DELPHOS: Computational Tool for Selection of Relevant Descriptor Subsets in ADMET Prediction

AXEL JUAN SOTO; MARÍA J. MARTÍNEZ; ROCÍO LUJÁN CECCHINI; GUSTAVO ESTEBAN VAZQUEZ; IGNACIO PONZONI

Los Cocos, Provincia de Córdoba, Argentina

Conferencia; 1ª Reunión Internacional de Ciencias Farmacéuticas (RICiFA 2010); 2010

Facultad de Ciencias Bioquímicas y Farmaceúticas, Universidad Nacional de Rosario

INTRODUCTION The design of QSAR (Quantitative Structure-Activity Relationship) methods constitutes an important research topic in modern drug discovery. One of the first and most important steps is the selection of the subset of descriptors that are relevant to the activity or property under study. In general, the descriptor selection task could not be manually achieved by experts, given the inherent complexity and non-linearity of the structure-activity relationships. Moreover, the number of molecular descriptors that may be calculated for a single compound is huge. Thereby, it is important to have a computational method for the selection of the subset of molecular descriptors to be used in a QSAR model.MATERIALS AND METHODSIn this paper we present DELPHOS that is a computational tool that was recently made available to assist pharmacists who work with QSAR/QSPR prediction models. The first release of DELPHOS has been launched in May 2010 and may be found at "http://lidecc.cs.uns.edu.ar/". The computational kernel of DELPHOS is based on a novel approach which has been recently published in a QSAR-related journal[1]. DELPHOS makes use of a two-phase computational method where the first phase executes a multi-objective optimization, using evolutionary algorithms, and the second phase is a thorough validation of the results obtained in the first step.Currently the DELPHOS software has the following features:GUI. A Graphic User Interface designed for using the software without the need to know specific details of the code or the applied methods.Data handling. Input data can be fed to the method using the CSV file format or standard Matlab matrix files. Computation performed after any phase can be saved and later restored.First Phase: Feature Searching and Evaluation. This phase is responsible of doing a coarse searching and a fast evaluation among all feasible subsets of descriptors. Several different parameters could be set for this phase.Second Phase: Learning Method. Using the data computed in the first phase, a thorough evaluation is applied in order to determine which subsets of the coarse selection are the most relevant ones.Post-processing. After the second phase has been executed, tables showing final results and severa lstatistical metrics are presented.RESULTSDELPHOS was successfully tested on three physicochemical and biological properties that are important for pharmacokinetic screening of drugs: hydrophobicity, blood-brain barrier penetration and human intestinal absorption. Moreover, our method has overcome prediction capacity of subsets of descriptors obtained by other competitive methods.CONCLUSIONSThe multi-objective optimization of DELPHOS looks for models where accuracy and simplicity are maximized. As a result, the method aims at selecting descriptor subsets with few descriptors. Another interesting point of DELPHOS is that more than a single subset is suggested to the user, so that the scientist is able to select from subsets of similar relevance based on other criteria, such as interpretability.ACKNOWLEDGMENTSThis work is supported by grants CONICET PIP11220090100322 and UNS PGIs 24/ZN15-ZN16.