Novel time-domain NMR-based traits for rapid, label-free Olive oils profiling

The study addresses the problem of authenticating and grading olive oils (OO) in the face of frequent adulteration and complex variability in physicochemical properties such as fatty acid (FA) profiles and free fatty acid (FFA, acidity) content. Existing classifications (EVOO, VOO, refined) depend on these properties and processing history. High demand and profit margins make EVOO a common target for adulteration, necessitating rapid, simple, and non-destructive analytical tools. Prior laboratory methods (chromatography, spectroscopy, DNA analysis, high-field NMR) are accurate but often slow, expensive, and require specialized infrastructure and expertise. Building on recent work using two-dimensional time-domain NMR (TD-NMR) to classify edible oils by saturation level, this study proposes a low-field, benchtop, point-of-care TD-NMR approach that pairs T1 and T2 relaxation times as phenotypic “traits.” The hypothesis is that subtle differences in molecular microenvironment (from triacylglycerols, FFA, and trace compounds) produce distinct relaxation signatures that can be leveraged with machine learning to rapidly and accurately classify OO type and infer region of origin in a label-free, non-destructive manner.

The paper reviews established methods for OO authentication and adulteration detection: chromatographic analyses (e.g., GC/HPLC of sterols, fatty acid markers), various spectroscopies (UV-Vis, NIR, Raman, FT-Raman, fluorescence), DNA-based approaches (PCR/HRM barcoding), and high-field NMR. While effective, these approaches often demand costly equipment, extensive sample prep, laboratory settings, and specialized personnel. High-field NMR has demonstrated strong performance for authenticity and quality control but entails cryogenics and complex workflows. Prior work by the authors and others showed low-field TD-NMR with T1–T2 relaxometry can classify edible oils by FA saturation more accurately than conventional methods, suggesting potential for point-of-care use with minimal sample prep and rapid turnaround.

Samples: Thirty-six commercially available olive oils were analyzed for type classification: 21 extra-virgin (EVOO), 8 virgin (VOO), and 7 refined (refined OO). For region-of-origin analyses, EVOOs labeled from four European regions were used: 3 Greece, 4 Italy, 9 Portugal, 5 Spain. Oils were purchased in Braga, Portugal or online; brands are listed in Supplementary Table 2. No preprocessing was performed prior to measurements.

Time-domain NMR: 1H NMR measurements were performed at 21.7579 MHz using a portable permanent magnet (B = 0.5 T, Metrolab Instruments) and a benchtop console (Kea Magritek). Temperature was controlled at 30 °C. T1 relaxation times were acquired via inversion recovery (IR) and T2 via Carr-Purcell-Meiboom-Gill (CPMG) sequences. Parameters: echo time = 200 µs; number of echoes = 10,000; signal averaging = 32; recycle delay = 2 s. Measurements were performed in a single-blinded manner. For type classification, 10 repeats per oil yielded 360 total points; for origin assessment, 210 points were collected. Data were represented on a two-dimensional (T2, T1) “magnetic state” map.

Acid value (AV): Determined per EN ISO 660:2009. Ten milliliters of oil were diluted in ethanol (99%) with phenolphthalein; titrated with 0.1 mol/L KOH under stirring to persistent color change. Each sample measured twice. AV was calculated as W_AV = 56.1 × cV / m. Free fatty acids estimated as W_EFA = V_c M / (10 × m) ≈ 0.5 × W_AV, with M = 282.47 g/mol (oleic acid).

UV-Vis and NIR spectroscopy: UV-Vis spectra (SHIMADZU UV-2550) recorded from 200–800 nm at 1 nm resolution. NIR spectra (PerkinElmer LAMBDA 950) recorded from 500–2200 nm in 5 nm steps. Measurements used matched 1 cm path length quartz/optical glass cells with an empty cell reference. NIR spectra were cleaned using a despiking algorithm (cut-off = 6, threshold = 10). Each sample was measured three times and averaged.

Adulteration limit-of-detection (LOD): Sunflower oil was mixed into a selected EVOO at concentrations from 0% to 100% sunflower oil. T1–T2 relaxometry was measured in duplicate across five samplings per mixture to assess linearity and LOD.

Machine learning and statistics: Data processing used Orange 3.1.2 and/or R. Unsupervised hierarchical clustering (Euclidean distance on averaged T1, T2, A-ratio) generated dendrograms and heatmaps. Supervised models included k-nearest neighbors, logistic regression, Naive Bayes, neural networks, and random forests. Models were evaluated with leave-one-out cross-validation. Receiver Operating Characteristic (ROC) curves and AUC values assessed performance. Student’s unpaired t-tests (one- or two-tailed as noted) were used for statistical comparisons; significance generally set at P < 0.005 in figures. A power function fit y = a x^b with Levenberg–Marquardt iterations (chi-squared tolerance 1e-9) was used in ROC analyses to compare fitted AUC to averaged model AUCs.

• Rapid, label-free classification: TD-NMR relaxometry (T1, T2, A-ratio) enabled non-destructive classification of OO types (EVOO, VOO, refined) in under 5 minutes using microliter samples.
• Strong clustering and separability: Significant clustering of oil types was observed in T1–T2 space (P < 0.005), indicating inter-type variation exceeded intra-sample variance.
• Superior diagnostic performance vs spectroscopies: For oil-type classification, average ROC AUCs across supervised models were: NMR 0.949 (CA 0.872), NIRS 0.838 (CA 0.800), UV-Vis 0.731 (CA 0.633). Best individual models reached AUC up to 0.984 (logistic regression, NMR) and 1.000 (kNN, RF on NIRS), but overall NMR averaged highest.
• Physicochemical correlates: Mean values (Table D) by type: EVOO AV 0.52%; VOO AV 0.71%; refined AV 0.40%. EVOO T1 ≈ 168.0 ms, T2 ≈ 150.5 ms, A-ratio ≈ 1.12; VOO T1 ≈ 174.0 ms, T2 ≈ 153.6 ms, A-ratio ≈ 1.14; refined T1 ≈ 158.8 ms, T2 ≈ 146.3 ms, A-ratio ≈ 1.11. Fatty acid profiles (SAFA/MUFA/PUFA) also differed by class (e.g., EVOO 13.8/71.3/7.3%).
• Origin inference: Using EVOOs labeled from Greece, Italy, Portugal, and Spain, mean T1 (ms) were 166.3, 168.9, 168.9, 166.7; mean T2 (ms) were 147.7, 150.2, 151.0, 150.1 for Greece, Italy, Portugal, Spain respectively. Average ROC AUC for regional identification: NMR 0.706 (CA 0.496), NIRS 0.696 (CA 0.462), UV-Vis 0.694 (CA 0.513). Pairwise AUCs reflected geography: Greece–Portugal 0.89; Greece–Spain 0.84; Greece–Italy 0.74; Spain–Portugal 0.69; Italy–Spain 0.73; Italy–Portugal 0.60.
• Adulteration LOD: Mixing sunflower oil into EVOO yielded a strong linear relationship in T1–T2 traits vs sunflower fraction (r^2 = 0.93). Pure sunflower vs EVOO showed (T2, T1) ≈ (188.3, 202.9) vs (155.3, 174.6) ms. The NMR-based LOD for adulteration was ~1%, comparable to NIRS (~1%) and better than UV-Vis (~5%).
• Practicality: The integrated, benchtop TD-NMR approach is user-friendly, fast, point-of-care deployable, requires no solvents or complex sample prep, and offers lower per-assay cost than spectroscopic comparators (Table 3).

Findings support the hypothesis that TD-NMR relaxation traits encode the composite molecular environment of olive oils, enabling accurate, rapid, label-free classification. Higher overall unsaturation (e.g., increased PUFA/MUFA ratio, reduced SAFA) and elevated FFA (acidity) were associated with longer T1 and T2, consistent with a mechanism where disrupted packing and weakened van der Waals interactions increase molecular mobility and relaxation pathways. Compared with chromatographic and vibrational spectroscopies, the TD-NMR method achieved higher average AUC for oil-type classification, with minimal sample handling and faster turnaround. For geographic origin, average AUC ≈ 0.71 indicates moderate discriminative power; pairwise AUCs correlate with geographic proximity (e.g., higher similarity within the Iberian Peninsula), suggesting phenotypic influences from genetics, climate, and soil. The approach integrates well with machine learning, allowing near real-time processing and scalable deployment for routine quality control and anti-adulteration screening.

The study introduces a low-field, benchtop time-domain NMR relaxometry workflow that rapidly and non-destructively profiles olive oils using T1/T2-based phenotypic traits. Coupled with machine learning, it outperforms UV-Vis and NIR spectroscopy on average for oil-type classification (AUC ≈ 0.95), shows moderate success in inferring region of origin (AUC ≈ 0.71), and detects adulteration at ~1% levels. The platform is inexpensive per assay, requires no solvents or complex preparation, and is amenable to point-of-care deployment. Future work could expand sample diversity across cultivars and regions, evaluate additional adulterants and complex blends, standardize protocols across instruments and temperatures, and integrate larger multi-instrument datasets to further improve robustness and generalizability.

• Region-of-origin performance was moderate (average AUC ≈ 0.71) and based on EVOOs labeled from only four European regions; broader geographic validation is needed.
• Adulteration LOD was assessed using sunflower oil mixed into one EVOO; performance may vary with other adulterants and matrices.
• Origin ground truth relied on product labeling rather than independent verification.
• Measurements were conducted at a fixed temperature (30 °C) and specific instrument settings; relaxation times can be temperature- and hardware-dependent, potentially affecting transferability.
• Sample set size for VOOs and refined OOs was relatively small (8 and 7, respectively), which may limit generalizability of class-specific statistics.