Application of ensemble learning for predicting bioavailability of drugs in the Central Nervous System

JUAN FRANCISCO MORALES; MARÍA ESPERANZA RUIZ; ALAN TALEVI

Buenos Aires

Workshop; Biowaivers 2015: Implementación de las bioexenciones basadas en el BCS; 2015

International Pharmaceutical Federation (FIP) - Control de Calidad de Medicamentos, Fac. Cs. Exactas, UNLP

The aim of the present work is to develop in silico filters for the early prediction of brain bioavailability of drug candidates. We developed a series of QSAR classifiers models that were selectively assembled to improve its predictive ability. MATERIAL AND METHODS: We begin performing a literature search of mammalian Kpuu values obtained by the microdialysis and slice techniques, to compile a dataset with those compounds which Kpuu value was below 1.4 (it is considered that higher values indicate active transport of the compound across the BBB). A dataset of 89 drugs were obtained, and two categories were defined using a cutoff value of 0.4. Our dataset consisted of 31 drugs with good ability to passively diffuse across the BBB (category A) and 58 drugs with limited brain bioavailability (category B). Due to the small number of compounds in the "active" (A) category, the training set was integrated with all the compounds belonging to category A plus 31 compounds randomly selected from category B. The remaining 27 drugs of category B formed the test set. More than 2000 conformation independent molecular descriptors were calculated and randomly divided into 100 subsets of no more than 200. The models were derived with the Statistica 8.0 software, using a linear discriminant analysis. One-hundred classifiers models were thus obtained and all of them were validated by standard techniques of internal and external validation, and compared using receiver operating characteristic curves (ROC). The best models were combined with selective learning assembly (ensemble learning) using the Average, Minimum and Maximum operators as combination techniques. RESULTS: The best individual model showed 89% and 93% of good classifications in the training and test set respectively. The area under the ROC curve (AUC ROC) for the training set was 0.927. The combination of the best models using the Minimum operator showed superior performance in terms of specificity, resulting in an 89% and 93% of good classifications in the training set and test set, respectively, and AUC ROC of 0.969. CONCLUSIONS: An in silico filter was successfully developed and validated by computational tools. Our next goal is to verify these predictions by the experimental validation of the developed combined models.