Dioxin screening in fish product by pattern recognition of biomarkers.
Two alternative, cost- and time-effective dioxin screening methods relying on two categories of potential lipid biomarkers were investigated. A dioxin range varying from 1.1 to 47.1 pg PCDD/F TEQ-WHO/g lipid using 64 fish meal samples was used for model calibration. The methods were based on multivariate models using either (1) fatty acid composition monitored by GC-FID or (2) fluorescence landscape signals analysed using the PARAFAC model and in both cases predicting dioxin content as pgPCDD/F TEQ-WHO/g lipid. In both cases, Partial Least Squares (PLS) regression was performed for predicting the dioxin content of a sample. The GC-FID data analyses was based on automatic peak alignment and integration, enabling extraction of the area of 140 peaks from the gas chromatograms, as opposed to the 31 fatty acids usually considered for fish oil characterisation. In addition to classic PLS employing the whole dataset for calibration, a two-step local PLS modeling approach was performed based upon an initial selection of k number of calibration samples providing the best match to the prediction sample using a so-called k Nearest Neighbors (kNN) approach, then followed by PLS calibration on these kNN selected samples for dioxin prediction. Fluorescence spectroscopy offers a promising non-invasive and ultra-rapid technique, with less than two minutes analysis time. However, fluorescence spectroscopy using the pattern recognition "kNN-PLS" yielded a correlation of 0.76 (r2) and a high root mean square error of prediction of 11.4 pg PCDD/F TEQ-WHO/g lipid. The predictions were improved when the PLS calibration was performed on all the sample with a root mean square error of prediction of 7.0 pg PCDD/F TEQ-WHO/g lipid. Unfortunately, these results failed to demonstrate the potential of fluorophore monitoring as a screening method. In contrast, the overall best screening performance was obtained with the fatty acid profile, when the kNN-PLS combination employed for pattern recognition (kNN) all the areas of the 140 detected peaks and the PLS regression used the areas of 46 selected peaks. This "kNN-PLS" prediction with three latent variables and based upon the 12 nearest neighbors selected out of the 64 x 2 fatty acid profiles (duplicate analyses), yielded a correlation of 0.85 (r2) and a root mean square error of prediction of 2.1 pg PCDD/F TEQ-WHO/g lipid and resulted in a total analysis time of one and half hour per sample.