CGU: An Algorithm for Molecular Structure Prediction

A global optimization method is presented for predicting the minimum energy structure of small protein-like molecules. This method begins by collecting a large number of molecular conformations, each obtained by finding a local minimum of a potential energy function from a random starting point. The information from these conformers is then used to form a convex quadratic global underestimating function for the potential energy of all known conformers. This underestimator is an L1 approximation to all known local minima, and is obtained by a linear programming formulation and solution. The minimum of this underestimator is used to predict the global minimum for the function, allowing a localized conformer search to be performed based on the predicted minimum. The new set of conformers generated by the localized search serves as the basis for another quadratic underestimation step in an iterative algorithm. This algorithm has been used to predict the minimum energy structures of heteropolymers with as many as 48 residues, and can be applied to a variety of molecular models. The results obtained also show the dependence of the native conformation on the sequence of hydrophobic and polar residues.