Assessing Applicability Domains of Toxicological QSARs: Definition, Confidence in Predicted Values, and the Role of Mechanisms of Action

Abstract

There are many issues relating to the use of Quantitative Structure -Activity Relationships (QSARs) to make predictions for regulatory purposes. Among those issues, characterization of models and the development of suitable tools to determine applicability domains rank as the more important. With regard to aquatic toxicology, QSARs for acute effects (e.g., IGC 50 À1) often take the form of a hydrophobic [i.e., Logarithm of the 1-Octanol/Water Partition Coefficient (log P)]-electrophilic [e.g., Maximum Acceptor Superdelocalizability (A max )]-dependent, regression-based model. In this study, the applicability domain of a model for the toxicity of aromatic compounds to Tetrahymenawas assessed. The structural and physicochemical domains of the model were characterized using two test sets of toxicity data (one prescreened to be within the descriptor space and structural domain of the training set and the other to be outside the structural domain of the training set). For test set compounds inside the domain of the model, there was no relationship between absolute residue values for predictions and hydrophobicity; however, there was a linear relationship between absolute residue values and electrophilicity. It was concluded that predictivity in the region of the domain associated with higher electrophilicity, greater potency, and derivatives containing both halo-and nitro-groups is poorer than elsewhere in the domain, and therefore less confidence should be given to those values. Compounds in this region of the domain of the model are associated with the soft-, or pro-electrophilic mechanisms of toxic action. For the second test set, i.e., derivatives outside the structural domain, an examination of absolute residue values revealed that the observed toxicity is typically in excess of that predicted, especially for compounds in the structural space(s) of well-known electrophilic mechanisms of reactive toxicity. Caution is therefore urged in using statistical approaches to account for, and apply confidence to predictions from the applicability domain. An appreciation of the mechanism of toxicity appears to be critical to the determination of the most likely applicability domain.

Applicability domains and accuracy of prediction of soft sensor models

“…Studies on the ADs of statistical models have been performed primarily in the field of quantitative structure-activity relationship analysis. [19][20][21][22][23] Some excellent results were obtained in these studies. In this article, the distance to the average of training data and the distance to the nearest neighbor (NN) of training data are used as DMs.…”

Accurate Prediction of Aquatic Toxicity of Aromatic Compounds Based on Genetic Algorithm and Least Squares Support Vector Machines

“…The toxicities of 384 aromatic compounds to T. pyriformis together with their log P values were obtained from the literature [7] and are listed in Table 1. The toxicities of the molecules were expressed as the minus log unit of the 50% Growth Inhibitory Concentration (À logIGC 50 ) values.…”

“…In addition, to verify the predictive ability and reliability of derived model in a more creditable way, a completely new external data set consisting of 92 aromatic compounds, taken from Ref. [7] (actually there were 94 compounds with 2 repeated ones, so we excluded the repeated compounds and 92 were used in this paper), was used. The toxicities and log P values of these 92 compounds are listed in Table 3.…”

“…In the previous work [7], the log P was one of the most important descriptor correlated to the toxicity. The importance of the other two descriptors [Me 2 and Topological Polar Surface Area; TPSA (NO)] is similar according to their standardized regression coefficients 0.2827 and 0.2889, respectively.…”

“…The resulting QSTRs revealed that the 1-octanol/water partition coefficient (log K ow ), is the most important descriptor of aldehyde aquatic toxic potency [3]. Cronin and coworkers developed a two-parameter (log P and A max ) QSTR model for the toxicity of 384 aromatic compounds to Tetrahymena pyriformis [7]. Panaye et al predicted the toxicity of 203 nitro-and cyano-aromatic chemicals to T. pyriformis by radial basis function neural network, general regression neural network, and support vector machine from calculated molecular descriptors [8].…”