Large worldwide use of chemicals has caused great concern about their possible adverse effects on human health, flora and fauna. Increased production of new chemicals has also increased demand for their risk assessment. Traditionally, results from animal tests have been used to assess toxicity of chemicals. However, such methods are ethically questionable since they involve killing and causing suffering of the test animals. Therefore, new in silico methods are being sought to replace the traditional in vivo and in vitro testing methods. In this article we report on one method that can be used to build robust models for the prediction of compounds’ properties from their chemical structure. The method has been developed by combining a genetic algorithm, a counter-propagation artificial neural network and cross-validation. It has been tested using existing data on toxicity to fathead minnow (Pimephales promelas). The results show that the method may give reliable results for chemicals belonging to the applicability domain of the developed models. Therefore, it can aid the risk assessment of chemicals and consequently reduce demand for animal tests.

Drug-induced liver injury is a major concern in the drug development process. Expensive and time-consuming in vitro and in vivo studies do not reflect the complexity of the phenomenon. Complementary to wet lab methods are in silico approaches, which present a cost-efficient method for toxicity prediction. The aim of our study was to explore the capabilities of counter-propagation artificial neural networks (CPANNs) for the classification of an imbalanced dataset related to idiosyncratic drug-induced liver injury and to develop a model for prediction of the hepatotoxic potential of drugs. Genetic algorithm optimization of CPANN models was used to build models for the classification of drugs into hepatotoxic and non-hepatotoxic class using molecular descriptors. For the classification of an imbalanced dataset, we modified the classical CPANN training algorithm by integrating random subsampling into the training procedure of CPANN to improve the classification ability of CPANN. According to the number of models accepted by internal validation and according to the prediction statistics on the external set, we concluded that using an imbalanced set with balanced subsampling in each learning epoch is a better approach compared to using a fixed balanced set in the case of the counter-propagation artificial neural network learning methodology.

Background: In the present study, we show the correlation of quantum chemical structural descriptors with the activation barriers of the Diels-Alder ligations. A set of 72 non-catalysed Diels-Alder reactions were subjected to quantitative structure-activation barrier relationship (QSABR) under the framework of theoretical quantum chemical descriptors calculated solely from the structures of diene and dienophile reactants. Experimental activation barrier data were obtained from literature. Descriptors were computed using Hartree-Fock theory using 6-31G(d) basis set as implemented in Gaussian 09 software. Results: Variable selection and model development were carried out by stepwise multiple linear regression methodology. Predictive performance of the quantitative structure-activation barrier relationship (QSABR) model was assessed by training and test set concept and by calculating leave-one-out cross-validated Q² and predictive R² values. The QSABR model can explain and predict 86.5% and 80% of the variances, respectively, in the activation energy barrier training data. Alternatively, a neural network model based on back propagation of errors was developed to assess the nonlinearity of the sought correlations between theoretical descriptors and experimental reaction barriers. Conclusions: A reasonable predictability for the activation barrier of the test set reactions was obtained, which enabled an exploration and interpretation of the significant variables responsible for Diels-Alder interaction between dienes and dienophiles. Thus, studies in the direction of QSABR modelling that provide efficient and fast prediction of activation barriers of the Diels-Alder reactions turn out to be a meaningful alternative to transition state theory based computation.

CPANNatNIC is software for development of counter-propagation artificial neural network models. Besides the interface for training of a new neural network it also provides an interface for visualisation of the results which was developed to aid in interpretation of the results and to use the program as a tool for read-across. The work presents the details of the program’s interface. Parts of the interface are presented and how they can be used. The examples provided show how the user can build a new model and view the results of predictions using the interface. Examples are given to show how the software may be used in read-across. CPANNatNIC provides a simple user interface for model development and visualisation. The interface implements options which may simplify read-across procedure. Statistical results show better prediction accuracy of read-across predictions than model predictions where similar compounds could be identified, which indicates the importance of using read-across and usefulness of the program.

The hepatotoxic potential of drugs is one of the main reasons why a number of drugs never reach the market or have to be withdrawn from the market. Therefore, the evaluation of the hepatotoxic potential of drugs is an important part of the drug development process. The aim of this work was to evaluate the relative abilities of different supervised self-organizing algorithms in classifying the hepatotoxic potential of drugs. Two modifications of standard counter-propagation training algorithms were proposed to achieve good separation of clusters on the self-organizing map. A series of optimizations were performed using genetic algorithm to select models developed with counter-propagation neural networks, X-Y fused networks, and the two newly proposed algorithms. The cluster separations achieved by the different algorithms were evaluated using a simple measure presented in this paper. Both proposed algorithms showed a better formation of clusters compared to the standard counter-propagation algorithm. The X-Y fused neural network confirmed its high ability to form well-separated clusters. Nevertheless, one of the proposed algorithms came close to its clustering results, which also resulted in a similar number of selected models.

Artificial neural networks (ANNs) are nowadays applied as the most efficient methods in the majority of machine learning approaches, including data-driven modeling for assessment of the toxicity of chemicals. We developed a combined neural network methodology that can be used in the scope of new approach methodologies (NAMs) assessing chemical or drug toxicity. Here, we present QSAR models for predicting the physical and biochemical properties of molecules of three different datasets: aqueous solubility, acute fish toxicity toward fat head minnow, and bio-concentration factors. A novel neural network modeling method is developed by combining two neural network algorithms, namely, the counter-propagation modeling strategy (CP-ANN) with the back-propagation-of-errors algorithm (BPE-ANN). The advantage is a short training time, robustness, and good interpretability through the initial CP-ANN part, while the extension with BPE-ANN improves the precision of predictions in the range between minimal and maximal property values of the training data, regardless of the number of neurons in both neural networks, either CP-ANN or BPE-ANN.