Computational Intelligence Methods for Physicochemical Property Prediction

SOTO AXEL JUAN; CECCHINI ROCÍO LUJÁN; PONZONI IGNACIO; VAZQUEZ GUSTAVO ESTEBAN

Córdoba

Conferencia; CLAFQO-9. 9a Conferencia Latinoamericana de Físico-Química Orgánica; 2007

Universidad de Córdoba

The ADMET properties determine the behavior of a drug in the human body. However, the rules that govern these properties are not clearly established. Additionally, ADMET properties are considered the main cause of failure in drug development. For these reasons, interest in QSAR/QSPR given by scientific and industrial community has grown considerably in last decades. QSAR methods developed by computer means are commonly referred as in silico methods. These methods are not pretended to be used in order to replace high quality in vitro experiments, but their consideration in the first stages of the development of a drug carries a series of important advantages. Some physicochemical properties are strongly related with ADMET rules, and given that they are simpler to understand than ADMET they are first considered to be modeled. Our developed work aims to systematize the process of predicting a physicochemical property of a chemical compound. This paper describes a set of algorithms and tools used in the overall process. In particular, our in silico methods are based on computational intelligence methods, i.e. algorithms that aim to learn the rules of a model in terms of existing experimental data. First, a front-end software that interacts with a database of chemical compounds was developed in order to query and retrieve the datasets of compounds and molecular descriptors that are necessary for the models. Evolutionary computing techniques were also applied for descriptor selection, that is, a technique for extracting relevant molecular features that result useful for the model. Finally, a property predictor method based on artificial neural network, was programmed in order to determine the numeric value of a physicochemical property. This same method was then improved by the incorporation of a cluster analysis method as a preprocessing step. All the aforementioned algorithms were tested for the hydrophobicity of a molecule. This property, traditionally express as the logarithm of the octanol-water partition coefficient (logP), was selected given their influence in ADMET properties.