Diagnostic of a Qspr Model: Aqueous Solubility of Drug-like Compounds

A diagnostic test for a qSPR (quantitative Structure-Property Relationship) model was carried out using a series of statistical indicators for correctly classifying compounds into actives and non-actives. A previously reported qSPR model, able to characterize the aqueous solubility of drug-like compounds, was used in this study. Eleven statistical indicators like those used in medical diagnostic tests were defined and applied on training, test and overall data sets. The associated 95% confidence interval under the binomial distribution assumption was also computed for each defined indicator in order to allow a correct interpretation. Similar results were obtained in the training and test sets with some exceptions. The prior probabilities of active and nonactive compounds proved not to be significantly different in the training and test sets. However, the probability of classification as active compounds proved to be significantly smaller in the training set as compared to the test set (p = 0.0042). The total fraction of correctly classified compounds proved to be identical in the training and test sets as well as in the overall set. Nevertheless, the overall model and the model obtained in the test set show a higher ability to correctly assign the non-active compounds to the non-active class while the model obtained in the training set has a higher ability to correctly assign the active compounds to the active class.