Chemosphere 85 (2011) 465–472

Contents lists available at SciVerse ScienceDirect

Chemosphere

journal homepage: www.elsevier .com/locate /chemosphere

Chemometric data analysis application to Sparus aurata samples from twooffshore farming plants along the Apulian (Italy) coastline

Roberto Miniero a,⇑, Gianfranco Brambilla a, Eugenio Chiaravalle b, Michele Mangiacotti b, Giulio Brizzi c,Anna Maria Ingelido a, Vittorio Abate a, Valeria Cascone c, Fabiola Ferri a, Nicola Iacovella a,Alessandro di Domenico a

a Laboratory of Toxicological Chemistry, Italian National Institute for Health, Viale Regina Elena 299 Roma, Italyb Istituto Zooprofilattico Sperimentale di Puglia e Basilicata, Via Manfredonia, 20 Foggia, Italyc Chlamis srl, Via VittorioVeneto 13 Trani (Bari), Italy

a r t i c l e i n f o

Article history:Received 26 November 2010Received in revised form 19 July 2011Accepted 29 July 2011Available online 6 September 2011

Keywords:Offshore farmsSea breamMediterranean seaPOPsDiscriminant analysis

0045-6535/$ - see front matter � 2011 Elsevier Ltd. Adoi:10.1016/j.chemosphere.2011.07.076

⇑ Corresponding author.E-mail address: roberto.miniero@iss.it (R. Miniero

a b s t r a c t

The levels of polychlorodibenzo-p-dioxins (PCDDs), polychlorodibenzofurans (PCDFs), dioxin-like poly-chlorobiphenyls (DL-PCBs), non-dioxin-like polychlorobiphenyls (NDL-PCBs), and polybromodiphenylethers (PBDEs) in fish collected from two marine offshore farming plants were determined. Each samplewas constituted by specimens of the same size collected at the same time in four different seasons alongthe farming year. The feeds given were of industrial origin and the plants were positioned in two differentsites respectively exposed to different environmental characteristics. A chemometric approach wasapplied to interpret the subtle differences observed in fish body burdens across the three chemical groupstaken into consideration. The approach consisted in a stepwise multivariate process including a hierar-chical cluster analysis (CA) and a linear discriminant analysis (DA). The two main clusters determinedby CA were subjected to the canonical DA, backward and forward selection procedures to select the bestdiscriminative functions. A clear temporal and spatial discrimination was found among the samples.Across the three chemical groups, the monthly separation seemed to depend on the growth processand the main exposure was due to the feed. In addition, the two plants differed significantly from theenvironmental point of view and the most important discriminating group of chemicals were the NDL-PCBs. The approach resulted really effective in discriminating the subtle differences and in individuatingsuggestions to improve the quality of culturing conditions.

� 2011 Elsevier Ltd. All rights reserved.

1. Introduction

Worldwide, aquaculture is more and more becoming the mainsource of fishery products, thus playing a significant role in creatingemployment opportunities in local coastal communities (FAO,2006). Within the enlarged European Union, the aquaculture indus-try accounts for a total of 1.3 million tonnes of fishery products ayear (http://ec.europa.eu/fisheries), with fish species-specific differ-ences among the areas dedicated to farming, depending on theenvironmental characteristics of the sea basin. In particular, thesea bream is the most important farmed species is in the Mediterra-nean area.

It is generally acknowledged that fish farming may increase theintake of omega-3 and omega-6 fatty acids in general population,thus helping to prevent the occurrence of cardio-vascular diseases,

ll rights reserved.

).

and in general promoting a good health status. However, suchdietary benefits may be challenged by the levels of pollutants accu-mulated by fish, such as polychlorodibenzo-p-dioxins (PCDDs),polychlorodibenzofurans (PCDFs), dioxin-like polychlorobiphenyls(DL-PCBs), non-dioxin-like polychlorobiphenyls (NDL-PCBs), andpolybromodiphenyl ethers (PBDEs). Their occurrence in fish, withrespect to other foods of animal origin, can differ of a factor 10in concentration, thus representing the most contributing factorto the daily intakes of the aforesaid contaminants (Hites et al.,2004). One way to maximize the farmed fish intake benefits, is toreduce their contamination levels by a sound use of feed itemsand the choice of an appropriate site for fish farming.

This work is aimed to characterize the differences in the accu-mulated PCDDs, PCDFs, DL-PCBs, NDL-PCBs and PBDEs in twofarming plants, possibly due to different feed regimes and the sitechoice. To this end, a chemometric approach was used to analyzethe data set consisting of farmed sea bream samples. The two off-shore farms considered were in the southern Adriatic sea and weremanaged according to a Good Agricultural Practice (GAP) regime.

466 R. Miniero et al. / Chemosphere 85 (2011) 465–472

This, in turn, means that all the factors possibly inducing a qualityreduction in the fish products are taken into consideration andpossibly eliminated. Therefore, the fish contaminant body burdenwould be expected to be under control. However, during the farm-ing period, there is a turn-over of feeds for optimal sustenance offish growth, along to a seasonal variation in feed consumption asa consequence of the diverse water temperature between winterand summer. Additionally, the two plants are exposed to differentenvironmental factors possibly including a chemical impact fromthe surrounding environment.

2. Materials and methods

2.1. Sampling and analysis

Sea bream was chosen for this study mostly because of its com-mercial relevance and its leading role in the Mediterranean aquacul-ture scenario. Fish samples were collected from two offshorefarming plants respectively called Giovinazzo and Mattinata situatedalong the Apulia coastline in southern Italy. The Giovinazzo plant isset at 3 miles from the coast at a site characterized by an 80 m depthand a fast sea stream. The Mattinata plant is relatively close to thecoast at a depth not exceeding 10 m. Sampling was performed be-tween August 2006 and April 2007, in the months of August, Novem-ber, February, and April. Fish of similar size were collected at eachsampling month and ten fillets, each one taken from a single fish,were pooled together to form the analytical sample. The fillets wereexcised from the fish body adopting the prescriptions of the EC Reg-ulation 1883/2006 (EC, 1883/2006). This procedure remained un-changed in spite of the fish weight increase due to growth (from100 to 350 g). The above-mentioned operations were performedimmediately after sampling and each individual analytical samplewas then stored separately at –20��C until pre-treatment.

Thirty ortho-substituted NDL-PCBs, 17 PCDD and PCDF congen-ers, eight mono-ortho DL-PCBs, the non-ortho-substituted PCBs 77,81, 126 and 169, and PBDE 17, 28, 47, 66, 99, 100, 138, 153, 154,183 and 209 were determined. The analytical procedure wasadapted from the USEPA Method No. 1613 (1994) for PCDD andPCDF determination. An adaptation of such procedure was de-scribed by De Felip and Miniero (1999).

Upon delivery to the laboratory for pretreatment the analyticalsamples were allowed to thaw and rinsed with distilled water.Then, they were added with 13C-labeled standards, allowed to restfor hours, and homogenized with anhydrous Na2SO4. An aliquot ofeach sample was extracted using a Soxhlet apparatus with a 50%mixture of acetone and n-hexane. The extract was concentratedto 20 mL using a rotary evaporator; a 2 mL aliquot was used for li-pid content determination by a gravimetric method. Clean-up wascarried out by filtration through Extrelut impregnated with con-centrated sulfuric acid, resting on a silica gel layer (di Domenicoet al., 1992). The extract was then analyzed for PCDDs, PCDFs,DL-PCBs, NDL-PCBs and PBDEs, with three different Power-Prepseparation programs.

Quantification was performed by high-resolution gaschromatography coupled with low-resolution mass spectrometry(HRGC-LRMS) used in the selected ion monitoring mode (SIM) forNDL-PCBs and PBDEs. HRGC-HRMS(SIM) was employed to deter-mine PCDDs, PCDFs, and DL-PCBs operating at 10,000 mass resolu-tion. A procedural blank was run together with three to fivesamples. Reliable measurements were allowed above the limit ofdetermination with a repeatability in the order of |±10%| (extendeduncertainty, |±20%|). The recovery rates of labeled ISs were ac-cepted within 40–120%; values outside this range led to specificevaluation, possibly rejection of trial.

2.2. Data treatment

Most multivariate methods require variables to conform to anormal distribution, thus the normality of the distribution of eachvariable was checked by analyzing kurtosis and skewness beforemultivariate statistical analysis. The original data were demon-strated to be almost normally distributed and positively skewedwith standardized Kurtosis coefficients greater than zero(p = 0.005). After data log-transformation, all skewness and kurto-sis values were significantly reduced, therefore log-transformationwas adopted for data analysis. For each sample, congener concen-trations were normalized against the pertinent cumulative analyt-ical concentration, the latter obtained by adopting the mediumbound approach when dealing with limits of determination. Afterdata validation, congeners with a high incidence (>40%) ofnon-detects throughout the samples were removed to minimizethe non-detect impact on the statistical approach (Miniero et al.,2007).

2.3. Chemometric approach

The levels of the above-mentioned chemicals were analyzed bya stepwise statistical approach including a hierarchical clusteranalysis (CA) and a linear discriminant analysis (LDA). The CA usesiterative approaches to find structure in data by identifying theirnatural groupings whereas the LDA procedure is designed to dis-tinguish between two or more groups of data based on a set of pobserved quantitative variables. It does so by constructing discrim-inant functions that are linear combinations of the variables. Thediscriminant scores, a single new composite variable, are calcu-lated by these functions which have the form

D ¼ w1Z1 þw2Z2 þw3Z3 þ . . . wpZp

where D is the discriminant score, wp the weighting (coefficient) forvariable p, Zp is the standardized coefficient for variable p.

An essential requirement of the LDA procedure is the normallydistributed data (Fielding, 2007). The goal of this approach is todisplay the most significant patterns, looking for possible group-ings and sources of data variation, as well as for their temporaland spatial distributions, through resolution and modeling of data(Goncalves et al., 2006; Zhou et al., 2007). For CA, a squared Euclid-ean distance was always used as the interval measure for cluster-ing using the following distinct linkage methods: between-grouplinkage, within-group linkage and Ward’s methods. The groups ofvariables individuated by the CA were singly evaluated by thecanonical, forward (FS) and backward (BS) LDA, with the objectiveto distinguish the ones which have driven the formation of thegroups of closed related samples. The forward and backward DAprocedures adopted to construct the best discriminating functionstook into account the statistical significance of the variables and inthe data evaluation a critical F value of 4 was used (F-to-enter andF-to- remove). The two classification factors considered, the sam-pling month (August, November, February, May) and the fish plant(Giovinazzo, Mattinata) were known as possible sources of datavariation due to the influence of different feed types and of differ-ent environmental characteristics.

3. Results and discussion

3.1. Data description

The lipid-base data on PCDD, PCDF, DL-PCBs NDL-PCB, andPBDE congeners are reported in Table 1. Among the PCDD andPCDF congeners the most important ones defining the sample-spe-cific profile are D1, D2, F1, F2, F3, F4, and F5. The highest

Table 1Lipid-based concentrations of PCDDs, PCDFs, DL-PCBs, NDL-PCBs and PBDEs determined in fish tissues collected from the two farming plants in the months of August, November, February and April.

Samples Contamination levels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16Planta G G M M G G M M G G M M G G M MSampling month August November February April

PCDD + PCDF(pg/g lb)D1b 0,20 0,20 0,15 nd 0,23 0,20 0,13 0,19 0,27 0,22 0,20 0,17 0,39 0,31 0,32 0,28D2 0,33 0,32 0,24 nd 0,34 0,32 0,22 0,12 0,45 0,36 0,28 0,29 0,50 0,48 0,35 0,34D3 0,06 nd nd nd nd nd nd nd nd nd nd 0,11 nd 0,15 nd 0,16D4 0,12 0,25 nd nd 0,15 nd nd nd nd 0,18 nd 0,20 nd 0,23 nd 0,22D5 nd nd nd nd nd nd nd nd nd nd nd 0,13 nd 0,09 nd 0,11D6 0,13 0,19 0,26 0,46 0,72 nd 0,18 nd nd 0,11 nd 0,28 nd 0,20 nd 0,29D7 nd 0,04 nd nd nd nd nd nd nd nd nd nd nd nd nd ndF1 4,56 4,37 3,85 4,00 5,35 4,31 3,19 2,98 4,48 4,36 4,25 4,05 6,44 4,82 4,93 4,15F2 0,59 0,66 0,49 0,74 0,67 0,55 0,39 0,41 0,87 0,78 0,62 0,62 1,08 0,83 0,70 0,72F3 1,21 1,25 0,66 0,88 1,20 1,08 0,58 0,55 1,62 1,56 0,90 1,03 1,96 2,06 1,08 1,12F4 0,22 0,25 0,13 0,28 0,31 0,20 0,09 0,11 0,49 0,17 0,23 0,19 0,44 0,27 nd 0,27F5 0,09 0,13 0,10 nd 0,11 0,12 0,06 nd 0,25 0,19 0,11 0,17 0,26 0,21 nd 0,15F6 nd nd nd nd 0,09 nd nd nd nd nd nd nd nd nd nd ndF7 nd 0,10 0,08 nd 0,08 nd nd nd 0,24 0,19 0,19 0,24 0,31 0,26 nd 0,24F8 nd nd nd nd 0,20 nd nd nd 0,25 nd 0,29 nd 0,51 nd 0,31 ndF9 nd nd nd nd 0,10 nd nd nd nd nd nd nd nd nd nd ndF10 nd nd nd nd 0,41 nd nd nd nd nd nd nd nd nd nd ndP

2,3,7,8-PCDD + PCDFc 8,1 8,5 7,3 11,1 20 8,8 6,6 5,8 12 8,8 9,5 8,3 15 10,6 12 9,3

DL-PCB (pg/g lb)non-ortho PCBT4CB 77 175 134 169 110 218 132 176 131 233 195 205 202 283 235 307 234T4CB 81 6,3 5,2 7,0 4,9 9,6 5,4 6,7 4,9 8,1 7,7 7,3 7,6 10,3 9,4 9,6 9,2P5CB 126 59,6 54,8 55,7 44,6 66,1 52,1 56,6 44,7 95,1 59,0 67,9 55,8 111 74,8 104 70,8H6CB 169 8,4 10,2 7,7 8,8 9,1 8,4 7,5 6,3 12,0 11,7 10,5 12,0 16,6 14,5 13,7 14,5

mono-ortho PCBP5CB 105 3432 3257 2979 3011 4249 3520 3299 2935 5560 4147 3973 3348 6671 4974 5139 4469P5CB 114 158 181 159 144 205 157 160 153 230 187 198 183 271 200 202 190P5CB 118 11 375 11 358 10 998 10 207 15 207 12 068 11 654 10 194 16 639 12 602 12 865 11 283 21 342 15 924 16 464 12 612P5CB 123 598 650 629 618 1078 655 628 602 850 600 778 490 1228 767 840 609H6CB 156 935 973 856 856 1345 955 1018 850 1873 1229 1345 1229 2131 1572 1857 1503H6CB 157 247 259 242 244 321 251 252 215 468 335 380 291 581 393 502 392H6CB 167 748 714 657 638 975 693 759 561 1530 829 956 1009 1667 1184 1254 1050H7CB 189 81,6 89,0 75,0 88,5 103 69,4 89,6 70,3 168 112 127 150 200 140 188 152P

WHO-TEQ PCDD+PCDFc 1,68 1,71 1,18 1,18 1,82 1,58 1,02 0,95 2,18 1,92 1,49 1,52 2,73 2,47 1,85 1,76

PWHO-TEQ DLPCB

c 8,29 7,84 7,76 6,59 9,74 7,63 8,05 6,55 13,3 8,66 9,66 8,11 15,8 10,9 14,1 10,1P

WHO-TEQ totalc 9,96 9,55 8,95 7,77 11,6 9,21 9,06 7,49 15,5 10,6 11,1 9,64 18,5 13,4 16,0 11,8

NDL-PCB (ng/g lb)PCB 18 nd nd 0,57 nd 0,38 nd 0,5 nd nd nd nd nd nd nd nd ndPCB 28 2,61 1,41 2,91 nd 2,68 1,67 2,81 1,34 nd nd nd nd nd nd nd ndPCB 31 1,35 0,77 1,4 nd 1,85 0,93 1,81 0,59 nd nd nd nd nd nd nd ndPCB 33 nd nd 0,6 nd 0,51 nd 0,71 nd nd nd nd nd nd nd nd ndPCB 41 1,4 1,17 1,35 1,31 1,74 1,27 1,56 1,05 nd 0,99 nd nd nd 1,33 nd 0,99PCB 44 0,64 0,5 0,73 0,69 0,93 0,66 0,84 0,56 nd nd nd nd nd nd nd ndPCB 47 0,99 0,78 1, 0,84 1,08 0,85 1,12 0,76 nd nd nd nd nd 0,76 nd ndPCB 49 0,82 0,69 1,04 0,93 1,03 0,81 1,08 0,79 nd nd nd nd nd nd nd 0,55PCB 51 nd nd 0,02 0,04 nd nd 0,03 nd nd nd nd nd nd nd nd ndPCB 52 4,34 3,75 3,93 3,79 4,7 4,15 4,19 3,41 4,53 4,11 4,16 3,55 5,83 5,0 5,18 4,42PCB 56/60 0,97 0,79 0,88 0,74 1,62 0,91 0,99 0,69 nd 0,6 nd nd nd 0,82 nd 0,63PCB 61 1,46 1,25 1,2 1,12 1,81 1,38 1,34 1,05 2,02 1,78 1,63 1,45 2,61 2,19 2,03 1,69

(continued on next page)

R.M

inieroet

al./Chemosphere

85(2011)

465–472

467

Table 1 (continued)

Samples Contamination levels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16Planta G G M M G G M M G G M M G G M MSampling month August November February April

PCB 66 2,33 1,95 1,95 1,8 2,94 2,11 2,17 1,72 2,61 2,37 2,24 2,01 3,52 2,94 2,83 2,43PCB 70 2,75 2,41 2,21 2,13 3,67 2,59 2,45 1,92 2,14 2,04 1,87 1,68 2,84 2,61 2,39 1,99PCB 74 nd nd nd nd nd nd nd nd nd nd nd nd nd nd nd ndPCB 87 2,05 1,89 1,64 1,71 3,07 1,98 1,93 1,95 3,34 2,84 2,78 2,19 4,58 3,09 3,68 2,73PCB 91 0,37 0,31 0,37 0,36 0,48 0,38 0,41 0,36 0,5 0,37 0,44 0,35 0,57 0,43 0,56 0,44PCB 95 3,3 2,86 2,83 2,69 5,4 3,19 3,38 2,8 3,94 3,28 3,64 2,79 4,91 3,65 4,37 3,59PCB 99 4,62 4,33 4,07 4,01 5,21 4,53 4,4 4,21 7,75 5,85 6,28 5,23 10,15 6,75 8,15 6,41PCB 101 9,79 8,99 8,75 8,35 15,89 9,69 9,68 8,32 13,4 10,2 10,6 8,09 16,4 11,9 14,1 10,6PCB 110 3,85 3,4 3,26 3,24 6,02 3,6 3,55 3,67 6,09 4,99 5,28 4,08 8,49 5,94 7,08 5,23PCB 128 1,88 1,92 1,62 1,67 2,36 1,89 1,75 1,59 3,8 2,33 2,9 2,17 4,38 2,82 3,82 2,63PCB 129 0,22 0,23 0,21 0,23 0,39 0,22 0,22 0,2 0,32 0,23 0,3 0,23 0,41 0,27 0,33 0,29PCB 138 22,1 21,9 19,5 19,97 30,4 21,7 21,9 18,1 38,4 27,2 30,6 26,0 45,8 33,0 40,8 30,6PCB 141 1,4 1,35 1,26 1,21 2,87 1,28 1,49 1,11 2,52 1,57 1,91 1,43 2,8 1,96 2,64 1,8PCB 146 3,44 3,38 3,03 3,1 4,3 3,13 3,32 2,74 6,13 4,17 4,78 3,89 7,2 5,07 6,31 4,81PCB 149 7,08 6,7 6,19 6,11 12,7 6,59 7,07 5,57 10,93 7,74 8,85 6,47 12,72 9,45 11,25 8,33PCB 151 2,42 2,27 2,09 2,06 4,24 2,18 2,33 1,83 3,75 2,86 3,17 2,47 4,47 3,44 3,95 3,1PCB 153 21,8 22,2 20,3 20,8 29,3 20,6 22,5 18,6 37,6 26,5 30,8 25,5 44,0 32,1 40,0 31,4PCB 170 1,49 1,53 1,67 1,59 1,94 1,34 1,75 1,43 2,25 1,73 2,32 2,5 2,42 2,45 3,01 2,78PCB 174 1,06 1,14 1,08 1,05 1,92 0,95 1,18 0,9 1,58 1,13 1,77 1,34 2,35 1,39 2,27 1,6PCB 177 1,3 1,38 1,1 1,19 1,78 1,13 1,27 1,01 2,21 1,52 1,8 1,55 2,8 1,88 2,42 1,81PCB 180 4,92 5,21 5,74 5,81 6,6 4,57 6, 4,97 11,2 7,11 10,5 9,18 12,9 8,56 13,7 10,9PCB 183 1,39 1,49 1,56 1,55 2,18 1,27 1,6 1,31 2,97 1,76 2,7 2,15 3,88 2,14 3,68 2,53PCB 185 0,1 0,11 0,12 0,1 0,19 0,09 0,12 0,08 0,22 0,14 0,2 0,16 0,29 0,16 0,26 0,18PCB 187 4,44 4,74 4,44 4,55 5,41 4,0 4,59 3,78 8,93 5,71 7,75 6,43 11,81 6,8 10,5 7,45PCB 191 0,1 0,05 0,05 0,06 0,08 0,04 0,06 0,05 0,09 0,07 0,11 0,1 0,11 0,08 0,09 0,11PCB 193 0,42 0,46 0,39 0,47 0,47 0,34 0,46 0,39 0,5 0,44 0,39 0,58 0,52 0,58 0,56 0,62PCB 194 0,31 0,33 0,41 0,38 0,36 0,25 0,38 0,32 nd 0,42 nd 0,65 nd 0,54 nd 0,76PCB 201 0,61 0,61 0,73 0,71 0,74 0,38 0,78 0,6 0,73 0,68 0,78 1,07 0,87 0,86 1,05 1,22PCB 202 0,26 0,25 0,26 0,26 0,28 0,2 0,28 0,21 nd 0,29 nd 0,37 nd 0,35 nd 0,42PCB 196/203 0,66 0,66 0,73 0,71 0,8 0,41 0,83 0,62 0,75 0,66 0,72 0,98 0,85 0,82 0,99 1,1PCB 206 0,11 0,12 0,18 0,12 0,16 0,05 0,14 0,07 nd 0,13 nd 0,21 nd 0,16 nd 0,23PCB 208 0,08 0,07 0,09 0,05 0,08 0,02 0,07 0,03 nd 0,06 nd 0,1 nd 0,07 nd 0,11PCB 209 0,07 0,05 0,07 nd 0,08 nd 0,08 nd nd 0,09 nd 0,1 nd 0,13 nd 0,14P

PCBc 121 115 113 107 171 113 125 101 181 134 151 127 220 162 198 156

PBDE (ng/g lb)BDE 28 0.253 0.215 0.573 n.d. 0.269 0.216 0.208 0.141 0.370 0.226 0.212 0.174 0.389 0.249 0.219 0.191BDE 47 5.41 5.22 3.96 4.00 5.18 4.83 4.25 3.68 7.97 6.63 5.11 4.32 10.6 7.20 6.14 5.37BDE 66 0.279 0.262 0.199 0.193 0.261 0.250 0.206 0.187 0.369 0.348 0.229 0.210 0.436 0.350 0.278 0.255BDE 99 0.449 0.489 0.364 0.367 0.553 0.495 0.399 0.352 0.571 0.415 0.419 0.377 0.665 0.457 0.504 0.398BDE 100 1.21 1.30 0.936 1.07 1.20 1.06 0.959 0.822 1.97 1.51 1.22 1.17 2.53 1.74 1.52 1.36BDE 153 0.138 0.159 0.104 0.113 0.139 0.114 0.110 0.0866 0.213 0.169 0.134 0.166 0.302 0.226 0.190 0.178BDE 154 0.488 0.488 0.394 0.458 0.523 0.425 0.436 0.333 0.974 0.809 0.612 0.765 1.53 0.992 0.883 0.793BDE 183 n.d. 0.197 n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d.BDE 209 n.d. n.d. 0.374 n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d. n.d.P

BDEc 8,3 8,4 6,9 6,3 8,2 7,5 6,6 5,7 12,5 10,2 8,0 7,2 16,6 11,3 9,8 8,6

a plant: G, Giovinazzo,; M, Mattinata.b D1, 2,3,7,8-T4CDD; D2, 1,2,3,7,8-P5CDD; D3, 1,2,3,4,7,8-H6CDD; D4, 1,2,3,6,7,8-H6CDD; D5, 1,2,3,7,8,9-H6CDD; D6, 1,2,3,4,6,7,8-H7CDD; D7, O8CDD; F1, 2,3,7,8-T4CDF; F2, 1,2,3,7,8-P5CDF; F3, 2,3,4,7,8-P5CDF; F4, 1,2,3,4,7,8-

H6CDF; F5, 1,2,3,6,7,8-H6CDF; F6, 1,2,3,7,8,9-H6CDF; F7, 2,3,4,6,7,8-H6CDF; F8, 1,2,3,4,6,7,8-H7CDF; F9, 1,2,3,4,7,8,9-H7CDF; F10, O8CDF.c Medium bound.

468R

.Miniero

etal./Chem

osphere85

(2011)465–

472

(a)

(b)

0

10

20

30

40

50

60

0

10

20

30

40

50

60

Fig. 1. Dendrograms of the DL-Compounds and NDL-PCBs identifying two majordata sets. A complete linkage method was adopted to evaluate them.

R. Miniero et al. / Chemosphere 85 (2011) 465–472 469

chlorinated ones — D7, F8, F9, and F10, in spite of their high Kow

values were not found in appreciable quantities probably becausetheir steric hindrance affecting their bioavailability (Bruggemanet al., 1984). The highest WHO-TEQ (Van den Berg et al., 1998) va-lue was determined in a sample collected from the Giovinazzofarming plant in November (20.6 pgWHO-TE/g lb). However,

Table 2Temporal discrimination. Summary of the first discriminant function data for the chemicalsfactors are the sampling months.

Discriminant strategy ka Variance percentage Wilks Lambda

DL Compounds

First group of variablesCanonical – – –Backward 42.2 90.28 0.00347Forward 4.69 100 0.176

Second group of variablesCanonical 6.01 84.86 0.0649Backward 3.65 100 0.215Forward 3.65 100 0.215

NDL PCBs

First group of variablesCanonical – – –Backward 498 86.49 0.0000169Forward 56.4 99.97 0.0171

Second group of variablesCanonical – –Backward 419 95.6 0.0000709Forward 17.7 89.22 0.0469

BDE compoundsCanonical 11.4 82.46 0.0176Backward 10.6 99.28 0.0801Forward 10.6 99.28 0.0801

a Eigenvalue.b Chi-square.

considering all theP

WHO-TEQ total values, any specific trend attime and spatial levels cannot be individuated. The non-orthoand mono-ortho DL-PCBs components were always determined inall the specimens analyzed, a phenomenon observed by otherauthors (Abad et al., 2003). Indeed, in wild as well as farmed fishthey usually contribute at a ratio of 3:1 with respect to theWHO-TEQ contribution from PCDDs + PCDFs. In fish feed their con-tribution is approximately at the same level (Abalos et al., 2008a);therefore, at least in farmed fish this source can significantly influ-ence the total WHO-TEQ body burden. NDL-PCB congeners weredetermined in all the samples analyzed (Table 1). In particular, inthe specimens collected in the months of February and April con-centrations of the three- and the four-chlorine substituted NDL-PCBs 28, 31, 33, 41, 44, 47, and 49 were below their respectivedetection limits. From the

PPCB values, it seems to exist a tempo-

ral and spatial trend which suggests an increase of estimates in thelast two sampling periods and, at spatial level, in the Giovinazzoplant. Among the PBDEs, PBDE 47 is the most abundant compound.Its widespread diffusion in different environmental matrices sug-gests that the environment can constitute a significant source forfish body burden (Hale et al., 2001) besides the feed component(Easton et al., 2002). Its concentration together with the ones fromthe other PBDE compounds are higher in the Giovinazzo than inthe Mattinata samples. In addition, as observed for the PCDD andPCDF group, the highest brominated PBDEs 183 and 209 werenot found in fish tissues.

3.2. Cluster analysis

An initial exploratory approach involved the use of hierarchicalCA on standardized data. The dendrogram in Fig. 1a shows the re-sult obtained by the clustering method of Further Neighbor. TheWard’s and Group Average methods gave the same subdivision;however the above-mentioned method had the best discriminatingpower for the variables. Sample clustering (not reported here)shows a separation clearly referred to a temporal trend, the firstcluster dealing with August and November samples whereas the

groups considered and for the two cluster of variables showed in Fig. 1. The classifying

v2b df p % Classified Variables Number

– – – –16.138 0.0074 87.5 921.73 3 0.0001 100 1

28.72 15 0.0175 87.5 519.21 3 0.0002 62.5 119.21 3 0.0002 62.5 1

– – – –76.90 36 0.0001 100 1248.81 6 0.0000 43.75 2

– – –66.88 6 0.013 75 1236.72 6 0.0000 75 2

38.38 21 0.0116 93.75 730.30 6 0.0000 75 230.30 6 0.0000 75 2

AugustAprilFebruaryNovemberCentroids

++++

+

AugustNovember

FebruaryApril

-5

0

5

AugustAprilFebruaryNovemberCentroids

++++

+

AugustNovember

FebruaryApril

AugustAprilFebruaryNovemberCentroids

+

+ +

AugustNovemberFebruary

April

-6 -3 0 3 6 9 12

-9 -6 -3 0 3 6 9-3

-1,5

0

1,5

3AugustAprilFebruaryNovemberCentroids

+

+ +

AugustNovemberFebruary

April

AprilFebruaryNovemberCentroids

++

+

++

August

November

FebruaryApril

-5 0 5-2

-1

0

1

2AugustAprilFebruaryNovemberCentroids

++

+

++

August

November

FebruaryApril

First function

Seco

nd f

unct

ion

(a)

(b)

(c)

Fig. 2. Maps of class predictions. Labels (?) represent the original class, whereas thecircles define the actual ones. Axes are the discriminant functions and thecoordinates are the case scores. (a) DL-compounds, (b) NDL-PCBs, and (c) PBDEs.

470 R. Miniero et al. / Chemosphere 85 (2011) 465–472

second with February and April. The variables clustering in Fig. 1also suggests grouping them into two main groups, which do notexhibit any chemical similarity among the variables. Therefore,the separation appears to be driven by a strong impact maybedue to chemicals present in fats of feed components. Indeed, theapparent lack of chemical similarity among variables clusteredclose to each other could depend on the various ingredients andprocessing methods adopted in industrial feed formulations. Thefat sources for feed components are diverse and during the feedpreparation they are processed by physical and chemical refiningmethods (Abalos et al., 2008a); therefore the original contamina-tion patterns can result deeply modified. The highest chlorinatedcompounds, O8CDD, O8CDF, and 1,2,3,4,7,8,9-H7CDF are not in-cluded in these two groups due to a high incidence (>40%) ofnon-detects. In other words, these three compounds which arepresent in feed fats (Abalos et al., 2008a), apparently are not signif-icantly accumulated by fish maybe for their steric hindrance(Bruggeman et al., 1984). The hierarchical clustering of NDL-PCBsshown in Fig. 1b appears to be slightly different from the oneshown in Fig. 1a. Here, the clustering process appears to be drivenalso by chemical similarity characteristics, in particular the chloro-substitution degree. Two large clusters can be clearly identified asobserved for the DL-compounds, therefore the most important fac-tor driving the NDL-compounds partitioning is probably the same.

However the cluster specific chemical similarity seems to suggest afactual association between this factor and additional factorswhich counterbalance in part the strong influence of the former,e.g. feed and environment.

3.3. Discriminant analysis

3.3.1. Temporal trendIn order to distinguish the role of variables in classifying the

samples, each one of the two groups, identified by the CA was ana-lyzed by Linear Discriminant Analysis (LDA). In particular, to pro-vide the best discriminating functions, the canonical, BS and FSprocedures were adopted to evaluate the CA clusters in relationto temporal variations. The parameters reported in Table 2 summa-rize the results concerning the temporal discrimination. On aver-age, the discriminant function one is almost ten times as large asthat for function two, suggesting greater separation between theclasses, therefore data on the second one are not reported. The per-centage of variance indicates its relative importance. The Wilk’sLambda is a measure of the discriminating power remaining inthe variables, and that values close to zero indicate high discrimi-nating power. At p 6 0.05 the BS and the FS procedures resulted al-ways significant, the canonical one being important only for thesecond variable cluster of the DL-compounds and the PBDEs.Among the three chemical groups, only the two NDL-PCBs variableclusters appear to be really discriminative, though the first clusterfunction significance is higher than the second one. On average theChi-Square values suggest the temporal variation in the data set tobe significant. On the whole, there is a clear significance increasefor function one moving from the canonical to the forward proce-dure. This increase is correlated to a decrease of the number of dis-criminating variables not followed by a parallel increase of classseparation. Therefore, it seems that the optimal number of separa-tion variables lies between the values indicated for the BS and FS,although the variable role (e.g. DL-compounds vs NDL-PCBs vsPBDEs) can significantly influence it. However, the function signif-icance increase is in general agreement with the ratios cases/pre-dictors which increases proportionally. Indeed, the scientificliterature suggests that ratios higher than 3:1 would give very goodseparation interpretations (Fielding, 2007). Therefore, from thispoint of view the BS and FS procedure approaches the theoreticalbest ratios for NDL-PCBs and PBDEs.

By plotting the first group of discriminating functions (Fig. 2), thefirst function discriminates among the months by coupling themsubstantially in two groups whereas the second one (not shown inTable 2) does not. In the figure, the results of the DL-compoundsBS and of the NDL-PCB and PBDE FS are reported. Across the threechemical groups and the adopted DA procedures a common trendcan be observed. There is a net separation of February and April fromAugust and November, and for the DL-compounds this separation iscomplete even between November and August. As to the other twochemical groups and for the couple February and April of the DL-compounds there is an almost complete overlapping. Remarkably,concerning the month- and chemical-specific profiles (not shownhere), DL-compound, NDL-PCB, and PBDE contributions, in qualita-tive terms are the same across the months, although they signifi-cantly differ in terms of standardized discriminant functioncoefficients. The profile composition could depend on two main fac-tors: different exposure sources and fish metabolism. Differentsources can induce two types of variation: different congener com-position and/or different quantities of the same congener. In thiscase, the importance of the first cause could be minimized becausethe composition across the months is the same whereas the secondcan contribute to the observed variation which is supposed to varyamong the months. Therefore, this behavior appears to be relatedto a common exposure source (e.g. feed), eventually modulated by

Table 3Spatial discrimination. Summary of the first discriminant function data for the chemicals groups considered and for the two cluster of variables showed in Fig. 1. The classifyingfactor are the farming plants.

Discriminant strategy ka Variance percentage Wilks Lambda v2b df p % Classified Predictors number

DL compounds

First group of variablesCanonical – – – – – – –Backward – – – – – – –Forward 7.19 100 0.122 26.28 3 0.0000 100 3

Second group of variablesCanonical 1.85 100 0.351 12.03 5 0.0343 5Backward 1.29 100 0.437 10.34 3 0.0159 87.5 3Forward – – – – – – 87.5 –

NDL PCBs

First group of variablesCanonical – – – – – – –Backward 1.14E3 100 0.000663 51.23 14 0.0000 100 14Forward 6.63 100 0.131 25.4 3 0.0000 100 3Second group of variablesCanonical 1.36 – – – – –Backward – 100 0.0106 31.84 0.0042 14Forward – 100 0.653 34.10 0.0000 5

PBDE CompoundsCanonical 14.7 100 0.0852 27.08 7 0.0001 100 7Backward 7.6 100 0.116 26.90 3 0.0000 93.73 3Forward 7.14 100 0.123 26.21 3 0.0000 100 3

a Eigenvalue.b Chi-square.

Table 4Standardized discriminant function coefficients of the FS procedure and structurecoefficients for the discriminant functions showed in Table 3. In this table only theresults for the first cluster of variables is shown.

Standardized canonical discriminant functioncoefficients

Classified functioncoefficients

Giovinazzo Mattinata

DL compoundscb 167 1.36 1.14E + 3 1.09E + 3cb 189 2.36 �992 �933f3 1.41 780 725NDL-PCBscb 201 3.42 775 873cb 206 �2.95 �278 �311cb 31 1.09 99.7 111PBDEBDE 153 2.52 283 242BDE 154 �4.78 �495 �437BDE 66 2.64 825 750

R. Miniero et al. / Chemosphere 85 (2011) 465–472 471

a temporal common factor shared by all fish (e.g., the growth metab-olism). As to the discriminating variables, only PCB 156 is selectedamong the DL-compounds whereas PCBs 44 and 70 and PBDEs 99and 154 are respectively selected among NDL-PCBs and PBDEs, allof them being chemicals frequently determined in different fishfeeds (Easton et al., 2002; Kim et al., 2007; Abalos et al., 2008a,b). Be-yond the cited ones, all the potential predictors have structure coef-ficients and they can be still be used to interpret differences amongthe months (Fielding, 2007) but they have not been reported here.The above-mentioned assumption about a common exposure sourcevarying among the months seems to be confirmed by the behavioracross chemical groups which is almost the same. In a farming con-dition, February and April are the months characterized by a growthrecovery from the winter starvation and the highest growth rate.This probably induces a dilution of fish contaminant body burden(Brambilla et al., 2007; Abalos et al., 2008b) and therefore the lowestscores of the function one shown in Fig. 2. On the contrary, Augustand November are respectively characterized by the highest feed in-take and reduced intake and growth (Abalos et al. 2008b). These fac-tors may influence, in turn, the body burden by increasing thecontaminant loads and therefore determining the positive scoresof Fig. 2.

3.3.2. Spatial trendThree spatial discrimination appears to be more effective for the

first group of variables and across the three chemical groups (Table3). Also in this case, the conclusions of the application of the DAstrategies on temporal discrimination appear to be confirmed.Among the chemical groups, the NDL-PCBs confirm their discrimi-nating role. For the DL-compounds, only the FS procedure resultseffective in discriminating between the two farming plants, theprocedure best complying with the above-mentioned criteriumon ratio case/predictor. The discriminating predictors are PCBs167 and 189 and F3 (Table 4), which has the highest WHO-TEQcontribution in the fish oil samples used for preparing fish feed(Abalos et al., 2008a). Between the two plants, Giovinazzo isassociated with the highest function coefficients, therefore it

confirms the detection of the highest WHO-TEQ values in fish tis-sues (Table 1). The industrial production of feed is based on ingre-dients that are processed to reduce the contaminants carryover toanimals; therefore, the original differences among the raw materi-als can result significantly reduced by these processes (Abaloset al., 2008a,b) together with the influences on fish body burdens.In this two plants there is a large overlapping between the diets,however the differences in the use of different brands and feed re-gimes can result in some different chemical-specific body burdens.

About the NDL-PCBs, the FS and the BS procedures separate thetwo plants, with a clear decrease of eigenvalues and a correspond-ing increase of the Wilk’s Lambda values of the first group vari-ables. From the Table 4, the discriminating predictors are PCBs31, 201 and 206, for this group and the FS procedure. The loadingsare higher for the Mattinata plant although, in qualitative term, thecomposition is the same. The variables selected by the BS proce-dure (not shown in the table) are the PCBs 91, 101, 129, 149,170, 191, 193, 194, 196#203, 201, 202 and 206. They show

472 R. Miniero et al. / Chemosphere 85 (2011) 465–472

differences in qualitative as well as in quantitative terms betweenthe two plants, thereby suggesting a differential impact on them. Inparticular, the variables show on average the highest loadings forthe Mattinata plant, thus highlighting a possible impact from theenvironment because this plant is placed at 1 km from the coaston a shallow sediment bed (�10 m of depth). This situation can in-crease the sediment role in contaminating the water column and,therefore, in the transfer of chemicals through the local foodchains. This congener assemblage appears to follow the Fig. 1bCA clustering because all of them selected by the BS procedure, ex-cept PCB 91, belong to the first three sub-clusters. All these cong-eners are represented in fish feed with variable amounts (Eastonet al., 2002), this suggesting a major contribute from feed anyway,it cannot be excluded an environmental influence because the twoplants are significantly differentiated. Therefore, this evaluationappears not in contrast with the trends suggested in the ‘‘DataDescription’’ because the profiles describe the qualitative natureof data.

As well as for the temporal discrimination, all PBDEs com-pounds were analyzed together because the CA did not show anyseparation in two groups. In Table 3, the Wilks’ Lambda and theChi-Square statistics suggest that there is a discrimination betweenthe two plants (indeed 100% of the samples were classified). Thepredictors determined by the FS procedure are shown in Table 4and their function coefficients suggest a differential impact onthe plants of the same type as the DL-compounds, because trendsamong the coefficients are the same. This feature again suggests acommon factor eliciting it, i.e. the feeds, which is more evident forthe Giovinazzo plant. Indeed, the PBDEs feed contents appear tosignificantly influence fish body burden (Easton et al., 2002), how-ever, considering both the factors of overlapping diets and localexposure highlighted respectively by DL-compounds and NDL-PCBs an environmental influence cannot be excluded a priori.

4. Conclusions

On the whole, the major separation of the variable groups ofFig. 1 determined by CA reflects the most discriminating variablesevidenced by the DA. On average, the first group of variables showsthe best performances in terms of eigenvalues, Wilk’s Lambda, ChiSquare, significance and % cases classified. In addition, the cluster-ization process seems to be driven also by chemical-specific differ-ences such as the chlorination degree. The DA appears to be a reallyeffective tool to distinguish between different influencing factors,although all the above procedures (Canonical, BS and FS) need tobe taken into consideration. An additional useful criterium to as-sess the critical issue of DA performance is the ratio case/predictor.

At the low level exposure experienced by the fish, body growth,feed intake rate and the farming plant site appear to be of criticalimportance in determining the fish body burden. All these factorsneed to be maintained under a close control to avoid possible risksof accumulating chemicals by fish thereby focusing the attention

on their well-being and on a careful application of Best AgriculturalPractice. For example, despite the implementation of internationaltreaties such as the Stockholm Convention (UNEP, 2001), and thePCB phasing out in many countries, PCBs still seem to representa threat for the environment.

References

Abad, E., Abalos, M., Calvo, M., Nombela, J., Caixach, J., Rivera, J., 2003. Surveillanceprogram on dioxins and dioxin-like PCBs in fish and shellfish consumed inSpain. Organohalogen Compounds 62, 49–52.

Abalos, M., Parera, J., Abad, E., Rivera, J., 2008a. PCDD/Fs and DL-PCBs in feeding fatsobtained as co-products or by-products derived from the food chain.Chemosphere 71, 1115–1126.

Abalos, M., Abad, E., Estevez, A., Solè, M., Buet, A., Quiros, L., Pina, B., Rivera, J.,2008b. Effects of growth and biochemical responses in juveniles gilthead seabram Sparus aurata after long-term dietary exposure to low levels of dioxins.Chemosphere 73, S303–S310.

Brambilla, G.F., Dellatte, E., Fochi, I., Iacovella, N., Miniero, R., di Domenico, A., 2007.Depletion of selected polychlorinated biphenyl, dibenzodioxin, anddibenzofuran congeners in farmed rainbow trout (Onchorhyncus mykiss): ahint for safer fish farming. Chemosphere 66, 1019–1030.

Bruggeman, W.A., Opperhuizen, A., Wijbenga, A., Hutzinger, O., 1984.Bioaccumulation of super-lipophilic chemicals in fish. Toxicological andEnvironmental Chemistry 7, 173–189.

De Felip E., Miniero R., 1999. Procedimenti Analitici Adottati per il Rilevamento diMicrocontaminanti in Sedimenti Lagunari. ISTISAN99/28. Istituto Superiore diSanità, Roma.

di Domenico, A., De Felip, E., Ferri, F., Iacovella, N., Miniero, R., Scotto di Tella, E.,Turrio Baldassarri, L., 1992. Determination of the composition of complexchemical mixtures in the soil of an industrial site. Microchemical Journal 46,48–81.

Easton, M.D.L., Luszniak, D., Von der Geest, 2002. Preliminary examination ofcontaminant loadings in farmed salmon, wild salmon and commercial salmonfeed. Chemosphere 46, 1053–1074.

FAO, 2006. The state of food insecurity in the world, Food and AgricultureOrganization (FAO) Rome, Italy. <http://www.fao.org>.

Fielding, A.H., 2007. Cluster and classification tecniques for the biosciences.Cambridge University Press, Cambridge. GB.

Goncalves, C., Esteves de Silva, C.G., Alpendurada, M.F., 2006. Chemometricinterpretation of pesticide occurrence in soil samples from an intensivehorticulture area in north Portugal. Analytica Chimica Acta 560, 164–171.

Hale, R.C., La Guardia, M.J., Harvey, E.P., Mainor, T.M., Duff, W.H., Gaylor, M.O., 2001.Polybrominated diphenyl ether flame retardants in Virginia freshwater fishes.Environmental Sciences and Technology 35, 4585–4591.

Hites, R., Foran, J.A., Carpenter, D.O., Hamilton, M.C., Knuth, B.A., Schwager, S.J.,2004. Global assessment of Organic Contaminants in Farmed Salmon. Science 9,303 (5655), 226–229.

Kim, M.K., Kim, S., Yun, J.S., Kwon, J.W., Son, W.S., 2007. Evaluation of PCDD/Fscharacterization in animal feed and feed additives. Chemosphere 69, 381–386.

Miniero, R., Ingelido, A.M., Iamiceli, A.L., Ferri, F., Iacovella, N., Fulgenzi, A.R., DeFelip, E., di Domenico, A., 2007. High concern chemicals in top layer sedimentsof the northern Adriatic seabed as markers of old waste dumpings.Chemosphere 68, 1788–1797.

UNEP, 2001. The Stockholm Convention. <http://www.unep.org>.Van den Berg, M., Birnbaum, L., Bosveld, A.T., Brunstrom, B., Cook, P., Feeley, M.,

Giesy, J.P., Hanberg, A., Hasegawa, R., Kennedy, S.W., Kubiak, T., Larsen, J.C., vanLeeuwen, F.X., Liem, A.K., Nolt, C., Peterson, R.E., Poellinger, L., Safe, S., Schrenk,D., Tillitt, D., Tysklind, M., Younes, M., Waern, F., Zacharewski, T., 1998. Toxicequivalency factors (TEFs) for PCBs, PCDDs, PCDFs for humans and wildlife.Environmental Health and Perspectives 106, 775–792.

Zhou, F., Guo, H., Liu, Y., Jiang, Y., 2007. Chemometrics data analysis of marine waterquality and source identification in Southern Hong Kong. Marine PollutionBulletin 54, 745–756.