Accelerated Monte-Carlo Simulations for All-Atom Protein Folding

Abstract

Recent advances in the computational description of proteins facilitate the atomic-level characterization of fast folding proteins using molecular-dynamics simulations[1]. However, due to the femtosecond timestep of an MD simulation, 10 ˆ 10 computationally expensive energy evaluations are required to reach timescales of the order of single folding events even for ultrafast-folding proteins. Stochastic simulation techniques can overcome this limiting factor by generating thermodynamic conformational ensembles without incorporating atomic vibrations, but most of these methods are inherently sequential. Here we investigate a parallel extension to the conventional Metropolis-Hastings algorithm, the Multiple-Try-Metropolis, enabling parallel simulation of reversible Markov-chains[2,3]. We successfully characterize the thermodynamic landscape of the Villin-headpiece-subdomain using a generalized MTM approach adapted to protein systems. We parameterized the polypeptide using the AMBER99SB-STAR-ILDN-forcefield with an implicit solvent model. In our simulations we could observe approximately two folding transitions per day on standard hardware, with a total of 109 transition events during 3∗10 ˆ 9 energy evaluations. The MTM approach thus facilitates the efficient thermodynamic characterization of proteins on common computer architectures. The comparison with experimental resolved folding times of 8µs translates to a time equivalent of 16ps per MTM step.[1] Shaw, Science-2010.[2] Liu, JASA-2000.[3] Pandolfi, JMLR-2010.View Large Image | View Hi-Res Image | Download PowerPoint Slide