QSPR analysis of threshold of odor for the large number of heterogenic chemicals

Quantitative structure–property relationships for odor thresholds based on representation of the molecular structure by the simplified molecular input-line entry system were established using the CORAL software. The total set of compounds with numerical data on the so-called arithmetic odor thresholds (\(n=1259\)) was distributed into the training and validation sets, three times. The average statistical quality of these models is (1) for training set \(\tilde{n}=967\pm 20({\approx }\,80\%), {\mathop {{r}}\limits ^\frown }^{2}=0.62\pm 0.02\); and (2) for validation set \(\tilde{n}=290\pm 20({\approx }\,20\%), {\mathop {{r}}\limits ^\frown }^{2}=0.62\pm 0.04\). Thus, the predictive potential of this approach was confirmed for three different splits into training and validation sets. Domain of applicability and mechanistic interpretation of these models are defined from the probabilistic point of view. The suggested models are built up according to OECD principles.