Binary classification models for endocrine disrupter effects mediated through the estrogen receptor.

Endocrine disrupters (EDs) form an interesting field of application attracting great attention in the recent years. They represent a number of exogenous substances interfering with the function of the endocrine system, including the interfering with developmental processes. In particular EDs are mentioned as substances requiring a more detailed control and specific authorization within REACH, the new European legislation on chemicals, together with other groups of chemicals of particular concern. QSAR represents a challenging method to approach data gap which is foreseen by REACH. The aim of this study was to provide an insight into the use of QSAR models to address ED effects mediated through the estrogen receptor (ER). New predictive models were derived to assess estrogenicity for a very large and heterogeneous dataset of chemical compounds. QSAR binary classifiers were developed based on different data mining techniques such as classification trees, decision forest, fuzzy logic, neural networks and support vector machines. The focus was given to multiple endpoints to better characterize the effects of EDs evaluating both binding (RBA) and transcriptional activity (RA). A possible combination of the models was also explored. A very good accuracy was reached for both RA and RBA models (higher than 80%).

Ecotoxicity prediction by adaptive fuzzy partitioning: comparing descriptors computed on 2D and 3D structures.

Classification models were established on four endpoints, i.e. trout, daphnia, quail and bee, including from 100 to 300 pesticides subdivided into 3 toxicity classes. For each species, five separate sets of molecular descriptors, computed by several software, were compared, including parameters related to 2D or 3D structures. The most relevant descriptors were selected with help of a procedure based on genetic algorithms. Then, structure-activity relationships were built by Adaptive Fuzzy Partition (AFP), a recursive partitioning method derived from Fuzzy Logic concepts.Globally, satisfactory results were obtained for each animal species. The best cross-validation and test set scores reached values of about 70-75%. More important, the relationships derived from the descriptors calculated from 2D structures were superior or similar to those computed from 3D structures. These results underline that the long computational time employed to compute 3D descriptors is often useless to improve the prediction ability of the ecotoxicity models. Finally, the differences in the prediction ability between the different software used were quite reduced and show the possibility to use different descriptor packages for obtaining similar satisfactory models.

Predicting logP of pesticides using different software.

We compared experimental and calculated logP values using a data set of 235 pesticides and experimental values from four different sources: The Pesticide Manual, Hansch Manual, ANPA and KowWin databases. LogP were calculated with four softwares: HyperChem, Pallas, KowWin and TOPKAT. Crossed comparison of the experimental and calculated values proved useful, especially for pesticides. These are harder to study than simpler organic compounds. Structurally they are complex, heterogeneous and similar to drugs from a chemical point of view. They offer an interesting way to verify the goodness of the different methods. Other studies compared several logP predictors using a single set of experimental values taken as a reference. Here we discuss the utility of the different logP predictors, with reference to experimental data found in different databases. This offers three advantages: (1) it avoids bias due to the assumption that one single data set is correct; (2) a given predictor can be developed on the same data set used for evaluation; (3) it takes account of experimental variability and can compare it with the predictor's variability. In our study Pallas and KowWin gave the best results for prediction, followed by TOPKAT.

Factors influencing predictive models for toxicology.

Comparisons of different models to predict toxicity and evaluation of the predictive power of a model are affected by the variability of the data. We assessed this problem by considering experimental toxicity data and chemical descriptors. We evaluated several toxicological end-points (Oncorhynchus mykiss, Daphnia magna, Acceptable Daily Intake, Anas Platyrhynchos, Colinus virginianus and Muridae) in the case of pesticides and also considered the availability of toxicological data. We calculated hundreds of molecular descriptors (divided into constitutional, electrostatic, geometrical, quantum-chemical and topological ones) for the selected compounds using CODESSA, HyperChem and Pallas. Molecular descriptors may vary depending on the conformation of the molecules and on the software used. We evaluated the extent of this variability, and compared it with the variability of the experimental toxicological values.