The European Water Framework Directive implementation implies to assess, then to maintain or restore the good ecological status of aquatic ecosystems in the Member States. Hence, distinguishing natural ecological variations from human related ones becomes crucial: the present work aims to focus on the biogeographical frame of diatom assemblages within natural or near-natural conditions. First, 233 diatom samples collected from unimpacted or low impacted sites of the French hydrographic network were classified according to similarities of their assemblages using a Self Organizing Map algorithm. The results showed 5 different natural community types, corresponding to specific environmental conditions making sense with the existing French hydro-ecoregions (Wasson et al., 2002). Second, those communities were predicted with a set of environmental parameters through a multilayer perceptron (MLP) with backpropagation learning algorithm. The predictability was compared with a discriminant function analysis and a regression tree, and the best results were obtained with MLP. The importance of environmental variables to predict diatom community types was evaluated through a sensitivity analysis of MLP model. In conclusion it seems possible for each French hydro-ecoregion to predict how the community should be like out of anthropogenic pressure. According to the WFD requirements, this work gives a first qualitative idea of the diatom assemblages representative of the good ecological status, taking into account ecoregional particularities of natural environment.