Development of predictive models by adaptive fuzzy partitioning. Application to compounds active on the central nervous system

摘要

A new data mining method, derived from Fuzzy Logic concepts, was developed in order to classify biochemical databases and to predict the activities of large series of untested compounds. This technique, called Adaptive Fuzzy Partition (AFP), builds relationships between molecular descriptors and biochemical activities by dynamically dividing the descriptor hyperspace into a set of fuzzily partitioned subspaces. These subspaces are described by simple linguistic rules, from which scores ranging between 0 and 1 can be derived. The latter values define, for each compound, the degrees of membership of the different biological properties analyzed. The prediction ability of AFP was evaluated by analyzing a training set of 377 central nervous system (CNS)-active molecules subdivided into eight receptor classes. After selecting the most relevant descriptors by a procedure combining genetic algorithms and stepwise techniques, the best AFP model was selected and validated by a validation set. Furthermore, its robustness was confirmed by predicting a test set of 102 compounds never used to define the AFP models. Encouraging validation ratios of about 80% were obtained in the prediction of the experimental CNS activities. Finally, a comparison between the results obtained by AFP and by other classic techniques showed that AFP improved sensibly the prediction power of the proposed models.