Convex Global Underestimation for Molecular Structure Prediction

Key problems in computational biology, including protein and RNA folding and drug docking, involve conformational searching. Current search methods — Monte Carlo, Molecular Dynamics, Simulated Annealing, and Genetic Algorithms — are too slow for protein folding by many orders of magnitude. They get stuck in kinetic traps. We describe a global optimization method, the CGU method, which appears to be very promising. We know the method always finds the same conformation from 100 different starting points, indicating that it finds the unique global minimum for the many different sequences we have tried. We know the CGU doesn’t get stuck in kinetic traps because the search time is independent of the shapes of the landscapes (amino acid sequence and composition). We know that the method is much faster than a standard Simulated Annealing algorithm that we have tested: the SA method doesn’t find global minima for chains longer than 10 residues, and the performance advantage of the CGU method increases with chain length. And computational results show that the computer time scales with n 4 where n is the number of degrees of freedom, and we consistently reach the global minimum of the model energy function for PPT, a 36-amino acid peptide (n = 72), in less than 3 hours on a 32 processor Cray T3E.