The recent unveiling of the human genome -- and its surprising conclusion that the number of genes encoded in human DNA is much lower than expected -- places all the more importance on the study of proteins and their interactions. Recently we have been involved in using large-scale optimization techniques to model the energy function that underlies the folding process of proteins. Linear programming is used to identify parameters in the energy function model, the objective being that the model recognize the structure of known proteins correctly. Such trained functions can then be used either for ab-initio prediction or for recognition of unknown structures. In order to obtain good energy models we need to be able to solve dense LPs with tens of millions of constraints in a few hundred parameters, which we achieve by tailoring and parallelizing our code PCx.