QSAR modeling of acute toxicity by balance of correlations

Optimal descriptors based on the simplified molecular input line entry system (SMILES) have been utilized in modeling of acute toxicity towards rats. Toxicity of 61 benzene derivatives has been modeled by means of balance of correlations for sets of the training (n = 27) and calibration (n = 24). The obtained models were evaluated with the external test set (n = 10). Comparison of models based on the balance of correlations and models which were obtained on base of the total training (i.e., in case of utilization both training and calibration sets as the united training set) has shown that the balance of correlations gives improvement of statistical quality for the external test set. Predictions based on the one-variable model (based on the correlation balance) are better that the results obtained by the multiple linear regression analysis based on topological and quantum chemical descriptors. A QSAR analysis showed that the electronegativity of the molecule plays an important role in acute toxicity of benzene derivatives studied; presence of electronegative groups increasing toxicity. The presence of nitrogen-containing groups (mostly NH groups) increasing the toxicity that confirmed by both approaches.