Conformational restriction by fragment assembly and guidance in molecular dynamics are alternate conformational search strategies in protein structure prediction. We examine both approaches using a version of the associative memory Hamiltonian that incorporates the influence of water-mediated interactions (AMW). For short proteins (<70 residues), fragment assembly, while searching a restricted space, compares well to molecular dynamics and is often sufficient to fold such proteins to near-native conformations (4Å) via simulated annealing. Longer proteins encounter kinetic sampling limitations in fragment assembly not seen in molecular dynamics which generally samples more native-like conformations. We also present a fragment enriched version of the standard AMW energy function, AMW-FME, which incorporates the local sequence alignment derived fragment libraries from fragment assembly directly into the energy function. This energy function, in which fragment information acts as a guide not a restriction, is found by molecular dynamics to improve on both previous approaches.fragment assembly ͉ associative memory Hamiltonian ͉ protein folding ͉ annealing ͉ molecular dynamics I t is useful to categorize protein structure prediction schemes into two classes: template-based modeling and de novo prediction. Template-based modeling depends on the existence, and identification of, at least one experimentally-solved structure with significant global structural similarity to the target to be predicted, usually a sequence homolog. The identification can be made either by a global sequence-sequence alignment or a global sequencestructure alignment (1). Finding the proper template is a search problem but unlike folding, a search highly restricted to a relatively modest number of possibilities. After finding a template, the homolog structure acts as a global constraint which again severely restricts the remainder of the relevant conformational space to be searched. This leads overall to a much simpler optimization problem to solve. Various energy functions can be used which often lead to successful predictions defined by significant improvement relative to input homolog information (2).However, modeling a protein structure when no experimentallydetermined homologs exist to match the structures globally (or none are recognized to exist) is quite challenging. Such de novo structure prediction can employ all-atom molecular mechanics or hybrid models. Molecular mechanics methods are based on physico-chemical interactions such as van der Waals, electrostatics, hydrogen bonding, solvation energy, and basic backbone steric constraints (covalent bond lengths and angles and torsion angle preferences). Model parameters are generally inferred from experimental measurements and/or quantum chemical calculations on small organic molecules (3, 4). Based on such data, one can generate a transferable energy function (5). The resulting energy function can be used in a variety of search procedures, including template-based modeling. Ultimately, a...