Modeling Relationships between Baltic Sea Herring (Clupea

Nov 4, 2010 - Modeling Relationships between Baltic Sea Herring (Clupea harengus) Biology and Contaminant Concentrations Using Multivariate Data Analy...
2 downloads 8 Views 296KB Size
Download PDF
Environ. Sci. Technol. 2010, 44, 9018–9023

Modeling Relationships between Baltic Sea Herring (Clupea harengus) Biology and Contaminant Concentrations Using Multivariate Data Analysis K A T R I N L U N D S T E D T - E N K E L , * ,†,‡ RICKARD BJERSELIUS,§ LILLEMOR ASPLUND,| KERSTIN NYLUND, YANG LIU,† AND MATHIAS SÖDERVALL† Environmental Toxicology, Department of Organismal Biology, Evolutionary Biology Centre, Uppsala University, Norbyva¨gen 18A, SE-752 36, Sweden, Department of Contaminant Research, Swedish Museum of Natural History, P.O. Box 50007, SE-104 05 Stockholm, Sweden, Swedish National Food Administration, Box 622, SE-751 26 Uppsala, Sweden, and Department of Applied Environmental Science, Stockholm University, SE-106 91 Stockholm, Sweden

Received July 29, 2010. Revised manuscript received October 13, 2010. Accepted October 19, 2010.

Baltic Sea herring (Clupea harengus) is a pelagic, zooplanktivorous fish and young (2-5 years old) individuals of this species are sampled annually in the Swedish marine monitoring program. This study determined concentrations of organochlorines (OCs) and brominated flame retardants (BFRs) in dorsal muscle from herring (n ) 60) of varying age (2-13 years), weight (25-200 g), and body length (16-29 cm) caught at three locations in the Swedish part of the Baltic Proper. In order to ensure that the fish biology was as varied as possible, though still similar from all sampling sites, the fish to be chemically analyzed were selected from a large number of fish with determined biology using Multivariate Design. In statistical evaluation of the data, univariate and multivariate data analysis techniques, e.g. principal components analysis (PCA), partial least-squares regression (PLS), and orthogonal PLS (OPLS), were used. The results showed that the fish are exposed to a cocktail of contaminants and levels are presented. Significant OPLS models were found for all biological variables versus concentrations of OCs and BFRs, showing that fish biology covaries with fish contaminant concentrations. Correlation coefficients were as high as 0.98 for e.g. βHCH concentration (wet weight) versus the lipid content. Lastly, the OC concentrations in herring muscle were modeled against the BFR concentrations to determine whether concentrations of either could be used to predict the other. It was found that OPLS models allowed BFR concentrations to be predicted from OC * Corresponding author phone: +46 18 471 64 98; fax: +46 18 471 64 25; e-mail: [email protected]. † Uppsala University. ‡ Swedish Museum of Natural History. § Swedish National Food Administration. | Stockholm University. 9018

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 44, NO. 23, 2010

concentrations with high, but varying, accuracy (R2Ys between 0.93 to 0.75). Thus, fish biology and contaminant concentrations are interwoven, and fish biological parameters can be used tocalculate(predict)contaminantconcentrations.Itisalsopossible to predict the BFR concentrations in an individual fish from its concentrations of OCs with very high accuracy.

Introduction Herring (Clupea harengus) is the most dominant commercial fish species in the Baltic Sea as well as being an important food source for several predators in the marine environment. It is a pelagic, zoo-planktivorous fish and young (2-5 years old) herring from the Baltic Sea have been analyzed for their content of numerous contaminants within the Swedish marine monitoring program since 1972 (1). As herring is pelagic, it is thought that the fish from the Baltic Proper contaminant pattern reflects the background occurrence of the contaminants and not local point sources (1). In a specific ecosystem, contaminants move between abiotic and biotic compartments and form a ‘contaminant web’. In this web, fluctuations in contaminant concentrations are caused by variations in abiotic factors such as temperature and availability of nutrients (2, 3) as well as in biotic factors such as species composition, age, and health status of the participating organisms (4). Overall, this leads to variations in contaminant concentrations in any given organism at any given time (5). When modeling contaminant levels, transport, and biomagnification in an ecosystem, thorough knowledge is needed of variations in concentrations in the different compartments (abiotic as well as biotic). The primary aim of the present study was to determine the concentrations of organochlorines (OCs) and brominated flame retardants (BFRs) in Baltic Sea herring of varying biology, e.g. animal age, weight, and length. Our second aim was to analyze the relationships between herring biological characteristics (hereafter referred to as biological variables) such as age and lipid content and residual levels of OCs and BFRs. For instance, it is known that some persistent contaminants bioaccumulate with increasing animal age (5-7), but not much is known about how other biological variables influence contaminant accumulation (8). In earlier work it was shown that concentrations of OCs and BFRs in the muscle of guillemot (Uria aalge) and in guillemot eggs covaried, forming two groups that each included OCs and BFRs, and that contaminants within the respective group covaried strongly (9, 10). Those studies showed that concentrations of OCs in a specific sample could be used to calculate/predict (hereafter referred to as ‘calculate’) the concentrations of BFRs in that sample with varying accuracy (R2Y between 0.92 for BDE99 and 0.50 for HBCD). The third aim of the present work was to investigate whether the concentrations of OCs and BFRs in herring muscle covaried with each other in a similar way, so that concentrations of OCs could be used to calculate concentrations of BFRs and vice versa. In order to get samples that spanned numerous biological variables, i.e. fish covering a broad span of ages, body weights, body lengths, etc., a large number of fish were caught initially and their biology determined. From this large number, multivariate design (MVD) (11-13) was used to select individuals for further chemical analyses of their contaminant concentrations. In total, 60 individuals from three sites in the Baltic Sea were selected from 381 individuals for which biological variables were determined. By using MVD we also ensured that the fish from the three different sampling sites 10.1021/es102448b

 2010 American Chemical Society

Published on Web 11/04/2010

were as similar as possible regarding their biology. As far as we know, this represents a novel use of the MVD method for sample selection in environmental research. In addition, as many variables regarding fish biology and concentrations of contaminants are highly dependent, we addressed our three main aims by applying multivariate data analysis techniques to extract additional and essential information from our data that could not be obtained otherwise. We term this way of working ‘computational environmental toxicology’.

Materials and Methods Fish. Herring (Clupea harengus membras L.) were collected from fish catches at three sampling sites in the Baltic Sea during December 2000 (see the Supporting Information (SI) for a map and coordinates). These three sites were all in the Baltic Proper: Landsort (L) situated in the northern part, Gotland (G) in the central part, and Utla¨ngan (U) in the southwestern part. The fish were collected at the same time as fish intended for the Swedish national marine monitoring program, and all collection and sample preparation were carried out and recorded in a standardized manner (1). For the present work, immediately after landing a large number of fish (approximately 300 per site) were individually deepfrozen in polythene bags, and all subsequent work was carried out on this deep-frozen material. Fish biological variables were measured as described previously (1). These included the following: body weight (BW) to the nearest 0.1 g, body length including tail fin (BL) to the nearest mm, age (in years), condition factor (Cond F ) BW (g) * BL-3 (cm) * 100), sex and maturity of the gonads, hereafter referred to as ‘reproductive phase’ (Rep Ph). The ages of the fish was determined in two ways, from scales and/or from otoliths (see the SI). The reproductive phase, i.e. state of gonadal maturity (scale 1-5), was determined by ocular inspection of the gonads through a small incision on the fish abdomen (see the SI). Only fish with a clearly determined sex and Repr Ph were included in the MVD selection. No fish showed Repr Ph 1 or 5, only 2, 3, and 4. With the use of MVD (11-13) and including the biological variables BW, BL and Age, a total of 60 female herring were selected (20 fish per site) for further chemical analyses from the 381 individuals with determined biology. The MVD was based on a principal component analysis (PCA) that included fish biological variables. The selection of individuals was made from the PCA score plot (with the 381observations) in such a way that the original score plot surface was covered as closely as possible by the 60 observations (the selected fish). The selection process also sought to ensure that the property domains for the biology (the score plot) were as similar as possible for all three sampling sites, resulting in 20 fish per sampling site. This meant that the biology of the selected fish from each site reflected the full diversity of the original large sample of fish, while also being as similar as possible regarding their biology (see the SI for more on MVD). Chemical Analysis. Ten grams of frozen dorsal muscle from each selected fish was placed in glass jars (prerinsed with ethanol/acetone) and stored at -20 °C pending chemical analysis. The chemical analyses of the OCs were carried out at the Special Analytical Laboratory (RSL), Swedish Museum of Natural History, Stockholm, Sweden, in cooperation with the Department of Applied Environmental Science (ITM) at Stockholm University, which performed the BFR analyses. All samples were homogenized, the lipids were extracted, and the lipid content (F%) was determined gravimetrically at ITM. The determination of OCs was then performed at RSL and the determination of BFRs at ITM. Sample treatment and analytical method for quantification of individual OCs and BFRs isomers/congeners as well as laboratory QA/QC procedures have been previously described in detail for OCs (1, 14-16) and BFRs (1, 9, 17). In short, the OCs were

determined by high resolution gas chromatography (GC) using an electron capture (EC) detector and the BFRs by GC connected to a mass spectrometer operating in the EC/ negative ion mode. By using a combination of different capillary columns, it was possible to separate all compounds except for CB138, where the interfering peak caused by CB163 meant that CB138 could be overestimated by ∼20-30%. For the sake of simplicity, only CB138 is mentioned in the remainder of this paper. One sample was lost during the OC determinations, so n ) 59 for the OCs and n ) 60 for the BFRs. The concentrations are expressed as ng/g lipid weight (lw) and wet weight (ww) (see the SI for more regarding the chemical analyses). ΣDDT was calculated as the sum of p,p′DDT, p,p′DDE, and p,p′DDD concentrations, ΣPCB as the sum of ICES 7 marker PCBs: CB28, CB52, CB101, CB118, CB138, CB153, and CB180 concentrations; ΣHCH as the sum of RHCH, βHCH, and γHCH concentrations; and ΣPBDE as the sum of BDE47, BDE99, BDE100, BDE153, and BDE153 concentrations. Concentrations below Level of Quantification. For the congener CB52, 29 of the 60 individuals tested had nonquantifiable concentrations of CB52. When performing the data analyses, all 23 contaminants analyzed were included, using the individual LOQ (upper bound method) for the fish with nonquantifiable concentrations of CB52 (see the SI for more regarding LOQ). Statistics. For univariate and bivariate statistics, the software GraphPad Prism 5.02 (18) was used. This included column statistics (e.g., mean ( standard deviation (SD), geometric mean (GM) and 95% confidence interval (95% CI) of GM), nonlinear regression, and Kruskal-Wallis nonparametric ANOVA test. As most of the contaminant data deviated significantly from the normal distribution according to the Kolmogorow-Smirnov test, Spearman’s correlation analysis was used, for instance to correlate the chemical contaminants with the biological variables. The significance level was set to p < 0.05 for all tests. For the multivariate data analyses (MVDA), principal component analysis (PCA), partial least-squares projection to latent structures (PLS), and orthogonal PLS (OPLS) the software SIMCA-P 12.0.1 (19) were used. For all MVDA a significance level of p < 0.05 was used, and data were centered and scaled (to variance 1) prior to modeling (20). For each model an R2 and Q2 value were calculated, where 2 R shows the dispersion of the data from the model and Q2 shows the structure (stability) in the data. The cross-validation term Q2 is calculated by excluding 20% of the data in crossvalidation rounds (21) and an R2 value of >0.7 and a Q2 value of >0.4 denote an acceptable model when analyzing biological data (22). PLS was used here to determine whether there were significant relationships between the biological variables and contaminant concentrations and to model the covariation of OCs and BFRs. OPLS was used to model the biological variables one at a time versus the multivariate contaminant concentrations in order to explore correlated variation and uncorrelated variation, which from a biological viewpoint may be of equal importance. Furthermore, OPLS was used to model the BFR concentrations one at a time versus the multivariate OC concentrations (see the SI for more regarding MVDA).

Results and Discussion Biological Variables. The use of MVD in selection of fish from the three sampling sites was successful, since a PCA that included the five biological variables known before the chemical analysis (BW, BL, Age, Rep Ph, and Cond F) showed that there were no differences between the L, G, and U fish in their biology (for details regarding the biological data, see Table S1, SI, for additional information regarding the MVD VOL. 44, NO. 23, 2010 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

9019

see p 3, the SI). Statistical analysis of the variables used for the MVD sample selection confirmed this; there were no significant differences in fish biology between the three sampling sites (Kruskal-Wallis nonparametric ANOVA test, p-values between 0.3-0.4). However, for F%, which could not be included in the MVD as it was not determined until the chemical analysis, the fish from G were significantly fatter (higher F%) than the fish from U (p < 0.001, Kruskal-Wallis test followed by Dunn’s Multiple Comparison Test). When the concentrations were modeled as lw instead of as ww, the fatty fish from G showed lower contaminant concentrations, while the lean fish from U showed very high concentrations. We believe that this is due to the ‘lipid dilution effect’ (9, 10). Because of these differences in F% between the sampling sites, for all the modeling and comparisons regarding the contaminant concentrations, the concentrations in ww were used. When the data generated are multivariate, i.e. when three or more variables are determined per observation, it is valuable to get an initial general overview of how the different observations and variables are related (20). In the present study using the six biological variables, the PCA obtained (R2 ) 0.57 and Q2 ) 0.39, one significant component) revealed that Repr Ph had no importance for the model (R2 ) 0.10 and Q2 ) -0.002, data not shown). Excluding Rep Ph gave a new improved PCA model (R2 ) 0.87 and Q2 ) 0.59, two significant components) (Figure S3, SI). The model overview showed that e.g. BW, BL, and Age had high R2 and Q2 values, meaning that these variables covaried, so that their respective values could be calculated fairly correctly from the other variables. However, the Cond F and F% showed lower Q2 values, indicating that these two variables varied more independently than the other biological variables and were thus more difficult to calculate. The PCA showed two groups of highly correlated variables, one with BW, BL, and Age and the other with Cond F and F%. Because of this, the following correlations were performed. Correlations Biological Variables. 1) BW vs BL. The correlation between BW (g) and BL (cm) followed the exponential equation Y ) a*exp(b*X)

(1)

where Y ) BW (g), X ) BL (cm), a ) 3.193 (2.542-3.844 as 95% CI), and b ) 0.1430 (0.1347-0.1512 as 95% CI). The correlation for the relationship between BW and BL showed a very high R2 of 0.96 (Table S2, Figure S4, SI). When the natural logarithmic BW and BL were used instead the relationship was linear, with the correlation between BW (as Y, in g) and BL (as X, in mm) modeled as ln(Y) ) a + bln(X)

(2)

where a ) -13.60 ( 0.4837, b ) 3.326 ( 0.09094, and p < 0.0001. Cardinale and Arrhenius (23) used eq 2 for a very large number of herring (n ) 3500 to 15000) from several sampling sites in the Baltic Sea over 11 years (1986-1996) and found a decreasing weight-at-age for the herring during that period. The differences between our results and those shown in that paper are interesting. First, we found a slightly lower R2 (0.96) than reported by them (0.96-0.98), most probably because of the lower number of samples used here (n ) 60). Second, our finding of a higher b (slope) of 3.33 ( 0.091 indicated that the growth rate had increased significantly compared with the growth rate reported by them. For fish caught in ICES areas 25, 27, and 28 (our U, L, and G, respectively) the lowest b reported in the Cardinale and Arrhenius study was 2.67 ( 0.03 for 1994, and the highest was 3.15 ( 0.02 for 1989 (23). The increase in growth rate can also be examined by calculating the expected weight of a 200 mm herring. Using 9020

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 44, NO. 23, 2010

eq 2, a 200 mm long herring would have a calculated mass of 55.82 g. The corresponding weights calculated for 200 mm herring from L, G, and U areas in the Cardinale and Arrhenius paper were all below 50 g during the latter part of the study period (1986-1996). The increase in growth rate indicated by the higher b in the present study, and the calculated higher weight for the 200 mm herring must of course be validated with more samples from more years before it can be concluded that the weight-at-age has increased. However, it is nonetheless interesting to compare our results with the earlier material, and it would be a great advantage if the same equations were used in all studies. Unfortunately, this is not always the case. 2) BW vs Age and BL vs Age. The correlations between BW and Age and between BL and Age both followed a third order polynomial equation Y ) B0 + B1X + B2X2 + B3X3

(3)

where Y ) BW (g) or BL (cm), X ) Age (years), and B0, B1, B2, and B3 are calculated coefficients. The resulting equations for BW and BL correlations to Age were BW ) 29.61 - 8.319*Age + 2.898*Age2 - 0.1131*Age3

(4)

(R2 ) 0.61) BL ) 15.88 - 0.3164*Age + 0.2233*Age2 - 0.009818 *Age3 (5) (R2 ) 0.64) (see Tables S3 and S4 and Figures S5 and S6, SI). The correlations between BW/BL and Age showed similar patterns in that fish up to 5 years of age had slower weight and length gain than fish 6 years and older, which grew more rapidly and gained weight and length faster (Figures S5 and S6, SI). This indicates a shift in food or feeding habits with age. Data from the literature confirm that the diet of younger fish consists of smaller zooplankton and that the growth spurt shown here coincides with an increasing part of the diet consisting of larger mysids and amphipods (24, 25). 3) Cond F vs F%. The regression between Cond F (as Y) and F% (as X) followed a ‘one site binding’ relationship (hyperbola) that is similar to a Michaelis-Menten equation Cond F ) (Cond Fmax F%)/(Km + F%)

(6)

where Cond Fmax ) 0.8024, F% ) lipid content, Km ) 0.3673, and (R2 ) 0.44). The correlation shows that fish with low F% also have low Cond F and that an increase in F% leads to a corresponding rapid increase in Cond F. However, the relationship flattens out, so that when F% is ∼2% the condition factor reaches a level of ∼0.6-0.8, and further increases in the lipid content do not lead to a corresponding increase in Cond F (Table S5, Figure S7, SI). Cond FMax is 0.80 (95% CI 0.76-0.84). In terms of fish biology, the PCA showed that the selection of individuals was successful and that the biology of fish from the sampling sites overlapped to a great extent, i.e. the fish biology was as similar as possible between the three sampling sites. The PCA also showed that BW, BL, and Age, and F% and Cond F, formed two groups in the PCA loading plot with highly correlating variables. Contaminant Concentrations. The concentrations of OCs and BFRs in young (2-5 years) and old (6 years and older) herring from L, G, and U are presented in Tables S6 (ng/g lw) and S7 (ng/g ww) in the SI. As the present paper deals with the modeling of relationships between e.g. fish biology and contaminant patterns, the levels of contaminants are

not discussed in detail (for more regarding contaminant concentrations see the SI). The contaminant concentrations in L and U fish aged between 2-5 years are presented in great detail by the Swedish marine monitoring program (1), which for herring has been running since 1978 for L (3-5 year old fish) and since 1980 for U (2-4 year old fish). It is worth noting that the contaminant concentrations in the young L and U fish (3-5 years of age) reported in the present work did not differ significantly from those reported for L and U fish of the same age analyzed in the Swedish monitoring program (p ) 0.10-0.82). As the OC concentrations in the present work were analyzed by RSL and the OC concentrations in the monitoring program were (and still are) analyzed by ITM, this means that the two laboratories are in accord and give results that are fully comparable. Furthermore, and most importantly, it means that the selection of fish using MVD did not introduce a bias toward fish with any different contaminant pattern. In terms of fish contaminant concentrations, it can be concluded that the OCs, especially ΣDDT, dominated over the BFRs and that contaminant patterns formed in the fish. These patterns are discussed further below (see also the SI). PCA, Contaminant Concentrations. Including the 17 OCs and six BFRs in herring muscle (ww) gave a PCA model (R2 ) 0.88 and Q2 ) 0.80, three significant components). In the loading plot clear groupings of contaminants could be seen, showing that most contaminants covaried positively to a high degree, although some variation was discernible (Figure S8, SI). In general, a fish with a higher concentration of one contaminant also had higher concentrations of all other contaminants. This was also apparent from the model overview, where most of the contaminants showed high R2 and Q2 values with a mean ((SD) value for the crossvalidation term Q2 of 0.84 ((0.08) (n ) 23). These high R2 and Q2 values indicate that the concentration of one contaminant in a fish can be calculated fairly correctly from the other contaminant concentrations. One contaminant (CB180) was different, with a low Q2 value of 0.05, showing that this PCB congener varied more independently than the others and was thus more difficult to calculate than the variables with higher R2 and Q2 values. OPLS, Fish Biological Variables versus Contaminants. The relationships between the biological variables of herring and the residual levels of OCs and BFRs in the fish dorsal muscle were also analyzed. Separate analyses were performed to model the correlations between biological variables and contaminant concentrations. OPLS regression was employed for these models using one biological variable (as Y) and the concentrations of OCs and BFRs (as X) for each biological variable separately. The results are discussed below, after all the OPLS models are presented. BW as Y and contaminant concentrations (wet weight) as X gave an OPLS model (R2X ) 0.76; R2Y ) 0.25; Q2 ) 0.24, one component) which showed that all contaminants except HCB increased in concentration with an increase in herring body weight (Figure S9A, SI). The highest VIP value (the importance of the variable on the OPLS projection) was shown by BDE100, followed by CB156, BDE154, BDE47, CB153, CB138, DDE, and so on in decreasing order (Table S8, SI), and the lowest by HCB. Further testing (Spearman correlation test) showed that the increase in concentration with increasing BW was significant for all contaminants except HCB (p ) 0.24). BL as Y and contaminant concentrations (in wet weight) as X gave an OPLS model (R2X ) 0.76; R2Y ) 0.24; Q2 ) 0.23, one component) which showed that in general contaminants increased in concentration with an increase in herring length (Figure S9B, SI). As for BW, the highest VIP values were shown by the BDEs, the larger CBs and DDE (Table S9, SI). Further testing (Spearman correlation test) showed that the increase

in concentration was significant for all contaminants except HCB (p ) 0.30), CB28 (p ) 0.09), and p,p′-DDD (p ) 0.08). Age as Y and contaminant concentrations (in wet weight) as X gave an OPLS model (R2X ) 0.75; R2Y ) 0.22; Q2 ) 0.19, one component) which showed that in general contaminants increased in concentration with an increase in herring age (Figure S9C, SI). As for BW and BL, the highest VIP values were shown by the BDEs, the larger CBs and DDE (Table S10, SI). Further testing (Spearman correlation test) showed that the increase in concentration was significant for 15 of the contaminants, though not for eight of them: HCB (p ) 0.52), CB28 (p ) 0.17), p,p′DDD (p ) 0.15), HBCD (p ) 0.10), p,p′DDT (p ) 0.09), CB52 (p ) 0.07), t-NoCl (p ) 0.07), and γHCH (p ) 0.054). F% as Y and contaminant concentrations (in wet weight) as X gave the strongest OPLS model (R2X ) 0.76; R2Y ) 0.75; Q2) 0.74, one component) which showed that all contaminants increased in concentration with an increase in herring lipid content (Figure S9D, SI). Interestingly, the pattern differed compared with BW, BL, and Age in that the highest VIP values were shown by βHCH, γHCH, RHCH, CB52, and CB101 (Table S11, SI). Further testing (Spearman correlation test) showed that the increase was significant (p < 0.001) for all contaminants. Repr Ph as Y and contaminant concentrations (in wet weight) as X gave a very weak OPLS model (R2X ) 0.76; R2Y ) 0.07; Q2 ) 0.05, one component). Cond F as Y and contaminant concentrations (in wet weight) as X gave a slightly stronger OPLS model (R2X ) 0.76; R2Y ) 0.26; Q2 ) 0.23, one component) than the corresponding model with Repr Ph. The two models showed that in general, all the contaminants increased in concentration with an increase in herring Rep Ph and Cond F (Figure 9E and 9F, SI). As for F%, the pattern shown by Repr Ph and Cond F differed from that shown by BW, BL, and Age (Tables S12 and S13, SI). The concentrations of ΣDDT, ΣPCB, and ΣPBDE also increased significantly with all biological variables (p ) 0.010.0001), respectively. The results from the respective OPLS with fish biological variables versus contaminants showed that fish biology covaried to a high degree with fish contaminant concentrations. This meant that most contaminant concentrations could be calculated with high accuracy from the fish biology. As shown, herring BW, BL, and Age covaried strongly, and, accordingly, the OPLS models with BW, BL, and Age as Y were similar in that those biological variables covaried most strongly with the concentrations (ww) of the BDEs and the larger CBs but not as strongly with e.g. HCB, CB28, CB52, and HBCD. These results are in line with those reported by Pikkarainen and Parmanne (26), who found (unfortunately without any statistics) an almost linear age dependence for accumulation of ΣDDT, ΣPCB, and the larger CB-congeners. This relationship could not be seen in that study for CB28 or HCB. Furthermore, the dioxin-like PCBs have been shown to have strong age dependence in Baltic Sea herring, with a R2 of 0.76 when modeling the age dependence of WHOPCBTEQ (8). That study also showed that the higher chlorinated CBs had higher R2 values (fitted better to the model) than the lower-chlorinated CBs. It is clear that for herring exposed to a cocktail of chemicals via food and water, bioaccumulation of some contaminants increases with age but not that of some other contaminants. It would be interesting to examine the effect on contaminant accumulation of the change in diet that occurs at around 20 cm herring BL, when the feed changes from smaller zooplankton to larger nektobenthos (24, 25). However, to our knowledge no chemical data are available on the contaminants in stomach contents from varying age classes. In a similar way F%, Cond F, and Repr Ph also show similarities in the respective OPLS models. The OPLS model VOL. 44, NO. 23, 2010 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

9021

relating herring lipid content (F%) to concentrations of OCs and BFRs showed the highest R2 and Q2 values, allowing individual fish chemical concentrations to be calculated to a high degree of accuracy from F%. For instance, the relationship between βHCH concentration (in wet weight) in all fish (n ) 59) and fish F% showed a correlation coefficient (Spearman’s rho, r) of 0.98 (Figure S13, SI). This indicates that a fatter fish also had a higher contaminant concentration (ww) than a lean fish. The reason behind this could lie in the chemical properties of the compounds that are fat soluble and accumulates in fatty tissue. The higher contaminant concentrations could be because the fatter fish had eaten more (consumed more lipids) and thus accumulated higher amounts of contaminants than a leaner fish. However, a starved fish, e.g. due to illness, has a higher contaminant concentration on a lipid weight basis, due to fat loss and slow excretion of the contaminants. It is therefore very important when discussing this to consider and specify whether the contaminant concentrations are given as lw or as ww. In our results we could see that F% increased significantly with increasing age but that the relationship was not altogether clear due to very large scatter (p ) 0.02, R2 ) 0.09). Interestingly, the environmental Obesogen hypothesis proposed recently suggests that environmental chemicals contribute to the development of obesity and that endocrine disrupting chemicals interfere with e.g. the biology of adipose tissue (27). If this were the case in herring, the reason for the fatter fish showing higher contaminant concentrations would be completely reversed, i.e. the fish would be fatter because of the higher contaminant concentrations. This could of course not be demonstrated in the present work, but it would be a very interesting field to explore in future experimental studies. Importantly, if these results regarding the strong covariance between fish biology and individual fish contaminant load are consistent, chemical concentrations could be calculated from fish biology for other herring samples. This would allow the cost of chemical analyses to be drastically reduced. However, it is imperative that such models are validated at regular intervals, as contaminant concentrations slowly change over time, and of course they cannot be used for new, rapidly emerging contaminants. Calculating BFR Concentrations from OC Concentrations. Lastly, to determine whether OCs and BFRs covaried with each other, the relationships between concentrations of OCs and BFRs in herring were modeled. PLS modeling with concentrations (ww) of BFRs analyzed in herring as Y and OCs as X gave a model (R2X ) 0.84, R2Y ) 0.81, and Q2 ) 0.78, two components) which showed that the OCs and BFRs covaried to a high degree. In order to further investigate these relationships, OPLS models with the respective BFRs as Y and OCs as X were calculated, one OPLS model for each BFR. The R2X, R2Y, and Q2 for the respective models were high, confirming that BFR and OC concentrations were correlated to a very high degree. The OPLS model with BDE47 concentrations as Y and OC concentrations as X (wet weight) gave an OPLS model that was very strong (R2X ) 0.91, R2Y ) 0.92, Q2 ) 0.91, one component) (Figure 1). This model showed that all the OCs increased in concentration with an increase in BDE47 concentration. The highest VIP values were obtained for DDE and the CBs and the lowest for βHCH, γHCH, RHCH, CB28 and HCB (Table 14, SI). For the corresponding OPLS models with the other BFRs the results were similar, although with varying R2 and Q2 values: BDE99 (R2X ) 0.79, R2Y ) 0.75, Q2 ) 0.74); BDE100 (R2X ) 0.92, R2Y ) 0.83, Q2 ) 0.82); BDE153 (R2X ) 0.79, R2Y ) 0.76, Q2 ) 0.75); BDE154 (R2X ) 0.90, R2Y ) 0.88, Q2 ) 0.84); and HBCD (R2X ) 0.92, R2Y ) 0.92, Q2 ) 0.90) (Figure S22, SI). 9022

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 44, NO. 23, 2010

FIGURE 1. Observed (analyzed) BDE47 concentrations (ng/g ww) versus the calculated (predicted) BDE47 concentration in herring (Clupea harengus) from the Baltic Sea, December 2000 (n ) 60). OPLS model with the BDE47 concentration as Y and OC concentrations as X. Individual fish from Landsort (L, black), Gotland (G, red), and Utla¨ngan (U, blue) with the numbers showing the fish age in years. R2Y ) 0.93; Q2 ) 0.91; RMSEE*) 0.14. *RMSEE ) Root Mean Square Error of Estimation. Further testing (Spearman correlation test) showed that all the correlations between single BFRs and single OCs were significant (p < 0.001) (Tables 14-19, SI). However, by arranging the VIP values for the respective OPLS models in descending order, it can be seen that the correlation patterns for the BFRs versus the OCs differ for the different BFRs. The patterns were similar for BDE47, BDE100, and BDE154 in that these were correlated most strongly to DDE and to most of the higher chlorinated CBs but not as strongly to CB28 and CB52 (Tables S14, S16, S18, SI). Another pattern was shown by the group consisting of BDE99, BDE153, and HBCD in that these BFRs were correlated most strongly to HCB, t-NoCl, CB28, CB52, CB101, p,p′DDE, p,p′DDD, and p,p′DDT. The patterns shown by BDE99 and BDE153 were almost identical, while HBCD differed slightly (Tables S15, S17, S19, SI). The contaminant patterns shown here in herring were different from the OC-BFR patterns in guillemot muscle and egg. In those earlier studies, it was also seen that OCs and BFRs covaried to a high degree, with similar R2 values as shown here for herring, although with lower Q2 values. For guillemot, using concentrations of OCs (as X) and BFRs (as Y) in muscle the values for the PLS model were R2X ) 0.94, R2Y ) 0.79, and Q2 ) 0.48 (three components) (9). For the corresponding PLS model but with guillemot egg concentrations of OCs (as X) and BFRs (as Y) the values were R2X ) 0.91, R2Y ) 0.79, and Q2 ) 0.64 (three components) (10). Therefore it was concluded that OC concentrations in an individual bird could be used to calculate its BFR concentrations with high accuracy (9). In the present study, we showed that the OC concentrations in an individual herring could be used to calculate its BFR concentrations, or vice versa, with even higher accuracy. The higher Q2 values found in the present work show that herring display higher consistency in the patterns than guillemot. As already mentioned, the groupings seen in herring were slightly different than the groupings in guillemot. For guillemot, the correlation patterns for BDE47, BDE99, BDE100, and BDE153 were similar and consistent in both egg and muscle and significantly positively correlated with the higher chlorinated PCBs; CB138, CB153, and CB180 (9, 10). Furthermore, in guillemot BDE154 and HBCD formed a group that covaried positively with βHCH, but this was not found in herring. These differences in the groupings were most likely caused by the varying physiochemical properties and structures of the contaminants and their interaction with

animal biology, leading to differences in uptake, distribution, biotransformation and excretion of contaminants in organisms, events that together can lead to bioaccumulation of a chemical, resulting in different patterns between different animal species. If these patterns shown by the covarying contaminants are consistent, i.e. are shown by other herring individuals from the same sampling sites, the cost of chemical analyses could be drastically reduced, as the concentrations of BFRs and OCs could be calculated from fish biology and/ or from the concentration of a few analyzed OCs or BFRs. As discussed above concerning the highly covarying variables regarding fish biology versus contaminant load, such models need to be validated and adjusted (recalibrated) at regular time intervals as the concentrations of BFRs in the Baltic Sea biota change slightly differently over the years than the concentrations of OCs (1).

Acknowledgments Thanks to MISTRA and FORMAS for financing KLE; Swedish Environmental Protection Agency for the monitoring programme; Erik Greyerz and Mats Hjelmberg, Dept. of Contaminant Research, Swedish Museum of Natural History for chemical analyses (EG) and for helping with fish aging (MH); Rolf Gydemo, Gotland County Administrative Board, for acquiring the Gotland herring; Carina Jernberg, Institute of Marine Research, Sweden, for aging via otoliths; Torbjo¨rn Lundstedt, Dept. of Medicinal Chemistry, Uppsala University for constructive discussions; Johan Trygg, Dept. of Chemistry, Umeå University for the OPLS method and more; and to all anonymous fishermen for the herring.

Supporting Information Available Additional material, map showing sampling sites, tables, and figures with results from the data analysis. This material is available free of charge via the Internet at http://pubs.acs.org.

Literature Cited (1) Bignert, A.; Danielsson, S.; Nyberg, E.; Asplund, L.; Eriksson, U.; Berger, U.; Haglund, P. Comments Concerning the National Swedish Contaminant Monitoring Programme in Marine Biota, 2009. http://www.nrm.se/download/ 18.6321786f122df65955f80002587/Marina+programmet+2009.pdf (accessed October 2010). (2) Olsson, M.; Bignert, A.; Eckhe´ll, J.; Jonsson, P. Comparison of temporal trends (1940s-1990s) of DDT and PCB in Baltic sediment and biota in relation to eutrophication. Ambio 2000, 29, 195–201. (3) Bignert, A.; Olsson, M.; Persson, W.; Jensen, S.; Zakrisson, S.; Litze´n, K.; Eriksson, U.; Ha¨ggberg, L.; Alsberg, T. Temporal trends of organochlorines in Northern Europe, 1967-1995. Relation to global fractionation, leakage from sediments and international measures. Environ. Pollut. 1998, 99, 177–198. (4) Bignert, A.; Olsson, M.; de Wit, C.; Litzen, K.; Rappe, C.; Reuthergårdh, L. Biological variation - an important factor to consider in ecotoxicological studies based on environmental samples. Fresenius J. Anal. Chem. 1994, 348, 76–85. (5) Bignert, A.; Go¨thberg, A.; Jensen, S.; Litze´n, K.; Odsjo¨, T.; Olsson, M.; Reutergårdh, L. The need for adequate biological sampling in ecotoxicological investigations: a retrospective study of twenty years pollution monitoring. Sci. Total Environ. 1993, 128, 121– 139. (6) Jensen, S.; Johnels, A. G.; Olsson, M.; Otterlind, G. DDT and PCB in herring and cod from the Baltic, the Kattegat and the Skagerrak. Ambio Spec. Rep. 1972, 1, 71–85. (7) Perttila¨, M.; Tervo, V.; Parmanne, R. Age dependence of the concentrations of harmful substances in Baltic herring (Clupea harengus). Chemosphere 1982, 11, 1019–1026.

(8) Parmanne, R.; Hallikainen, A.; Isosaari, P.; Kiviranta, H.; Koistinen, J.; Laine, O.; Rantakokko, P.; Vuorinen, P. J.; Vartiainen, T. The dependence of organohalogen compound concentrations on herring age and size in the Bothnian Sea, northern Baltic. Mar. Pollut. Bull. 2006, 52, 149–161. (9) Lundstedt-Enkel, K.; Johansson, A. K.; Tysklind, M.; Asplund, ¨ rberg, J. Multivariate data analyses L.; Nylund, K.; Olsson, M.; O of chlorinated and brominated contaminants and biological characteristics in adult guillemot (Uria aalge) from the Baltic Sea. Environ. Sci. Technol. 2005, 39, 8630–8637. (10) Lundstedt-Enkel, K.; Asplund, L.; Nylund, K.; Bignert, A.; ¨ rberg, J. Multivariate data analysis Tysklind, M.; Olsson, M.; O of organochlorines and brominated flame retardants in Baltic Sea guillemot (Uria aalge) egg and muscle. Chemosphere 2006, 65, 1591–1599. (11) Wold, S.; Sjo¨stro¨m, M.; Carlson, R.; Lundstedt, T.; Hellberg, S.; ¨ hman, J. Multivariate design. Skagerberg, B.; Wikstro¨m, C.; O Anal. Chim. Acta 1986, 191, 17. (12) Trygg, J.; Holmes, E.; Lundstedt, T. Chemometrics in metabonomics. J. Proteome Res. 2007, 6, 469–479. (13) Eriksson, L.; Andersson, P. L.; Johansson, E.; Tysklind, M. Multivariate biological profiling and principal toxicity regions of compounds: the PCB case study. J. Chemometr. 2002, 16, 497–509. (14) Jensen, S.; Reutergårdh, L.; Jansson, B. Manual of methods in aquatic environment research. Part 9 - Analyses of metals and organochlorines in fish. Analytical methods for measuring organochlorines and methyl mercury by gas chromatography. FAO Fish. Tech. Pap. 1983, 21–33. (15) Eriksson, U.; Ha¨ggberg, L.; Ka¨rsrud, A.-S.; Litze´n, K.; Asplund, L. Analytical method for determination of chlorinated organic contaminants in biological matrices; 59; ITM, Stockholm University, Sweden, 1997. (16) Lundstedt-Enkel, K.; Tysklind, M.; Trygg, J.; Schu ¨ller, P.; Asplund, L.; Eriksson, U.; Ha¨ggberg, L.; Odsjo¨, T.; Hjelmberg, M.; Olsson, ¨ rberg, J. A statistical resampling method to calculate M.; O biomagnification factors exemplified with organochlorine data from herring (Clupea harengus) muscle and guillemot (Uria aalge) egg from the Baltic sea. Environ. Sci. Technol. 2005, 39, 8395–402. (17) Sellstro¨m, U.; Bignert, A.; Kierkegaard, A.; Ha¨ggberg, L.; de Wit, C. A.; Olsson, M.; Jansson, B. Temporal trend studies on tetraand pentabrominated diphenyl ethers and hexabromocyclododecane in guillemot egg from the Baltic Sea. Environ. Sci. Technol. 2003, 37, 5496–5501. (18) GraphPadPrism. version 5.03; GraphPad Software, La Jolla, CA USA. (19) SIMCA-P+. version 12.0.1; Umetrics, Umeå, Sweden. (20) Wold, S.; Esbensen, K.; Geladi, P. Principal component analysis. Chemom. Intell. Lab. Syst. 1987, 2, 37–52. (21) Wold, S. Cross-Validatory Estimation of the number of components in factor and principal components models. Technometrics 1978, 20, 397–405. (22) Lundstedt, T.; Seifert, E.; Abramo, L.; Thelin, B.; Nystro¨m, Å.; Pettersen, J.; Bergman, R. Experimental design and optimization. Chemom. Intell. Lab. Syst. 1998, 42, 3–40. (23) Cardinale, M.; Arrhenius, F. Decreasing weight-at-age of Atlantic herring (Clupea harengus) from the Baltic Sea between 1986 and 1996: a statistical analysis. ICES J. Mar. Sci. 2000, 57, 882– 893. (24) Arrhenius, F.; Hansson, S. Food consumption of larval, young and adult herring and sprat in the Baltic Sea. Mar. Ecol.: Prog. Ser. 1993, 96, 125–137. (25) Casini, M.; Cardinale, M.; Arrhenius, F. Feeding preferences of herring (Clupea harengus) and sprat (Sprattus sprattus) in the southern Baltic Sea. ICES J. Mar. Sci. 2004, 61, 1267–1277. (26) Pikkarainen, A.-L.; Parmanne, R. Polychlorinated biphenyls and organochlorine pesticides in Baltic herring 1985-2002. Mar. Pollut. Bull. 2006, 52, 1304–1309. (27) Gru ¨ n, F.; Blumberg, B. Endocrine disrupters as obesogens. Mol. Cell. Endocrinol. 2009, 304, 19–29.

ES102448B

VOL. 44, NO. 23, 2010 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

9023