Structural genomics projects have revealed remarkable features of proteomes. But these are essentially of combinatorial nature---selected proteins interact within a complex, so that extending them to the structural level requires building 3D models of these complexes. Such models have recently been reconstructed for the Nuclear Pore Complex (NPC), based on the integration of diverse biophysical and biochemical data. Yet, a full synergy between them and the experimental data is not at play because the reconstructions are qualitative. This thesis makes three contributions addressing these limitations. First, we introduce toleranced models (TOM) to inherently represent uncertain shapes as a continuum of models. We show that a TOM is equivalent to a compoundly weighted Voronoi diagram, and develop the lambda-complex, the equivalent of the alpha-complex for such a diagram. Second, we use toleranced models to represent protein assemblies. We explain how a toleranced model can be used to assess stable contacts between proteins and to check the coherence between such a model and experimental data. Third, we propose tools to compare graphs encoding contacts within proteins, such graphs coming from a toleranced model on the one hand, and from atomic-resolution models on the other hand. All these concepts and tools are used to probe the aforementioned reconstructions of the NPC.
