A new methodology for 3D-QSAR studies, based on the combined use of global and local approaches, is proposed. Information on the whole molecular electrostatic potential distribution is obtained from a global approach and used to build global models, based on the grid weighted holistic invariant molecular (G-WHIM) descriptors, as well as to define an alignment criterion for developing local models. Moreover, a new technique based on local correlation indexes is proposed to perform the variable reduction. Both the global and the local models are obtained by a variable selection genetic algorithms technique. The methodology is applied to a benchmark steroid data set. The molecular electrostatic potential and the experimental binding affinity constants for the corticosteroid-binding globulin are used as variables. The QSAR models obtained compared with those obtained by the CoMFA method show higher predictive ability and give insight into the local molecular features that determine high binding affinity.