Real-Time Pork Breed and Boar Taint Classification Using REIMS

To address consumer demand for authentic, high-quality pork and combat food fraud, reliable traceability systems are essential. Current methods for identifying pig breeds and detecting boar taint—an off-odor affecting meat quality—are limited by high costs, slow processing times, and low sensitivity. A recent study explored rapid evaporative ionization mass spectrometry (REIMS) as a high-throughput, real-time alternative. By analyzing metabolomic fingerprints from pig neck fat, REIMS was combined with multivariate data analysis and machine learning algorithms to successfully classify pork breeds and accurately detect boar taint in both laboratory and slaughterhouse environments. LCGC International spoke to Lieselot Y. Hemeryck and Lynn Vanhaecke, two of the authors of the paper resulting from this work, and their expert colleague, Vera Plekhova, about the study and their findings.

Can you explain the ionization mechanism in rapid evaporative ionization mass spectrometry (REIMS) and how it differs from electrospray ionization (ESI) or matrix-assisted laser desorption–ionization (MALDI) in terms of sample preparation, ion types generated, and matrix effects?
REIMS is based on rapid thermal ablation leading to desorption and evaporation of biological matrices, initiated via RF diathermy. In this process, resistive heating of the sample causes rapid vaporization and the creation of a metabolite-containing aerosol. This aerosol is then transferred directly into the REIMS source where droplet disruption, declustering and additional ion formation occur through interaction with a heated surface. Ionization using REIMS is thus a type of thermal ablation, combined with thermal desorption. REIMS-based ionization is thus relatively harsh compared to the softer ESI and MALDI. Depending on the selected ionization mode of the mass spectrometer, REIMS can generate and detect either positively or negatively charged ions. In lipid-rich tissues analyzed in negative ion mode, spectra are typically dominated by negatively charged lipid-derived ions (e.g. deprotonated lipids, fatty acids, etc.).

In contrast to ESI and MALDI, REIMS does not require sample extraction or matrix application. As such, REIMS is suited for direct in-situ analysis, whereas ESI and MALDI are not. ESI, MALDI as well as REIMS are all subject to matrix effects. For ESI, this is partially mitigated by means of sample extraction and chromatographic separation. In MALDI variability arises primarily from co-crystallization processes and ion competition within matrix-analyte crystals. For REIMS specifically, matrix effects are inherent and unavoidable, implying that matrix effects become embedded within the characteristic metabolic fingerprint used for classification. As such, rather than being eliminated, matrix effects are controlled through standardization and robust modeling.

The study uses a Q-ToF mass analyzer for real-time metabolomic fingerprinting. Why is a Q-ToF particularly well suited for REIMS-based applications, and what trade-offs would arise if an Orbitrap or triple quadrupole were used instead?
A Q-ToF is particularly well suited for REIMS because it provides fast full‑scan acquisition and wide within-scan dynamic range in combination with good mass accuracy, which is indeed required for reproducible untargeted metabolomic/lipidomic fingerprinting in real time. An Orbitrap provides superior mass resolution and accuracy (useful for metabolite identification) compared to a Q-ToF. However, its scans more slowly, is prone to space-charge effects in dense ion populations and is more susceptible to contamination. Consequently, it is less suited for the dirty REIMS-generated aerosols and the envisioned high-throughput analysis. Triple quadrupoles are characterized by their robust design and high sensitivity, making them well-suited for targeted analysis. These instruments are engineered to provide low resolution and low scanning speed over extended mass ranges. Triple quads depend on MRM transitions of predefined analytes with established fragmentation patterns. However, for untargeted fingerprinting, high resolution is essential, and triple quads are not suitable for this application.

REIMS generates complex lipid-rich spectra from adipose tissue. What challenges does this pose for mass accuracy, ion suppression, and spectral reproducibility, and how can these issues be mitigated during acquisition or data preprocessing?
The analysis of adipose tissue by means of REIMS generates a very high ion flux, dominated by a few highly abundant lipid classes. These lipids can contribute to peak shifts and reduced mass accuracy through detector saturation. The presence of abundant lipid ions can potentially lead to ion suppression, thereby influencing the detectability of molecules with lower abundance. In adipose tissue specifically, changes in tissue composition may occur depending on factors such as the sampling site, resulting in variations and reduced reproducibility. During the acquisition process, measures can be implemented to mitigate this issue. These measures include the use of optimized, matrix-specific instrument settings to reduce overloading of the MS, as well as standardized sampling procedures to reduce spectral variability due to tissue-related differences. In addition, it is recommended to implement lock-mass correction, which involves the use of a stable reference ion (e.g., leucine enkephalin) to ensure continuous correction of mass drift. During the data preprocessing stage, it is essential to implement normalization and scaling. Additionally, the implementation of data binning may assist in mitigating the downstream consequences of technical inconsistencies while maintaining the integrity of biologically significant information.

Neck fat was selected as the sampling matrix for both breed classification and boar taint detection. From an MS perspective, what advantages does adipose tissue offer for metabolomic fingerprinting, and what limitations does it impose?
From an MS perspective, adipose tissue offers several advantages for metabolomic fingerprinting, primarily due to its high and relatively stable lipid content, which produces consistent and information-rich spectra. The low water content of fat contributes to stable ionization and reduced variability; however, this comes with a trade-off, as previously mentioned. Due to the significant ion suppression caused by the predominant lipid species, the direct detection of more polar or lower-abundance metabolites is hindered.

Nevertheless, depending on the research question, lipid-derived fingerprints are highly informative and discriminative for characteristics like genotype (e.g., breed), the production system, and overall physiology. In our study, we found that neck fat is particularly well-suited for detecting and classifying boar taint (an off odor that may be present in meat from uncastrated male pigs) in adipose tissue. Additionally, the neck region is known to contain a high amount of fat, which makes it accessible during the slaughter process.

Boar taint compounds such as androstenone and skatole accumulate in fat but are present at low concentrations. How does REIMS detect discriminatory information despite limited sensitivity for specific low-abundance analytes, and what does this imply about targeted vs. untargeted MS workflows?
Classification of boar taint using REIMS is not based on the detection of a limited set of low concentration biomarkers, yet on the overall metabolic fingerprint and underlying metabolic state. The abundant lipid ions generate strong, stable signals that can be used to discriminate samples by means of multivariate statistics and/or machine-learning-based models, even when the concentrations of the boar taint compounds are low. This illustrates a key distinction between targeted and untargeted MS: targeted methods can be used to detect and quantify specific molecules of interest with high sensitivity, while untargeted methods classify metabolic phenotypes based on distinctly different metabolic patterns. As a result, untargeted MS excels in capturing complex biological states and can be used to perform classification in real‑time (and with high‑throughput), even without the dominance of causal biomarkers. Targeted MS remains useful still, albeit for different purposes, like e.g. more fundamental research.

The study relies heavily on multivariate and machine-learning models rather than identification of individual biomarkers. How does this shift the role of mass spectrometry from a compound-identification tool to a pattern-recognition platform, and what are the implications for method validation?
When an analytical workflow, like ours for example, is using multivariate and machine‑learning models, mass spectrometry may indeed shift from being a compound‑identification tool to a pattern‑recognition platform. Instead of detecting specific biomarkers, the MS system generates molecular fingerprints that describe the overall biochemical composition of a sample. This delivers a different type of output: a fingerprint and model classification instead of a report on the detection (yes/no) and concentration of the compounds of interest. As a result, method validation should not mainly focus on accuracy, precision or limits of detection for individual analytes, but rather on the model’s classification performance and robustness. This shift necessitates alternative validation strategies, including extensive cross-validation, the use of independent test sets, and monitoring of model stability and drift.

High-throughput slaughter-line analysis requires acquisition times on the order of seconds. What compromises are typically made in mass resolution, scan speed, or spectral range to achieve real-time MS performance, and how might these affect classification robustness?
High-throughput slaughter-line analysis indeed requires acquisition times in the order of seconds, which involves several compromises regarding instrument performance. Given the limited time available for sampling per carcass, the MS scan rate is set high to acquire sufficient data points within that timeframe. In combination with the ToF mass spectrometer, where resolution is largely independent of scan speed (unlike, e.g., Orbitrap), this still allows for HRMS acquisition. However, the limited acquisition window restricts the number of scans that can be averaged for each sample. This restricts ion statistics, lowers signal‑to‑noise ratios and increases spectral variability. In addition, the m/z range is often narrowed, thus ‘sacrificing’ a broader spectral range. As a result, classification models must rely on the overall metabolome fingerprint in a narrowed m/z range. Although these compromises indeed reduce the amount of information achieved on the level of individual compounds, robust multivariate models that were trained and cross-validated using real samples can nevertheless achieve good to excellent classification accuracy.

Chemometric methods such as PCA and OPLS-DA were compared with ML approaches like SVM and Random Forests. From an MS data perspective, what characteristics of REIMS spectra influence the choice between classical multivariate analysis and machine-learning classifiers?

Because REIMS produces spectral fingerprints that are dominated by lipids and lipid‑derived compounds, it may be expected that variables are highly correlated, which makes dimensionality‑reduction approaches like PCA or OPLS‑DA interesting. Also, classical chemometric methods such as PCA and OPLS-DA are well suited for exploratory analysis and visualization of trends. In contrast, the application of PCA and OPLS-DA for modelling non-linear relationships is somewhat limited. Machine-learning methods are more adept at modeling complex, non-linear patterns. Consequently, they are expected to handle noise and redundant features more efficiently, which is pertinent to REIMS data. The final model of choice is contingent upon the underlying biochemical processes that are being modeled. These processes are difficult to predict in advance. Therefore, it is necessary to test different data modeling approaches to evaluate which is the optimal fit.

Ambient MS techniques are often criticized for limited inter-laboratory reproducibility. What sources of variability are specific to REIMS, and how can standardization be approached in a regulatory or industrial setting?
REIMS is a process that is influenced by various factors during the sampling and ionization stages. These factors include the settings of the probe, the speed and depth of sampling and ionization, indoor climate variables (e.g., humidity and temperature), and the individual operator. It is important to note that sample-related factors may also play a role and may vary between labs. These factors may include, for example, the temperature of the sample and the specific sampling spot. It should be noted that differences in instrumental settings may further contribute to the variability, as different settings will result in a different output. In industrial or regulatory settings, it is advisable to implement harmonized Standard Operating Procedures (SOPs) for instrument calibration, cleaning, data preprocessing, and model training. These SOPs should emphasize standardization and control of sampling and analysis. Additionally, the use of Quality Control materials to monitor instrument performance is recommended. Centralized model training is an effective strategy for standardizing classification models across sites. To ensure long-term robustness and reproducibility, periodic retraining and performance monitoring of models is essential.

Compared with DNA-based traceability methods, MS-based metabolomic fingerprinting offers speed but indirect biological specificity. In your view, what role should MS-based methods like REIMS play alongside genomic approaches in future food authentication and traceability frameworks?
Each technology and methodology has its advantages and disadvantages. We could benefit from using them in a complementary way rather than as competing tools. Genomic methods can provide unambiguous species and breed identification, but they are slow, expensive, and lab-bound. Consequently, they are not appropriate for use at line. REIMS and other ambient fingerprinting techniques deliver rapid, on-site metabolomic fingerprinting. This can help identify potential anomalies or fraud. In a modern traceability framework, MS-based methods have the potential to serve as frontline screening tools. These tools would enable continuous monitoring, early detection, and rapid triage of samples. Should further characterization or confirmation be required, additional complimentary analyses can be performed. These analyses may also include genomic approaches.

References

1. Gkarane, V.; De Graeve, M.; Stephens, C. et al. Towards Real-Time Pork Breed and Boar Taint Classification Using Rapid Evaporative Ionisation Mass Spectrometry. npj Sci Food 2026. DOI: 10.1038/s41538-025-00685-4