Are you interested in finding the smallest contiguous region possible of the protein that confers activity or are you interested in peptidomimetics? The first commercial peptibody took a peptide mimetic approach and there was not much involved in terms of computational approach except for trying to predict immunogenicity. The information in this example is company confidential.
If you are interested in protein minimization, you need to get structure-activity relationships in order to see what is important for the function and must be kept in the mimetic.
In my view, if it corresponds to a linear part (ie a loop) and simple binding, it can be done. Nevertheless, the affinity will decrease and further work and the peptide must be done (affinity maturation/stabilization). In the case of enzymatic activity (ie a binding pocket) it can be really tricky...