Sequential Monte Carlo Methods for Protein Folding

We describe a class of growth algorithms for finding low energy states of heteropolymers. These polymers form toy models for proteins, and the hope is that similar methods will ultimately be useful for finding native states of real proteins from heuristic or a priori determined force fields. These algorithms share with standard Markov chain Monte Carlo methods that they generate Gibbs-Boltzmann distributions, but they are not based on the strategy that this distribution is obtained as stationary state of a suitably constructed Markov chain. Rather, they are based on growing the polymer by successively adding individual particles, guiding the growth towards configurations with lower energies, and using "population control" to eliminate bad configurations and increase the number of "good ones". This is not done via a breadth-first implementation as in genetic algorithms, but depth-first via recursive backtracking. As seen from various benchmark tests, the resulting algorithms are extremely efficient for lattice models, and are still competitive with other methods for simple off-lattice models.