Selection plays an important role in estimation of distribution algorithms. It determines the solutions that will be modeled to represent the promising areas of the search space. There is a strong relationship between the strength of selection and the type and number of dependencies that are captured by the models. In this paper we propose to use different selection probabilities to learn the structural and parametric components of the probabilistic graphical models. Customized selection is introduced as a way to enhance the effect of model learning in the exploratory and exploitative aspects of the search. We use a benchmark of over 15,000 instances of a simplified protein model to illustrate the gains in using customized selection.