Multi-Objective Feature Selection in QSAR Using a Machine Learning Approach

Authors

  • Axel J. Soto,

    1. Laboratorio de Investigación y Desarrollo en Computación Científica (LIDeCC), Departamento de Ciencias e Ingeniería de la Computación, Universidad Nacional del Sur, Av. Alem 1253, 8000 Bahía Blanca, Argentina
    2. Planta Piloto de Ingeniería Química (PLAPIQUI), Universidad Nacional del Sur, CONICET, Camino La Carrindanga km.7, CC 717, Bahía Blanca, Argentina
    Search for more papers by this author
  • Rocío L. Cecchini,

    1. Laboratorio de Investigación y Desarrollo en Computación Científica (LIDeCC), Departamento de Ciencias e Ingeniería de la Computación, Universidad Nacional del Sur, Av. Alem 1253, 8000 Bahía Blanca, Argentina
    Search for more papers by this author
  • Gustavo E. Vazquez,

    1. Laboratorio de Investigación y Desarrollo en Computación Científica (LIDeCC), Departamento de Ciencias e Ingeniería de la Computación, Universidad Nacional del Sur, Av. Alem 1253, 8000 Bahía Blanca, Argentina
    Search for more papers by this author
  • Ignacio Ponzoni

    1. Laboratorio de Investigación y Desarrollo en Computación Científica (LIDeCC), Departamento de Ciencias e Ingeniería de la Computación, Universidad Nacional del Sur, Av. Alem 1253, 8000 Bahía Blanca, Argentina
    2. Planta Piloto de Ingeniería Química (PLAPIQUI), Universidad Nacional del Sur, CONICET, Camino La Carrindanga km.7, CC 717, Bahía Blanca, Argentina
    Search for more papers by this author

Abstract

The selection of descriptor subsets for QSAR/QSPR is a hard combinatorial problem that requires the evaluation of complex relationships in order to assess the relevance of the selected subsets. In this paper, we describe the main issues in applying descriptor selection for QSAR methods and propose a novel two-phase methodology for this task. The first phase makes use of a multi-objective evolutionary technique which yields interesting advantages compared to mono-objective methods. The second phase complements the first one and it enables to refine and improve the confidence in the chosen subsets of descriptors. This methodology allows the selection of subsets when a large number of descriptors are involved and it is also suitable for linear and nonlinear QSAR/QSPR models. The proposed method was tested using three data sets with experimental values for blood-brain barrier penetration, human intestinal absorption and hydrophobicity. Results reveal the capability of the method for achieving subsets of descriptors with a high predictive capacity and a low cardinality. Therefore, our proposal constitutes a new promising technique helpful for the development of QSAR/QSPR models.

Ancillary