Microbial communities are complex systems composed of various species of microorganisms interacting with each other and with their environment. Systems biology offers a framework for their study, combining experimentation, high-throughput data generation and the integration of the latter into computer models. Understanding these ecosystems requires the analysis of their metabolism and the exchange of molecules between members, which can have positive or negative impacts on each other. Metabolism is a set of biochemical reactions that can be represented as a genome-scale metabolic network through the association of genes and reactions in an organism. These networks can be used to build metabolic models, mathematical representations of organism behavior under given environmental conditions. To scale-up to the analysis of a community, composed of a few species under controlled conditions, or several hundreds uin natural environments, raises methodological difficulties in the construction of models. This thesis manuscript deals with the construction of computational models for the analysis of metabolism and metabolic interactions in microbial ecosystems, with a particular attention on explaining the cellular mechanisms underlying bacterial interactions. Numerical solutions are mainly used - ensuring the accuracy of results - but they face combinatorial issues generated by bacterial interactions in large-scale communities. Discrete approaches overcome this problem, but are limited to pairwise analysis. In order to identify a putative methodological tradeoff reconciling the advantages of both approaches, i.e. finding a hybrid approach, a first contribution focuses on the development of a dynamic and accurate numerical model of a cheese bacterial community composed of three strains. Our iterative strategy enables the integration of heterogeneous data through refinement and dynamic calibration steps. This back-and-forth between knowledge and model has ensures the accurate prediction of metabolite concentrations and bacterial densities during cheese production. We propose as a second contribution a reasoning-based model for deciphering cooperative and competitive potentials in bacterial communities. This model relies on the inference of logical rules motivated from biology in order to evaluate and compare community interaction potentials. Ecosystem-specific interactions potentials are retrieved, and the fast execution of the resoning-based approach facilitates the screening of collections of communities. Finally, the third contribution is a reflection on the enrichment of the logic model.  We propose a prototype based on the inference of logical rules, enabling (i) the selection of the best community based on biological constraints and (ii) the inference of a temporal notion, which can impact interaction potentials. Through this thesis, we demonstrate that the construction of a hybrid model of metabolism is not required but that a hybrid approach, using numerical models for small communities and discrete models for rapid analysis of full-size communities, seems to be relevant.