Prediction of ecotoxicity of hydrocarbon-contaminated soils using physicochemical parameters^†

Soil collection and characterization

The eight hydrocarbon-contaminated soils and associated control soils were collected from different field locations across the United States. All the soils were contaminated with crude oil except one, which contained lube oil (soil 18). One soil was also contaminated with brine (soil 8), and another was also contaminated with produced water (soil 14). The metals concentrations found in the soils were within the range of naturally occurring concentrations reported by Deuel and Holliday [10] except soil 9. The spill age ranged from 3 months to more than 5 years at the time of collection.

Physical and chemical characterization of the contaminated and control soils was conducted (Table 1). The analytical methods used in these characterizations were described by Saterbak et al. [9]. Briefly, physical parameters, such as pH, electrical conductance, porosity, bulk and particle densities, total organic carbon, and soil texture (% sand, clay, and silt), were determined along with trace metal analyses. Hydrocarbon contamination was characterized using TPH, oil and grease, and GC boiling-point distribution. Total petroleum hydrocarbons were determined both gravimetrically and by GC (tphgc). Freonextractable and carbon disulfide-extractable gravimetric oil and grease were determined. The carbon disulfide extracts were also used to speciate hydrocarbons by GC into the following classes: total saturates, consisting of n- and iso-saturates and ring-saturates; total aromatics; polars; and asphaltenes. The TPH and oil and grease measurements were considered as crude measurements in this study and the speciated hydrocarbon classes represented fractionated components of the oil and grease carbon disulfide extracts. Sixteen target polyaromatic hydrocarbons and benzene, toluene, ethylbenzene, and xylenes (BTEX) were also quantified.

Toxicity test endpoints

Toxicity tests with earthworms (Eisenia fetida) and four plant species, corn (Zea mays variety “Trucker's Favorite”), lettuce (Lactuca sativa variety “Great Lakes”), mustard (Brassica rapa), and wheat (Triticum aestivum variety “Taylor”), were conducted. No-observed-effect concentrations (NOECs) were estimated for earthworm avoidance, earthworm survival (in 7-d, 14-d, and chronic tests), earthworm reproduction, seed germination, and root growth. Concentrations that caused 25% inhibition (IC25) in earthworm survival and reproduction, seed germination, and root growth were also calculated if the concentration response data allowed (Table 2).

Statistical analyses

Correlations between physicochemical parameters and toxicity. Correlation analysis, using Pearson product-moment correlation coefficients, was used to explore possible linear relationships between physicochemical parameters and toxicity endpoints. Additionally, correlations between the physical and chemical variables were examined. Highly correlated (p ≤ 0.05) physical and chemical parameters were identified along with statistically significant soil property-toxicity relationships. Screening the physicochemical parameters was necessary because the number of observations (soils tested) was small relative to the number of physicochemical measurements. The relationships between the chemical parameters were used to help select variables for univariate model development.

Univariate models for toxicity prediction. Only hydrocarbon measurements were used as predictor variables. The physical parameters and metal concentrations were generally similar between the contaminated and control soils and were not used in model development. Using only hydrocarbon parameters as predictors also helped to reduce the number of variables.

A series of steps was used to build, if possible, a common prediction model for a group of toxicity endpoints (earthworm avoidance and survival, worm reproduction, seed germination, and root growth). For each toxicity endpoint (y), candidate predictors (x's) were selected from some of the chemical parameters that were correlated and from some that were not correlated. The maximum number of x's selected was limited by the number of soils tested, which ranged from five to eight. In stepwise regression with forward selection, the significance levels for variable entry and removal from the model were p = 0.05. To avoid overfitting the model, a minimum of 3 df for the model residuals was used. Model assumptions of normality and randomness were assessed using the residuals of the fitted model. Subsequently, the common predictor variables within each group of toxicity endpoints were identified. For each predictor variable, the slopes and intercepts of the individual regression models were then tested using the SAS general linear models procedure [11] to determine whether they were significantly different (p ≤ 0.05) from each other. If the slopes and intercepts were not significantly different, a regression with common slope and intercept, y = α₀ + βx, was obtained. If only intercepts were significantly different, a regression with common slope and different intercepts, y = α_i + βx, where α_i = different intercepts for different toxicity endpoints, was obtained. Prediction error of the models was calculated as the square root of the mean square error from analysis of variance. Correlation analysis and univariate model development were done using STATGRAPHICS Plus software [12].

Cluster analysis to group soils. Cluster analyses were used to group soils based on physicochemical parameters. The analyses were done using the average, complete, or single linkage methods in SAS clustering procedures [11]. Single linkage minimizes the distance between points, whereas complete linkage maximizes these distances. Average linkage, on the other hand, determines clustering on the basis of averaged distances between clusters [13]. The criteria used in determining the number of clusters were coefficient of determination (R²) and graphical separation using the root-mean-squared distance for the average clustering method, maximum distance for the complete clustering method, and minimum distance for the single clustering method. Analyses were performed on both the original values and those that had been standardized for mean and variance. Standardization was used to decrease variability between the parameters because the range of concentrations within each parameter differed.

The physical and chemical parameters were analyzed both separately and jointly for the contaminated soils. Certain physical parameters and metals, such as bulk and particle densities, silver, cadmium, and mercury, were not included in the analysis because the concentrations of these variables were essentially the same across all soils. All chemical parameters and subsets of these parameters were used in separate analyses. Subsets included BTEX and the crude hydrocarbon measurements (e.g., gravimetric TPH) and BTEX with the fractionated components (e.g., polars).

Principal component analysis to group soils. Principal component analysis was used to reduce the number of physicochemical parameters and toxicity endpoints (X) to a smaller set of representative variables or principal components (PCs). The PCs were linear combinations of the variables that explained the variability in the Xs, where the eigenvectors in this case were the generated loadings of the PCs, which displayed the influence of the variables. The influential Xs were those with higher coefficients in its eigenvectors and were the main contributors to the corresponding PCs. Principal component analysis was performed on the same groups of parameters that were used in cluster analyses. The variables were standardized to unit variance before analysis so that an equal weight was given to each variable. The fraction of the variability in the Xs explained by a PC (R²X) was used to determine the number of significant components. Graphical representation of the PCA parameters for scores and loadings provided information about similarities of the soils and correlation structures of the variables, respectively.

Partial least squares models for toxicity prediction. The method of partial least squares projections to latent structures was applied to the data to explore quantitative relationships between multivariate toxicity endpoints (Y) and chemical parameters (X). Partial least squares is particularly useful in analyzing multicollinear data and in cases where the number of variables exceeds the number of observations [14]. All 11 chemical parameters were used as X variables, and the toxicity endpoints were separated into groups as in the univariate model development (i.e., worm avoidance and survival, worm reproduction, seed germination, and root growth). R²X and R²Y were used to evaluate how well the models simultaneously explained the variability in the X and Y variables, respectively. The normality assumption was assessed using the Y residuals. The cross-validation statistic, Q², provided an estimate of the predictive power of the PLS models. The statistical significance of a component was based on Q² exceeding the cross-validation threshold for that component. X variables with the greatest influence in explaining variability in Y, or variable importance in projection, were determined by the criterion of a value greater than 1. The PLS scores plots portrayed the relationships between X and Y, expressed by the latent variables t and u, for each soil. The loadings plots, on the other hand, showed the interrelationships between these variables within each component. Principal component analysis and PLS analyses were conducted using SIMCA® software [15].

Soil groupings

Using cluster and PCA analyses, soils were separated into groups on the basis of physicochemical parameters and toxicity endpoints. In general, soils were grouped similarly by cluster and PCA. Based on visual inspection, the soils tended to separate into more groups in the PCA scores plots than in cluster analysis. Only the results from PCA are presented (Table 3). Using physical parameters, soil 9 was different from the other soils on the first PC, whereas soil 2 was different on the second PC. The soil properties that distinguished soil 9 were electrical conductance, total organic carbon, porosity, arsenic, barium, chromium, and zinc (PC1), whereas the variables that distinguished soil 2 were pH and soil texture.

For chemical parameters, PCA separated soils 9 and 17, and soil 4 to a lesser extent, from the other soils when either crude or fractionated hydrocarbon parameters were used. The crude and fractionated hydrocarbon parameters, except n- and iso-saturates, in the first PCs, separated soil 9. For soil 17, BTEX was the discriminating variable; this variable also dominated PC2. The remaining five soils appeared to cluster together in the PCA scores plot, with soil 2 exhibiting more difference than the other four soils. When both physical and crude chemical parameters were used, the soil groupings were more similar to groupings based on physical parameters in that soils 4 and 17 were not differentiated.

The PCA scores plot showed that soils 8 and 18 were similar on the basis of earthworm toxicity (Fig. 1A). These two soils tended to be less toxic than average (higher than average NOEC) for worm avoidance and for 14-d acute and chronic survival (PC1). Soils 2 and 17 were observed to be in opposite regions of the scores plot. Soil 2 tended to be less toxic than average, whereas soil 17 was more toxic than average for all endpoints (both PCs). Soil 4 was more similar in toxicity to soil 17, whereas soil 6 was more similar to soil 2. The worm avoidance and 14-d acute and chronic survival endpoints contributed most to PC1, and the juvenile production endpoints dominated PC2 (Fig. 1B).

Based on plant responses, soil 18, and soil 8 to a lesser extent, were separated from the other soils by PCA (Fig. 2A). Soil 18 was different from the other soils and tended to be less toxic than average for all plant endpoints. Of the five remaining soils, soil 9 was the most toxic, and soils 4 and 17 had similar toxicities. Most of the endpoints contributed approximately equally to PC1 (Fig. 2B). On the other hand, the endpoints formed two groups in PC2, and the variation between these two groups was mostly captured by PC2. Mustard germination was differentiated from corn and wheat in PC2 and was more sensitive. Mustard and corn root growth were also differentiated from wheat.

Interrelationships between soil properties and toxicity

Relationships between soil properties and between soil properties and toxicity were discernible from correlation analysis and PCA loadings plots. From correlation analysis, most chemical parameters were significantly correlated with each other (r ≥ 0.70, p ≤ 0.05) with the exception of BTEX and n- and iso-saturates. The BTEX compounds were not highly correlated (r < 0.50) to any other chemical parameters. The n- and iso-saturates were only significantly correlated to total saturates. A few physical parameters and metals also correlated to chemical parameters. Electrical conductance and total organic carbon were positively correlated (r ≥ 0.70) to all hydrocarbon parameters except BTEX and n- and iso-saturates. Barium and zinc were the only metals that showed significant correlations to some of the chemical parameters, such as polars and ring saturates. Principal component analysis loadings plots showed similar associations between the physical and chemical properties.

Earthworm avoidance and survival were correlated only to chemical parameters, whereas reproduction was correlated primarily to physical parameters. Significant negative correlations between worm avoidance and survival and TPH (as measured by GC), carbon disulfide-extractable oil and grease, polars, and n- and iso-saturates were found, whereas cocoon and juvenile production showed significant positive correlations to clay, arsenic, barium, chromium, and lead. Among the plant endpoints, only corn and wheat germination NOECs and wheat root growth NOEC showed significant correlations to physical and chemical parameters. Germination was negatively correlated to most hydrocarbon parameters, whereas root growth was less well correlated.

The relationships between highly correlated physicochemical parameters and biological endpoints were further examined using scatter plots. In many cases, the highly correlated relationships were due to data falling into two clusters, with endpoints reflecting either very low or very high toxicities at low and high concentrations of the physicochemical parameter, respectively. Because these relationships were based on limited data (six to eight soils) and were further limited by the use of NOEC endpoints from widely spaced test concentrations, they are not very surprising, and more data are needed to fill the gaps. Nevertheless, these relationships suggest possible threshold concentrations above which toxic effects are manifested.

Prediction models development

Univariate prediction models. Predictive models for four sets of toxicity endpoints (i.e., earthworm avoidance and survival, earthworm reproduction, seed germination, and root growth) have been constructed (Table 4). Total petroleum hydrocarbons measured by GC and polars were common predictors for earthworm avoidance and survival and seed germination. These hydrocarbon parameters appeared to have greater effect on earthworms than seed germination, as indicated by greater negative slopes in the earthworm models. The earthworm models have common slopes and intercepts and explained 65% of the variation seen in the five avoidance and survival endpoints. Common slopes and intercepts in the tphgc and polars models suggest that each of these parameters have similar effects on avoidance and acute and chronic survival of earthworms. The models for earthworm avoidance and survival using TPH-GC and polars as predictors are

(1)

(2)

Gravimetric TPH and n- and iso-saturates were identified as the “best” predictors of earthworm reproduction (Table 4). However, these variables accounted for only 20 to 40% of the variability observed. Consequently, the models have low predictability, with estimated prediction errors of approx. 60% for gravimetric TPH and 80% for n- and iso-saturates (relative to intercepts).

In addition to tphgc and polars, asphaltenes were also identified as predictors of seed germination (Table 4 and Fig. 3). The germination models, in contrast to the earthworm survival regressions, have common slopes but different intercepts (Fig. 3). The germination models using tphgc, polars, and asphaltenes as predictors are

(3)

where α = 98.54 for corn, 27.90 for mustard germination NOEC, 34.54 for mustard germination IC25, and 86.04 for wheat.

(4)

where α = 97.63 for corn, 26.91 for mustard germination NOEC, 33.56 for mustard germination IC25, and 85.13 for wheat.

(5)

where α = 104.93 for corn, 34.81 for mustard germination NOEC, 41.45 for mustard germination IC25, and 92.43 for wheat.

Different intercepts within each model suggest differences in interspecies sensitivity to the predictors, with mustard germination being more sensitive than corn and wheat. Furthermore, the intercepts in the asphaltenes, polars, and tphgc models were similar for each species, suggesting that these hydrocarbon parameters were affecting germination in a similar manner. The models explained 75 to 85% of the variation observed in the four germination endpoints for the three plant species. The estimated prediction error was 20 to 25% (relative to intercepts) for corn and wheat and 45 or 75% for mustard (depending on whether it was based on the NOEC or IC25 endpoint, respectively).

Asphaltenes did not appear to predict root growth as well as germination. The model for all three plant species explained only 43% of variability and has an associated prediction error of 40%.

Multivariate prediction models. Partial least squares models that relate worm avoidance and survival and seed germination to soil chemical parameters have been developed. Multivariate models relating worm reproduction and root growth to chemical parameters were less definitive. The first component in the worm reproduction and root growth models was not statistically significant, and the prediction power of the models was very poor (≤5%).

For worm avoidance and survival, there were three statistically significant components. However, the second component did not substantially increase the predictive power of the model (42 vs 48%). The third component did increase the predictive power to 61%, but the most influential X variables in the third component were also in the second component. The additional variable present in the second component that was not present in the first component was BTEX. Therefore, on the basis of these factors, a one-component model was adopted. The one-component model accounted for 69% of the variability in Xs and 63% of the variability in Ys. The model explained 73% of the avoidance data, 78% of the 14-d survival data, 70% of the chronic survival data, and 18% of the 7-d survival data. The predictive power of the model was, likewise, poor for 7-d survival but 60 to 70% predictive of the other endpoints. The most relevant chemical parameters (variable importance in projection > 1) in explaining the worm avoidance and survival endpoints were, in descending order, tphgc, polars, n- and iso-saturates, carbon disulfide-extractable oil and grease, and total saturates (Table 5), because these parameters loaded most heavily into the first component (Fig. 4B). The negative regression coefficient for each chemical variable indicated an inverse relationship between chemical concentration and toxicity endpoint. The coefficients for each chemical parameter were similar for avoidance, 14-d survival, and chronic survival, suggesting that each chemical parameter affected these endpoints in a similar manner. Likewise, for each toxicity endpoint, the coefficients for the five most influential chemical variables were not substantially different, suggesting that any of these five parameters can be used equally well for toxicity prediction. Within the one-component PLS model, a good linear relationship (r = 0.90, p < 0.01) exists between the scores from the chemical parameters and worm avoidance and survival for the seven soils studied, as seen in the scores plot, where X and Y are expressed as the latent variables t and u (Fig. 4A).

For seed germination, with 11 X variables and four Y variables, a one-component model explained 77% of the variability in Xs and 42% of the variability in Ys. A second component, which was not statistically significant, accounted for an additional 27% of the variability in Ys but only another 7% of the variability in Xs. For each Y in the one-component model, R² ranged from 9 to 73%, with the model explaining much of variability observed for corn and wheat germination but not for mustard germination. The overall predictive power of the one-component model was 29% for all germination endpoints, with 44 and 54% predictability for corn and wheat germination, respectively. The chemical parameters (Xs) that were most relevant in explaining the germination endpoints were, in descending order, asphaltenes, polars, tphgc, carbon disulfide-extractable oil and grease, total aromatics, and ring saturates (Table 6 and Fig. 5B). The coefficient for each of these chemical parameters indicated an inverse relationship with the germination endpoints. Because the corn and wheat endpoints contributed more to the first component than the mustard endpoints (Fig. 5B), the coefficients reflected greater effect of these chemical parameters on corn and wheat germination than mustard germination (Table 6). Corn and wheat germination was similarly affected by each chemical variable. As seen with worm avoidance and survival, the coefficients for the most influential chemical variables did not differ greatly, indicating that any of the six parameters can be used to predict toxicity to corn and wheat germination. The scores values of the chemical parameters and germination endpoints showed a good linear relationship (r = 0.83, p = 0.01) for the eight soils within the one-component PLS model (Fig. 5A).