Abstract : Phylogenetic networks generalize the tree concept to model Evolution, by allowing edges between branches inside the tree to reflect genetic material exchanges between coexisting species. Lots of combinatorial approaches have been designed to reconstruct networks from data extracted from a set of contradictory gene trees. These approaches can be divided into several categories depending on the kind of input, i.e. triplets, quartets, clusters and splits, and on the kind of structure restrictions they impose on reconstructed networks. We particularly analyze the structure of one class of such restricted networks, namely level-k phylogenetic networks, and adapt this level parameter to the unrooted context. We also give new combinatorial methods to reconstruct phylogenetic networks from clusters - implemented in Dendroscope - or quartets. We study the limits of combinatorial methods (complexity explosion, noise and silence in the data, ambiguity in the reconstucted network), and the way to tackle them, in particular with an appropriate data preprocessing. Finally we illustrate the results of these reconstruction methods on a dataset, and we conclude on how to use them in a global methodology which integrates statistical aspects.