A new method for multi-objective selection of molecular descriptors for QSAR/QSPR

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

Montevideo

Conferencia; ISCB-LA 2010 (First ISCB Latin-American Meeting); 2010

International Society for Computational Biology (ISCB) y Institut Pasteur (filial Montevideo)

We proposed a two-phase methodology aimed at selecting relevant descriptors for linear and nonlinear QSAR models. The first phase consists in a multi-objective (MO) evolutionary wrapper that aims both, to maximize predictive capacity and to reduce the number of selected descriptors. This phase is responsible for an initial coarse selection of the subset of descriptors by using simple prediction methods, which allow evaluating the combinatorial chemical search space in a fast way. The second  phase builds prediction models from the descriptor subsets chosen by the wrapper method by using neural network ensembles. The goal of this phase is to have a more rigorous assessment of the prediction capacity of the subsets pre-selected by the wrapper.