Predicting Physicochemical Properties for Drug Design Using Clustering and Neural Network Learning

AXEL JUAN SOTO; IGNACIO PONZONI; GUSTAVO ESTEBAN VAZQUEZ

Angra Dos Reis, Brasil

Simposio; Brazilian Symposium on Bioinformatics 2007; 2007

Prediction of physicochemical properties is of major concern for pharmaceutical research. In this context, machine learning methods are of great importance due to their contribution to the development of a plethora of models. Actually, many predictors exist but most of them do not correctly generalize when external data is presented. We present a novel framework for physicochemical property prediction, where training data is first clustered according to their structural similarity, and a classifier is trained for each conformed cluster. In this regard, the property prediction of a novel candidate drug is modelled by the classifier associated with the cluster that has more structurally-similar compounds with regard to the new putative drug. The generalization problem is not completely solved with the presented approach, but it allows to reduce the prediction error of the method. Artificial neural networks (NN) are used as classifiers and an analysis on logP (octanol-water distribution coefficient) is used to show the advantages of our proposal.