In the recent times, the race to cope with the increasing multidrug resistance of pathogenic bacteria has lost much of its momentum and health professionals are grasping for solutions to deal with the unprecedented resistance levels. As a result, there is an urgent need for a concerted effort towards the development of new antimicrobial drugs to stay ahead in the fight against the ever adapting bacteria. In the present report, antibacterial classification functions (models) based on the topological molecular computational design-computer aided ‘‘rational’’ drug design (TOMOCOMD-CARDD) atom-based non-stochastic and stochastic bilinear indices are presented. These models were built using the linear discriminant analysis (LDA) method over a balanced chemical compounds dataset of 2230 molecular structures, with a diverse range of structural and molecular mechanism modes. The results of this study indicated that the non-stochastic and stochastic bilinear indices provided excellent classification of the chemical compounds (with accuracies of 86.31% and 84.92%, respectively, in the training set). These models were further externally validated yielding correct classification percentages of 86.55% and 87.91% for the non-stochastic and stochastic bilinear models, respectively. Additionally, the obtained models were compared with those reported in the literature and demonstrated comparable results, although the latter were built over much smaller datasets and with much higher degrees of freedom. Finally, simulated ligand-based virtual screening of 116 compounds, recently identified as potential antibacterials, was performed yielding 86.21% and 83.62% of correct classification, respectively, and thus demonstrating the utility of the obtained TOMOCOMD-CARDD models in the search of novel compounds with desirable antibacterial activity.

The recently introduced, non-stochastic and stochastic quadratic indices (Marrero-Ponce et al. J. Comp. Aided Mol. Des. 2006, 20, 685-701) were applied to QSAR/QSPR studies of heteroatomic molecules. These novel bond-based molecular descriptors (MDs) were used for the prediction of the partition coefficient (log P), and the antibacterial activity of 34 derivatives of 2-furylethylenes. Two statistically significant QSPR models using non-stochastic and stochastic bond-based quadratic indices were obtained (R² = 0.971, s = 0.137 and R² = 0.986, s = 0.096). These models showed good stability to data variation in leave-one-out (LOO) cross-validation experiment (q² = 0.9975, sCV = 0.164 and q² = 0.947, sCV = 0.114). The best two discriminant models computed using the non-stochastic and stochastic molecular descriptors had globally good classification of 94.12% in the training set. The external prediction sets had accuracies of 100% in both cases. The comparison with other approaches (edge- and vertex-based connectivity indices, total and local spectral moments, quantum chemical descriptors as well as with other TOMOCOMD-CARDD MDs) revealed the good performance of our method in this QSPR study. The obtained results suggest that it is possible to obtain a good estimation of physical, chemical and physicochemical properties for organic compounds with the present approach.

En el presente reporte, se aplican los índices cuadráticos de relaciones de enlace introducidos recientemente (Marrero-Ponce et al. J. Comp. Aided Mol. Des. 2006, 20, 685-701) en estudios QSAR/QSPR de moléculas heteroatómicas. Estos descriptores moleculares de tipo enlace son usados en la predicción del coeficiente de partición (log P) y la actividad antibacterial de 34 derivados de los 2-furiletilenos. Dos modelos QSPR estadísticamente significativos fueron obtenidos usando índices no estocásticos y estocásticos (R² = 0.971, s = 0.137 y R² = 0.986, s = 0.096, respectivamente) en la modelación del Log P. Estos modelos mostraron una estabilidad adecuada en la validación interna LOO (q² = 0.9975, sCV = 0.164 and q² = 0.947, sCV = 0.114, respectivamente). Por otro lado, los dos mejores modelos discriminantes muestran un porcentaje de exactitud global de 94.12% en la serie de entrenamiento y de 100% en la data de predicción en la modelación de actividad bactericida. Finalmente, la comparación con otros enfoques computacionales (índices de conectividad de enlace y de átomo tanto 2D como 3D, momentos espectrales totales y locales, descriptores químicos cuánticos al igual que con otros índices implementados en el programa TOMOCOMD-CARDD) evidencia un buen comportamiento de nuestros nuevos índices. Los resultados obtenidos sugieren que el método propuesto permite obtener una adecuada estimación de propiedades fisicoquímicas y biológicas de moléculas orgánicas.