Abstract: In a post-genomic context, plenty of proteins identified by their sequence have no experimentaly resolved structure, and fall out the range of application of comparative modelling methods. The goal of my PhD thesis has been to explore a new de novo protein structure prediction approach. This approach is based on the concept of structural alphabet, i.e. of a local description of protein architecture by using a small number of prototype conformations. Starting from the amino acids sequence of the protein to model, we have developped a candidate fragments prediction method covering more than 98.6% of the protein structure with an average length of 6.7 residues. This set of predicted fragments can approximate the protein structures with a precision of less than 2.2 A. A greedy algorithm have been developped in the laboratory to assemble fragments. The OPEP force field has been optimized and then implemented in the greedy assembling procedure to evaluate the relevance of the generated models. Our participation to the CASP7 experiment came out some weaknesses of the method. For now, the improvement of the OPEP force field and the fragment assembly procedure leeds us to generate, in some cases, models as relevant or better than other famous protein structure prediction servers.