We introduce a new class of collective variables which allow forming efficiently beta-sheet structures in all-atom explicit-solvent simulations of proteins. By this approach we are able to systematically fold a 16-residue beta hairpin using metadynamics on a single replica. Application to the 56-residue SH3 and GB1 proteins show that, starting from extended states, in ∼100 ns tens of structures containing more than 30% beta-sheet are obtained, including parts of the native fold. Using these variables may allow folding moderate size proteins with an accurate explicit solvent description. Moreover, it may allow investigating the presence of misfolded states that are relevant for diseases (e.g., prion and Alzheimer) and studying beta-aggregation (amyloid diseases).