Abstract
In the eld of molecular evolution, so called Structurally constrained (SC) models have been developped. Expressed at the codon level, they explicitely separe the mutation (applied to the nucleotide sequence) and the selection (applied to the encoded protein sequence) factors. The selection factor is described as a function between the structure and the sequence of the protein, via the use of a statistical potential. However, the whole evolutionary model depends on the expression of this potential, and one can ask wether a potential would be better than another. In this thesis, is developped a probabilistic framework to optimize statistical potentials especially meant for protein design, using a maximum likelihood approach. The statistical potential used in this thesis is composed by a contact potential and a solvent accessibility potential, but the probabilistic framework can easily be generalized to more complex statistical potentials. In a rst part, the framework is dened, and then an algorithmical enhancement is proposed, and nally, the framework is modied in order to take into account misfolded structures (decoys). The framework dened in this thesis and in other works allows to compare dierent optimization methods of statistical potentials for SC models, using cross-validation and Bayes factor comparisons.