classification results using the predictive model are given for each sample of PSE-like muscle defect (DH, n = 6) and normal (NH, n = 7) hams. The classification assignment is given (NH/DH) and shown in gray in the case of misclassification. The score probability of classification assignment is given for each class, i.e., DH and NH.

Spatial variability of the DH defect

The spatial variability of the PSE-like defect that occurs in cooked hams was investigated by MSI by MALDI-TOF. To determine whether the defect progressed internally within the ham, indicating a local heterogeneity, the relationship between NH molecular maps with both DHi and DHo was studied. The training set used to build the predictive model was composed of the spectral data of the DHi and DHo. Then, the data from the NH were classified using this model to determine the class to which they were closer: DHi or DHo. The model’s performance was evaluated by cross-validation accuracy, which was 87%. The model’s performance can be expressed as its capability to describe the spectral classes, in percentage, and its use as a tool to determine the class in which the new spectral data belonged. The m/z intensities on which the model’s performance rely are those showing the highest scores: m/z 4128.7, m/z 4155.6, m/z 5442.5, and m/z 8565.0. These 4 peaks are represented in Figure 1A with stars on the mean spectra. The two first peaks, m/z 4128.7 and m/z 4155.6, are significantly more intense in the DHi compared with the DHo and those of the NH. This result confirmed the hypothesis according to which the structural defect is internal, and that the DHo present more similarities to the NH. Boxplots and all the individual intensities measured on each pixel are shown in Figure 1B, as well as an example of an ionic map for each ion. It appeared very clear that these two ions are co-localized and that their intensities are correlated. The mass shift of 27 Da could be due to oxidation of arginine (to glutamic acid) (https://abrf.org/delta-mass), and the oxidation product of arginine may be glutamic semialdehyde, a carbonyl that is a very well described modification in meat products (Estévez, 2011). Indeed, the carbonylation of proteins from chicken breast meat has been studied and linked to PSE meat (Estévez, 2015; Carvalho et al., 2017). Despite the absence of direct protein identification using MSI by MALDI-TOF, altogether these results could suggest that m/z 4128.7 and m/z 4155.6 are the same protein or peptide from the same protein, with the presence of an amino acid modification. The intensities plot of m/z 5442.5 was lower in the DHi, while those of m/z 8565.0 were higher in the DHo. Furthermore, the ionic map of these ions indicated that their intensities were not homogenous on the tissue section but showed spatial specificity, which could not be seen without imaging by mass spectrometry.

Figure 1.

Predictive model of the PSE-like areas. (A) Mean spectra acquired on the muscle sections from the “normal hams” (in blue, n = 10), on the muscle sections from the inner (in red, n = 10) and outer (in yellow, n = 10) parts of “PSE-like hams.” The peaks with the highest loading scores and biggest contribution to the predictive model are represented by stars. (B) Box plot and graphical representation of all the individual peak intensities observed on the muscle sections, and ionic map for m/z 4128.7, m/z 4155.6, m/z 5442.5, and m/z 8565.0 for representative tissue section samples from “DHo,” “DHi,” and “NH.” (C) NH spectral data classification using the predictive model: the spectra from NH classified as DHi are shown in red, and the spectra from NH classified as DHo are shown in yellow. Serial cross sections were stained with Sirius Red to reveal the connective network and with hematoxylin and eosin to reveal the tissue structure. DHi = inner parts of PSE-like hams; DHo = outer parts of PSE-like hams; NH = normal hams; PSE = Pale Soft Exudative.

Then, the predictive model, mainly based on these ion intensities, was used to classify the spectral data from NH. Thanks to this approach, the spectral data from NH could be classified as being closer to the DHi or the DHo. Spectral data using this predictive model were classified for every spectrum of the sections from NH and were represented as a color map: the spectrum belonging to the DHi in red and the spectra belonging to the DHo in yellow (Figure 1C).

Plasma protein fingerprint to predict the PSE-like muscle defect

The second part of the study aimed to assess the capacity of protein fingerprinting by MALDI-TOF mass spectrometry and ATR-FTIR spectroscopy to predict the PSE-like muscle defect in hams. Plasma samples were chosen since it is essential not to depreciate the carcass and to develop an easy-to-use method, blood sampling being easier than muscle biopsies. In the first part of this study, the prediction capacity of both spectral approaches was assessed separately using multivariate analysis. Then, both data matrixes were analyzed using the most discriminant spectra features to determine the synergy of both of these spectral methods. The predictive model built using neural network analysis from MALDI-TOF spectra showed a classification accuracy of 84.6% and a precision and recall of 88% and 84.6%, respectively, resulting in the correct classification of 100% of NH samples, whereas 2 DH samples were misclassified (Tables 1–2). This misclassification has to be moderated by the number of observations and by their corresponding score probability. Indeed, in both cases, the score probabilities were 56.4% and 65.5% for the DH class and 43.6% and 34.5% for the NH class. The neural network built with FTIR spectra showed a classification accuracy of 84.6% and a precision and recall of 88.5% and 84.6%, respectively. This resulted in the correct classification of 100% of DH samples, whereas 2 NH samples were identified as a false negative, classified as DH (Table 1), with a score probability of 66.1% and 96.9%.

To demonstrate the synergy of both spectral fingerprints in predicting the PSE-like muscle defect, both approaches were combined. The respective top 30 features found to be the most relevant in the neural networks obtained with MALDI-TOF and FTIR spectral fingerprints were analyzed to build a new predictive model. The results showed a classification accuracy of 100%, with a precision and recall of 100% and 100%. Thus, the classification results confirmed the relevance of such a combinatory approach since it improved the correct classification assignment up to 100%. All the individual spectra were correctly classified, meaning that the neural network successfully predicted the PSE-like muscle defect in all cases.