In nature, some proteins partially unfold under specific environmental conditions. These unfolded states typically consist of a large ensemble of conformations; their proper description is therefore a challenging problem. NMR spectroscopy is particularly well suited for this task: information on conformational preferences can be derived, for example, from chemical shifts or residual dipolar couplings. This information, which is measured as a time- and ensemble-average, can be used to model these states by generating large ensembles of conformations. The challenge is then to select a minimum representative set of conformations out of a large ensemble to represent the unfolded state. We have developed for this purpose an algorithm called MINOES (MINimum Optimal Ensemble Selection), which is based on an iterative procedure based on a driven expansion/contraction selection process. MINOES aims at selecting an optimal and minimal ensemble of conformations that, on average, maximizes the agreement between back-calculated and experimental (NMR) data, without any a-priori assumption about the required ensemble size. This approach is demonstrated by modeling the partially unfolded state of a deletion mutant of the Photoactive Yellow Protein, Delta25-PYP, which has been previously characterized by NMR (Bernard et al., Structure 2005;13:953-962).