DK-507K

Quantitative structure-activity relationship study on three diverse sets of structurally similar fluoroquinolones was performed using a comprehensive set of molecular descriptors. Multiple linear regression technique was applied as a preprocessing tool to find the set of relevant descriptors (10) which are subsequently used in the artificial neural networks approach (non-linear procedure). The biological activity in the series (minimal inhibitory concentration (μg/mL) was treated as negative decade logarithm, pMIC). Using the non-linear technique counter propagation artificial neural networks, we obtained good predictive models. All models were validated using cross validation leave-one-out procedure. The results (the best models: Assay1, R = 0.8108; Assay2, R = 0.8454, and Assay3, R = 0.9212) obtained on external, previously excluded test datasets show the ability of these models in providing structure-activity relationship of fluoroquinolones. Thus, we demonstrated the advantage of non-linear approach in prediction of biological activity in these series. Furthermore, these validated models could be proficiently used for the design of novel structurally similar fluoroquinolone analogues with potentially higher activity.

There are many different kinds of pathogenic bacteria species with very different susceptibility profiles to different antibacterial drugs. One limitation of QSAR models is that they consider the biological activity of drugs against only one species of bacteria. In a previous paper, we developed a unified Markov model to describe the biological activity of different drugs tested in the literature against some antimicrobial species. Consequently, predicting the probability with which a drug is active against different species of bacteria with a single unified model is a goal of major importance. The work described here develops the unified Markov model to describe the biological activity of more than 70 drugs from the literature tested against 96 species of bacteria. We applied linear discriminant analysis (LDA) to classify drugs as active or inactive against the different tested bacterial species. The model correctly classified 199 out of 237 active compounds (83.9%) and 168 out of 200 inactive compounds (84%). Overall training predictability was 84% (367 out of 437 cases). Validation of the model was carried out using an external predicting series, with the model classifying 202 out of 243 (i.e., 83.13%) of the compounds. In order to show how the model functions in practice, a virtual screening was carried out and the model recognized as active 84.5% (480 out of 568) antibacterial compounds not used in the training or predicting series. The current study is an attempt to calculate within a unified framework the probabilities of antibacterial action of drugs against many different species.