Machine learning-based models to predict modes of toxic action of phenols to Tetrahymena pyriformis

Abstract

The phenols are structurally heterogeneous pollutants and they present a variety of modes of toxic action (MOA), including polar narcotics, weak acid respiratory uncouplers, pro-electrophiles, and soft electrophiles. Because it is often difficult to determine correctly the mechanism of action of a compound, quantitative structure-activity relationship (QSAR) methods, which have proved their interest in toxicity prediction, can be used. In this work, several QSAR models for the prediction of MOA of 221 phenols to the ciliated protozoan Tetrahymena pyriformis, using Chemistry Development Kit descriptors, are reported. Four machine learning techniques (ML), k-nearest neighbours, support vector machine, classification trees, and artificial neural networks, have been used to develop several models with higher accuracies and predictive capabilities for distinguishing between four MOAs. They showed global accuracy values between 95.9% and 97.7% and area under Receiver Operator Curve values between 0.978 and 0.998; additionally, false alarm rate values were below 8.2% for training set. In order to validate our models, cross-validation (10-folds-out) and external test-set were performed with good behaviour in all cases. These models, obtained with ML techniques, were compared with others previously reported by other researchers, and the improvement was significant.

Keywords:

• Molecular descriptor
• machine learning technique
• QSAR
• mode of toxic action
• phenol derivative
• pollutant

Acknowledgements

J.A. Castillo-Garit and G.M. Casañola-Martin thank the program ‘Estades Temporals per an Investigadors Convidats’ for a fellowship to work at Valencia University in 2013. F. Torrens acknowledges support from the Spanish Ministerio de Economía y Competitividad (Project No. BFU2013-41648-P) and EU ERDF.