Protein folding and binding is commonly depicted as a search for the minimum energy conformation in a vast energy landscape. Indeed, modelling of protein complex structures by RosettaDock often results in a set of low-energy conformations near the native structure. Ensembles of low-energy conformations can appear, however, in other regions of the energy landscape, especially when backbone movements occur upon binding. What then characterizes the energy landscape near the correct orientation? We have applied a machine learning algorithm to distinguish ensembles of low-energy conformations around the native conformation from other low-energy ensembles. FunHunt, the resulting classifier, identified the native orientation for 50/52 protein complexes in a test set, and for all of 12 recent CAPRI targets. FunHunt is also able to choose the near-native orientation among models created by algorithms other than RosettaDock, demonstrating its general applicability for model selection. The features used by FunHunt teach us about the nature of native interfaces. Remarkably, the energy decrease of trajectories toward near-native orientations is significantly larger than for other orientations. This provides a possible explanation for the stability of association in the native orientation. The FunHunt approach, discriminating models based on ensembles of structures that map the nearby energy landscape, can be adapted and extended to additional tasks, such as ab initio model selection, protein interface design and specificity predictions.