The comparison of biological networks is now one of the most promising approaches that help in understanding the functioning of living organisms. It appears as the expected continuation of the comparison of biological sequences, whose study represents in reality only the genomic aspect of the information manipulated by biologists. In this thesis, we propose an innovative approach allowing to compare two biological networks modeled respectively by a directed graph D and an undirected graph G, and provided with a correspondence function f between the vertices of both graphs. The approach consists in extracting automatically a biologically significant structure in D whose vertices induce in G a biologically significant structure as well. We realize an algorithmic study of the problem arising in our approach by starting with its variant in which D is acyclic (DAG). We provide polynomial algorithms for several cases and we show that other cases are algorithmically difficult (NP-completes). In order to solve the difficult instances, we propose a reliable heuristic and an exact algorithm based on the branch-and-bound method. To deal with the case where D is cyclic, we introduce a method motivated by biological hypotheses and consisting in decomposing D into DAGs such that the vertices of each DAG induce in G a connected subgraph. We also study in this thesis, the problem of signaling pathways inference by combining the information on causes and effects of extra-cellular events. We model this problem by a problem of mixed graphs orientation and we perform a complexity study allowing to identify the easy and the difficult instances.
