Identification of milk from different animal and plant sources by desorption electrospray ionisation high-resolution mass spectrometry (DESI-MS)

The study addresses the need for reliable, rapid methods to identify milk species and detect adulteration across animal- and plant-derived milks. Rising demand for alternative milks (due to allergy, intolerance, lifestyle, or supply issues) coupled with economic pressures increases fraud risks, including adulteration of higher-value products with cow’s milk. Existing regulatory approaches (e.g., EU IEF of β/γ-caseins) and DNA-based PCR methods have limitations such as complex, time-consuming preparation, potential false positives, and indirect detection. Ambient mass spectrometry, particularly DESI-MS, offers direct, soft ionisation with minimal sample prep and rapid analysis. The purpose here is to develop a DESI-MS method to classify milk by species, identify lipid biomarkers, and detect low-level cow milk adulteration in non-cow milks rapidly and accurately.

The paper reviews milk fraud detection methods: the EU reference IEF casein method (complex prep, potential false positives in some species), and DNA-based PCR techniques (effective but require DNA isolation, primer design, are time-consuming, and risk contamination). High-resolution MS approaches for speciation and plant protein detection have shown promise, including foodomics-based untargeted profiling. MALDI-TOF MS can differentiate milk species but needs matrix preparation and pre-treatment. Ambient MS sources (DART, DESI) enable rapid, direct analysis; DART-HRMS has authenticated milk with moderate adulteration thresholds but often still uses extraction and consumables. DESI-MS is highlighted as soft, minimally destructive, suitable for small molecules and some proteins/peptides, and well-suited for rapid, environmentally friendly screening without chromatographic separation.

Study design: Five milk types were analysed: cow, goat, camel, oat, and soya. To capture within-group variability and validate models, samples were sourced across multiple farms, production systems, and suppliers: 103 cow milk samples from 30 farms/systems (UHT and pasteurised), 27 goat milks (5 semi-skim, 22 whole) from two farms, 36 camel milks from two distributors (five production dates), 34 oat milks from four suppliers, and 73 soya milks from five suppliers (various formulations). Each sample underwent 6–9 parallel repeat acquisitions, generating 2225 total spectra (cow 940, camel 516, goat 96, oat 230, soya 443). Sample preparation optimisation: Three treatments were compared: direct spotting, methanol dilution, and water dilution. Direct analysis was impractical due to film thickness and evaporation issues. Methanol caused loss of lipid features (precipitation and clean-up) and altered ionisation behaviour, removing many potential biomarkers. Water dilution preserved lipid groups and improved total ion current. Optimal dilution was milk:water = 1:4 (v:v). A 2 µL aliquot (0.4 µL milk + 1.6 µL water) was spotted on glass slides (3 mm diameter spot) and air-dried (~10 min). Instrumentation and DESI conditions: Analyses used a Waters G2-XS Q-ToF with a Prosolia 2D Omni-Spray DESI source. Spray solvent: 98% acetonitrile–water with 0.2% formic acid, flow 2 µL/min; N2 nebulising gas at 0.7 MPa; spray voltage 4.0 kV; spray angle 65°. Source temperature 50°C; cone voltage 50 V. Mass calibration used 0.5 mM sodium formate (90% IPA) at 5 µL/min; resolution ~15,000 FWHM at m/z 600. Acquisition in positive ion mode over m/z 100–2000; scan time 0.5 s; acquisition time ~10 s per sample. Throughput: ~25 min for 96 spots; prep time comparable to dilute-and-shoot LC-MS and faster than conventional preps. Chemometrics and validation: Spectra were averaged per spot and processed using Waters AMX (v1.0.1563.0) for PCA and LDA models across m/z 100–2000, intensity limit 1.0E6 counts, mass bin width 0.2 Da. Cross-validation used 20% out bootstrapping (80% train). AMX Recognition (v0.9.2092.0) supported rapid recognition; four partitions (80%) for training; outlier threshold at 5 standard deviations. Additional modelling used SIMCA 14.1 for OPLS-DA with mean centering and Pareto scaling; model validation via permutations, reporting R2Y and Q2. Biomarker identification used LipidMaps matching of accurate mass and MS/MS fragmentation for structural confirmation. Differential analysis employed coefficients, VIP, and S-plots to select robust markers. Adulteration simulation: Cow milk was mixed with: goat (0–100% v/v in steps of 5–100%), camel (0–100% including 0.5–50%), oat (0–100% including 0.5–100%), and soya (0–100% including 0.1–100%). Binary LDA models were built to classify adulteration levels.

- DESI-MS spectral features distinct across species: cow and goat spectra dominated by fatty acids, glycerophospholipids (GP), and sphingolipids (SP); camel, oat, and soya showed more glycerolipids (GL), particularly triacylglycerols. - Unsupervised PCA separated all five classes with limited proximity between cow and goat; supervised LDA achieved clearer separations. PCA model robustness: R2X=0.898, Q2=0.867; OPLS-DA for cow vs others: R2Y=0.965, Q2Y=0.964. - Classification accuracy: Overall LDA achieved 100% cross-validation classification across milk sources for species-level identification (as reported). Across specific adulteration models, correct classification rates were: camel–cow 93.9%, soya–cow 92.2%, oat–cow 84.3%, goat–cow whole 82.35%, goat–cow semi-skim 60.7%. - Adulteration detection limits (cow milk in other milks): soya 0.1% v/v; camel 0.5% v/v (note that 0.5% was challenging in one model description); oat 0.5% v/v; goat 5% v/v. Linear trends between LDA scores and adulteration levels were observed. - Feature space: ~9501 components detected across five species at 0.2 Da mass binning. - Biomarkers: 28 robust lipid markers identified (23 cow; 5 other species), confirmed by MS/MS and LipidMaps. Cow-milk biomarkers predominantly GP and SP classes (e.g., phosphatidic acids PA(24:0), PA(28:0), PA(30:0), PA(34:0), PA(36:1); ceramides; and sphingolipids M(IP)2C series at m/z ~1202.7–1342.7). Non-cow markers included GL/TG species such as TG(55:6), TG(56:8), TG(57:8), TG(54:5), and PC(42:1) in oat. - Operational benefits: Minimal sample prep (water dilution), no organic extraction, rapid analysis (~10 s per spot), and environmentally friendly solvent use; repeatable analysis due to soft ionisation.

The findings demonstrate that DESI-MS lipidomic profiling, coupled with multivariate analysis, reliably distinguishes milk from different animal and plant sources and sensitively detects cow milk adulteration at low levels. The identification of species-specific lipid biomarkers (rich GP and SP in cow milk vs. predominant TGs in plant/camel milks) provides mechanistic discriminants for classification and trace detection. Robust PCA/OPLS-DA metrics and cross-validated LDA performance indicate strong model generalizability across diverse suppliers, farms, and processing types. Practically, the workflow offers rapid, non-destructive screening without complex extraction or chromatography, aligning with needs for on-site or high-throughput authenticity testing. Although cow–goat discrimination is more challenging, binary models achieve effective separation, addressing a critical risk scenario. The approach supports both fraud detection and allergen risk mitigation in shared processing environments and could be deployed for routine QC with potential migration to simpler MS platforms.

This work presents the first demonstration of DESI-MS for rapid lipidomic profiling to identify milk species and detect adulteration across animal and plant milks. Approximately 9500 components were observed, and 28 robust lipid biomarkers were confirmed, primarily GP/SP for cow and TGs for others. DESI-MS with chemometrics achieved high classification accuracy with detection limits from 0.1% to ~5% cow milk in alternative milks. The method is simple, fast, and environmentally friendly, enabling effective screening for mislabeling and cross-contamination. Future research directions include quantitative lipid assays using triple quadrupole MS, expansion to more species and processed products, validation on lower-cost or portable MS platforms (e.g., single-quadrupole), and broader inter-laboratory validation to standardize workflows for regulatory adoption.

- Discrimination between cow and goat milk is less distinct than other pairings; semi-skim goat samples showed lower classification performance (60.7%). - For some models, detection at the lowest adulteration level (e.g., 0.5% in camel–cow) was reported as challenging/not feasible in specific analyses despite low stated limits elsewhere, indicating potential variability near detection limits. - Lower correct classification rate for oat models (84.3%) compared to others suggests matrix-dependent performance differences. - Models were built on samples from specific suppliers/farms and processing types; broader geographic and processing diversity may affect generalizability. - The approach relies on high-resolution MS for biomarker discovery; while translatable, performance on simpler instruments requires further validation. - The study is primarily qualitative/semi-quantitative; dedicated quantitative validation (e.g., with triple quadrupole) is suggested for routine threshold enforcement.