Ralf-Uwe Ebert

External Validation and Prediction Employing the Predictive Squared Correlation Coefficient — Test Set Activity Mean vs Training Set Activity Mean

The external prediction capability of quantitative structure-activity relationship (QSAR) models is often quantified using the predictive squared correlation coefficient, q (2). This index relates the predictive residual sum of squares, PRESS, to the activity sum of squares, SS, without postprocessing of the model output, the latter of which is automatically done when calculating the conventional squared correlation coefficient, r (2). According to the current OECD guidelines, q (2) for external validation should be calculated with SS referring to the training set activity mean. Our present findings including a mathematical proof demonstrate that this approach yields a systematic overestimation of the prediction capability that is triggered by the difference between the training and test set activity means. Example calculations with three regression models and data sets taken from literature show further that for external test sets, q (2) based on the training set activity mean may become even larger than r (2). As a consequence, we suggest to always use the test set activity mean when quantifying the external prediction capability through q (2) and to revise the respective OECD guidance document accordingly. The discussion includes a comparison between r (2) and q (2) value ranges and the q (2) statistics for cross-validation.

A new risk assessment approach for the prioritization of 500 classical and emerging organic microcontaminants as potential river basin specific pollutants under the European Water Framework Directive

Given the huge number of chemicals released into the environment and existing time and budget constraints, there is a need to prioritize chemicals for risk assessment and monitoring in the context of the European Union Water Framework Directive (EU WFD). This study is the first to assess the risk of 500 organic substances based on observations in the four European river basins of the Elbe, Scheldt, Danube and Llobregat. A decision tree is introduced that first classifies chemicals into six categories depending on the information available, which allows water managers to focus on the next steps (e.g. derivation of Environmental Quality Standards (EQS), improvement of analytical methods, etc.). The priority within each category is then evaluated based on two indicators, the Frequency of Exceedance and the Extent of Exceedance of Predicted No-Effect Concentrations (PNECs). These two indictors are based on maximum environmental concentrations (MEC), rather than the commonly used statistically based averages (Predicted Effect Concentration, PEC), and compared to the lowest acute-based (PNEC(acute)) or chronic-based thresholds (PNEC(chronic)). For 56% of the compounds, PNECs were available from existing risk assessments, and the majority of these PNECs were derived from chronic toxicity data or simulated ecosystem studies (mesocosm) with rather low assessment factors. The limitations of this concept for risk assessment purposes are discussed. For the remainder, provisional PNECs (P-PNECs) were established from read-across models for acute toxicity to the standard test organisms Daphnia magna, Pimephales promelas and Selenastrum capricornutum. On the one hand, the prioritization revealed that about three-quarter of the 44 substances with MEC/PNEC ratios above ten were pesticides. On the other hand, based on the monitoring data used in this study, no risk with regard to the water phase could be found for eight of the 41 priority substances, indicating a first success of the implementation of the WFD in the investigated river basins.

Group contribution methods to estimate water solubility of organic chemicals

Group contribution methods to estimate water solubility have been studied on the basis of a test set of 694 organic nonelectrolytes, consisting of 351 liquids and 343 solids with experimental data taken from literature after critical evaluation. Derivation of a new fragmentation scheme leads to a squared correlation coefficient of 0.95 and an average absolute calculation error of 0.38 log units, which is superior to other group contribution methods currently available. Differences in performance for individual classes of compounds are discussed in detail. Solubility prediction for liquids is generally better than for solids, and the results support the inclusion of a melting point term to account for the entropy of fusion of solids.

Acute to chronic ratios in aquatic toxicity--variation across trophic levels and relationship with chemical structure.

For fish, daphnids, and algae, acute to chronic ratios (ACRs) have been determined from experimental data regarding new and existing chemicals. Only test results in accord with the European Union Technical Guidance Document (TGD) and validated by authorities were considered. Whereas the median ACRs of 10.5 (fish), 7.0 (daphnids), and 5.4 (algae) are well below the ACR safety factor of 100 as implied by the TGD, individual ACRs vary considerably and go up to 4400. The results suggest that a safety factor of 100 is not protective for all chemicals and trophic levels. Neither a correlation between ACR and baseline toxicity as modeled through the logarithmic octanol-water partition coefficient nor an ACR correlation across trophic levels exists. Narcosis is associated with a preference for a low ACR; nevertheless, low ACRs are frequently obtained for nonnarcotics. Analysis of chemical structures led to the derivation of structural alerts to identify compounds with a significantly increased potential for a high ACR, which may prove to be useful in setting test priorities. At present, however, life-cycle tests are the only way to conservatively predict long-term toxicity.

Chemical domain of QSAR models from atom-centered fragments.

A methodology to characterize the chemical domain of qualitative and quantitative structure-activity relationship (QSAR) models based on the atom-centered fragment (ACF) approach is introduced. ACFs decompose the molecule into structural pieces, with each non-hydrogen atom of the molecule acting as an ACF center. ACFs vary with respect to their size in terms of the path length covered in each bonding direction starting from a given central atom and how comprehensively the neighbor atoms (including hydrogen) are described in terms of element type and bonding environment. In addition to these different levels of ACF definitions, the ACF match mode as degree of strictness of the ACF comparison between a test compound and a given ACF pool (such as from a training set) has to be specified. Analyses of the prediction statistics of three QSAR models with their training sets as well as with external test sets and associated subsets demonstrate a clear relationship between the prediction performance and the levels of ACF definition and match mode. The findings suggest that second-order ACFs combined with a borderline match mode may serve as a generic and at the same time a mechanistically sound tool to define and evaluate the chemical domain of QSAR models. Moreover, four standard categories of the ACF-based membership to a given chemical domain (outside, borderline outside, borderline inside, inside) are introduced that provide more specific information about the expected QSAR prediction performance. As such, the ACF-based characterization of the chemical domain appears to be particularly useful for QSAR applications in the context of REACH and other regulatory schemes addressing the safety evaluation of chemical compounds.

Quantitative read-across for predicting the acute fish toxicity of organic compounds.

Read-across enables the interpolation of a property for a target chemical from respective experimental data of sufficiently similar compounds. Employing a set of 692 organic compounds with experimental values for the 96 h fish toxicity toward the fathead minnow in terms of LC(50) (lethal concentration 50%) values, a read-across method has been developed that is based on atom-centered fragments (ACFs) for evaluating chemical similarity. Prediction of log LC(50) proceeds through reading across the toxicity enhancement over predicted narcosis-level toxicity in terms of the respective logarithmic ratio, log T(e), and adding the respective baseline narcosis LC(50) estimated from log K(ow) (octanol/water partition coefficient). Depending on the minimum similarity imposed on a compound to serve as read-across basis for the target chemical, three different standard settings have been introduced, allowing one to perform screening-level estimations as well as predictions with intermediate and good confidence. The respective squared correlation coefficients (r(2)) are 0.73, 0.78, and 0.87, with root-mean square errors (rms) of 0.73, 0.60, and 0.39 log units, respectively. As a general trend, increasing the ACF minimum similarity increases the prediction quality at the cost of decreasing the application range. The method has the potential to assist in the predictive evaluation of fish toxicity for regulatory purposes such as under the REACH legislation.

Predicting Fate-Related Physicochemical Properties

The environmental fate of chemical substances is determined by partitioning between environmental compartments, and by transport and degradation processes. In this context, long range transport potential and persistence are important characteristics of the compound behaviour [1-7]. Besides specialized models to address the compound fate in individual environmental compartments, multimedia fate models have become popular in exposure and fate assessment on global and regional scales. A main application of such models is the screening-level prediction of the fate of environmental chemicals under standardized emission scenarios, and more recently the focus has shifted to more detailed process descriptions including time-dependent concentration levels of compounds and consideration of spatial resolution. For a more detailed account of multimedia fate modelling, the reader is referred to Chapter 4. In general, environmental compartments are modelled in terms of thermodynamic phases. The compoundspecific disposition for a certain environmental behaviour is thus determined by physicochemical properties governing partition processes, by molecular reactivity that drives abiotic degradation, and by the accessibility of the compound for microbial degradation. It follows that compound properties are essential input values for models to simulate their environmental behaviour and fate. As an example, multimedia fate models typically exploit the fugacity of chemicals [8] in the respective compartments. Fugacity is a central thermodynamic characteristic that in turn can be estimated from, or related to, some of the compound properties introduced in the following sections of this chapter.

Model selection based on structural similarity-method description and application to water solubility prediction.

A method is introduced that allows one to select, for a given property and compound, among several prediction methods the presumably best-performing scheme based on prediction errors evaluated for structurally similar compounds. The latter are selected through analysis of atom-centered fragments (ACFs) in accord with a k nearest neighbor procedure in the two-dimensional structural space. The approach is illustrated with seven estimation methods for the water solubility of organic compounds and a reference set of 1876 compounds with validated experimental values. The discussion includes a comparison with the similarity-based error correction as an alternative approach to improve the performance of prediction methods and an extension that enables an ad hoc specification of the application domain.

Prediction of the Intrinsic Hydrogen Bond Acceptor Strength of Chemical Substances from Molecular Structure

Hydrogen bonding affects the partitioning of organic compounds between environmental and biological compartments as well as the three-dimensional shape of macromolecules. Using the semiempirical quantum chemical AM1 level of calculation, we have developed a model to predict the site-specific hydrogen bond (HB) acceptor strength from ground-state properties of the individual compounds. At present, the model parametrization is confined to compounds with one HB acceptor site of the following atom types: N, O, S, F, Cl, and Br that act as lone-pair HB acceptors, and pi-electron (aromatic or conjugated) systems with the associated C atoms as particularly weak HB acceptors. The HB acceptor strength is expressed in terms of the Abraham parameter B and calculated from local molecular parameters, taking into account electrostatic, polarizability, and charge transfer contributions according to the Morokuma concept. For a data set of 383 compounds, the squared correlation coefficient r2 is 0.97 when electrostatic potential (ESP) derived net atomic charges are employed, and the root-mean-square (rms) error is 0.04 that is in the range of experimental uncertainty. The model is validated using an extended leave-50%-out approach, and its performance is comparatively analyzed with the ones of earlier introduced ab initio (HF/6-31G**) and density functional theory (B3LYP/6-31G**) models as well as of two increment methods with respect to the total compound set as well as HB acceptor type subsets. The discussion includes an explorative model application to amides and organophosphates that demonstrates the robustness of the approach, and further opportunities for model extensions.

Application of neural networks to modeling and estimating temperature-dependent liquid viscosity of organic compounds.

Back-propagation neural network models for correlating and predicting the viscosity-temperature behavior of a large variety of organic liquids were developed. Experimental values for the liquid viscosity for 1229 data points from 440 compounds containing C, H, N, O, S, and all halogens have been collected from the literature. The data ranges covered are from -120 to 160 degrees C for temperature and from 0.164 (trans-2-pentene at 20 degrees C) to 1.34 x 10(5) (glycerol at -20 degrees C) mPa.s for viscosity value. After dividing the total database of 440 compounds into training (237 with 673 data points), validation (124 with 423 data points), and test (79 with 133 data points) sets, the modeling performance of two separate neural network models with different architectures, one based on a compound-specific temperature dependence and the second based on a compound-independent one, has been examined. The resulting former model showed somewhat better modeling performance than latter, and the model gave squared correlation coefficients of 0.956, 0.932, and 0.884 and root mean-squares errors of 0.122, 0.134, and 0.148 log units for the training, validation, and test sets, respectively. The input descriptors include molar refraction, critical temperature, molar magnetic susceptibility, cohesive energy, temperatures, and five kinds of indicator variables for functionalities, alcohols/phenols, nitriles, amines, amides, and aliphatic ring. The reliability of the proposed model was assessed by comparing the results against calculated viscosities by two existing group-contribution approaches, the method of van Velzen et al. and the Joback and Reid method.