Abstract : In silico prediction of protein structure from sequence is a major scientific challenge. It is now admitted that native 3D protein structures can be described by a limited set of recurring local structures. This observation led to the development of protein structure prediction techniques based on fragment assembly methods. Nowadays, these techniques are among the most effective. Protein local structure prediction is the first step toward the generation of global protein models. This thesis work mainly focuses on two major questions: (i) protein local structure prediction from sequence and (ii) the analysis of local structure predictability according to protein structure flexibility features. These analyses were based on a library - previously developed in the laboratory - of 120 3D structural prototypes encompassing all known local protein structures. An associated local structure prediction method from sequence had also been created and yielded a correct prediction rate of 51 %. Here, we achieved a balanced improvement of the prediction rate by coupling evolutionary information with Support Vector Machines. A very satisfying correct prediction rate of 63 % was obtained. Moreover, for directly estimating the quality of the prediction, we developed a confidence index which enables to identify regions that are more difficult to predict. In the same way, protein structures are not rigid macromolecules. Hence, we also extended our analysis and addressed the question of the structural predictability of a sequence with regards to its structural flexibility properties inside protein structures. We analyzed local structure flexibility features in proteins by relying on: (i) B-factors from X-ray experiments and (ii) backbone fluctuations observed in molecular dynamics simulations. Finally, an original flexibility prediction method from sequence was developed. Our different analyses are the first step toward the prediction of 3D global protein models.