*2.2. 1H and 13C NMR Spectra Acquisition*

The analysis was performed on a model AVANCE III 500 MHz NMR spectrometer equipped with a 5 mm 1H/D-BB probehead with z-gradient, automated tuning and matching accessory, and a BTO-2000 accessory for temperature control (Bruker BioSpin GmbH, Rheinstetten, Germany). Samples were measured at 300.0 K after a 5 min resting period for temperature equilibration. NMR spectra were acquired using Topspin 3.5 (Bruker). Automated tuning and matching, locking and shimming using the standard Bruker routines, ATMA (automatic tuning and matching in automatic mode), LOCK (frequency-field lock to offset the effect of the natural drift of the NMR's magnetic field B0) and TopShim, were used to optimise the NMR conditions. Samples were analysed using the zgpg30 pulse method for 13C NMR, while the zg30 and NOESY 1D noesypr1d NMR pulse sequence using a standard presaturation were used for 1H NMR. Every extract sample was run twice with a 1H NMR standard single pulse experiment zg30 for 100 scans. The samples were run twice automatically under the control of ICON-NMR. Each run had two prior dummy scans, resulting in 65,536 data points with a resolution of 0.305 Hz acquired with an acquisition time and a relaxation delay time of 3.27 and 4 s, respectively. The 90◦ flip angle for free induction decay was adjusted to 10 μs. In the case of one-dimensional Nuclear Overhauser Effect spectroscopy, 100 scans were acquired, each run having two dummy scans, which resulted in 32,768 data points with a resolution of 0.489 Hz, acquired with an acquisition time and relaxation time of ~2.04 and 4 s, respectively. In the case of 13C NMR, 250 scans were recorded for each sample, with an acquisition time of 21 s to allow sufficient time for complete relaxation of 13C nuclei between scans. The acquisition delay was set at 2 s. The receiver gain was set at 203, and the temperature was locked at 298.0 K by means of a BTO-2000 accessory. Broadband 1 H decoupling techniques were employed. The above parameters and settings could run samples with a turnover time of 1 h and 40 min each, excluding an initial 5-min temperature equilibration period.

Prior to Fourier transformation, the free induction decays (FIDs) were zero-filled to 64 k and a 0.3 Hz line-broadening factor was applied. The chemical shifts are expressed in d scale (ppm), referenced to the residual signal of chloroform. For 1H NMR, this was found at 7.24 ppm [21] whilst for 13C NMR, this was found as a triplet centerd around 77.01 ppm [22]. The corrected spectra were exported as ASCII files from Topspin 3.5 (TopSpin™ version 5, Bruker, Billerica, MA, USA) and imported directly into The Unscrambler X 10.3 (CAMO Software, Oslo, Norway) for all subsequent mathematical data processing. Each spectrum was automatically binned by the software into 32,768 buckets, each bucket being 0.0072223 ppm wide. The signal-to-noise ratio was calculated using the peak at 172.8 ppm for 13C NMR corresponding to C1 of the glycerol chain, which resulted in a signal-to-noise ratio of 520:1. For 1H NMR, the signal-to-noise ratio was calculated using the peak at 9.70 ppm, corresponding to the aldehyde proton in hexanal, and a signal-to-noise ratio of 1.26:1 and 1.46:1 was obtained for zg30 and NOESY pulse sequences, respectively.

The spectrum obtained was subjected to different spectroscopic signal processing techniques, which were evaluated and compared. The spectra were normalised, a transformation that put all spectra on the same scale, thus eliminating the fluctuations in intensities between spectra arising from slightly different sample concentrations. Both peak normalisation and area normalisation were carried out separately on the baseline corrected spectrum. Normalisation was followed by detrending and deresolving procedures. The detrend transformation removes the effects of nonlinear trends, showing only the absolute changes in values across spectra by removing the least-squares line of best fit from the data, thus focusing only on fluctuations between data. Deresolve is a noise-reducing transformation that operates by artificially lowering the resolution of the spectra. Other treatments applied to the baseline corrected spectrum include multiplicative and orthogonal scatter corrections (MSC and OSC), and standard normal variate (SNV). MSC was corrected for scaling effects by performing a regression of a spectrum against a reference spectrum, thereby correcting the spectrum using the slope of the

fit was obtained from the regression. OSC removes variance from the factors that is not related to the response, by finding directions in X that describe large variances while being orthogonal to Y and subtracting them from the data. The SNV transformation works similarly to MSC, however, it standardises each spectrum using data from the spectrum itself rather than data averaged from all the spectra. A number of derivatising procedures (1st and 2nd derivatives, Savitzky-Golay) were also carried out. The 1st derivative removes baseline effects while the 2nd derivative also removes the slope of the spectrum by measuring the change in slope, thereby sharpening spectral features. The Savitzky-Golay derivative fits a low-degree polynomial to adjacent points in a spectrum, thereby smoothing the spectrum while minimally affecting the signal-to-noise ratio.

#### *2.3. Data Analysis*

A principle component analysis (PCA) was carried out using Unscrambler X 10.3 in order to identify any gross outliers and determine any preliminary clustering reflecting the geographical origin. An inspection of the PCA loadings was carried out in order to determine whether the loadings had a spectral shape indicating that observed variation was due to the NMR spectra and not due to noise. PCA was carried out on all treated spectra to reduce all the spectral information down to seven principal components (PCs), which retained the information of the original dataset. The first PC accounted for most of the variation in the dataset, with successive principal components accounting for decreasing amounts of the variation. The resulting PC-1 vs. PC-2 plots could be examined for any clustering that might arise from each spectral pretreatment. Similarly, to PLS, PCA generates loading plots which indicate those x-values which are most responsible for the variability between the different spectra. The loading plots for the first two principal components (which explain most of the variability in the dataset) were used to determine which ppm values had the largest influence on the separation of PC-1 and PC-2. Following a PCA, supervised chemometric methods were carried out using JMP®, Version 10 (SAS Institute Inc., Cary, NC, USA), including the partial least squares discriminate analysis (PLS-DA). The whole dataset was split into two sets, termed the training and test sets (the former to build the model, the latter to validate it). In order to preserve the diversity in the training and test sets and to account for the fact that different pretreatments had to be tested, a unique sample splitting scheme was used.

In order to determine the suitability of the whole NMR spectra for discrimination of EVOOs of Maltese origin, an artificial neural network (ANN) analysis was carried out. The main advantage of a neural network model is that it can efficiently model different response surfaces due to its nonlinearity, allowing a better fit to the data given enough hidden nodes and layers, providing an accurate prediction for many kinds of data. Unlike other modeling and discriminate methods (PLS) the main disadvantage of a neural network model is that the results are not easily interpretable, due to the presence of several intermediate hidden layers. In this experiment, 25 iterations were carried out using a TanH activation function as the standard neuron activation function in JMP software. In the case of ANN, three different cross-validation techniques were employed in order to prevent model overfitting; the k-fold (CV-10), hold back (33.3%), and excluded rows (Venetian blinds). Thirty-three percent of the samples were held back from the model during holdback validation, which operates by randomly splitting the dataset into training and validation sets. Thirty-three percent of the data was thus 'held back' to form the validation set. Excluded rows holdback uses those rows that were excluded by the Venetian blinds method as the validation set. K-fold validation divides the dataset into 'k' number of subsets where each subset contains a fraction '1/k' of the data. Each of these sets is used to validate the model thereby fitting 'k' number of models. The best fitting model is presented as the final output. In this study, K-fold validation was carried out using 5 k-folds.
