Determining how proteins interact to form stable complexes is of crucial importance, for example in the development of novel therapeutics. Computational methods to determine the thermodynamically stable conformation of complexes from the structure of the binding partners, such as RosettaDock, might potentially emerge to become a promising alternative to traditional structure determination methods. However, while models virtually identical to the correct experimental structure can in some cases be generated, the main difficulty remains to discriminate correct or approximately correct models from decoys. This is due to the ruggedness of the free-energy landscape, the approximations intrinsic in the scoring functions, and the intrinsic flexibility of proteins. Here we show that molecular dynamics simulations performed starting from a number top-scoring models can not only discriminate decoys and identify the correct structure, but may also provide information on an initial map of the free energy landscape that elucidates the binding mechanism.