Protein flexibility is essential for enzymatic function, ligand binding, and protein-protein or protein-nucleic acid interactions. Normal mode analysis has increasingly been shown to be well suited for studying such flexibility, as it can be used to identify favorable structural deformations that correspond to functional motions. However, normal modes are strictly relevant to a single structure, reflecting a particular minimum on a complex energy surface, and are thus susceptible to artifacts. We describe a new theoretical framework for determining "consensus" normal modes from a set of related structures, such as those issuing from a short molecular dynamics simulation. This approach is more robust than standard normal mode analysis, and provides higher collectivity and symmetry properties. In an application to HIV-1 protease, the low-frequency consensus modes describe biologically relevant motions including flap opening and closing that can be used in interpreting structural changes accompanying the binding of widely differing inhibitors.