The NVIDIA GPU Award Finalists for San Francisco (Fall 2014)

Post date: Apr 21, 2014 6:39:13 PM

Acknowledgements: This work was supported by the National Natural Science Foundation of China [21073195, 21373227] and China's State Key Laboratory of Multiphase Complex Systems [MPCS-2012-A-05].

References

1. van Duin, A. C. T.; Dasgupta, S.; Lorant, F.; Goddard, W. A., J. Phys. Chem. A 2001, 105 (41), 9396-9409.

2. Zheng, M.; Li, X.; Guo, L., J. Mol. Graphics Modell. 2013, 41, 1-11.

3. Zheng, M. et al. Energy Fuels 2014, 28 (1), 522-534.

Harnessing the power of FEP with GPU cloud computing

Levi CT Pierce. Schrödinger, Department of Research and Methods Development, New York City, New York, 11102, United States

Obtaining highly accurate, reliable, and consistent binding free energies, across a diverse range of drug targets has been difficult to achieve for computational tools. Free energy perturbation (FEP) is one of the few tools that has been shown to achieve highly accurate binding free energies, but you have to pay a high computational cost. In the past this price equated to acquiring 1024 CPU cores on a massive cluster and now this same compute power is achieved by 4 Nvidia Tesla GPU's running under my desk. The high compute price for this technique limited exploration of its applicability and reliability across diverse ligand sets and protein targets. In this work we demonstrate our FEP/REST (Replica Exchange Solute Tempering) workflow, which harnesses the power of the NVidia GPU using a custom kernel engineered to achieve unprecedented throughput. Using our workflow engine we are now able to run more than 200 perturbations a day on our internal Tesla powered GPU cluster, allowing us to iteratively improve the method and explore new domains for FEP, antibody design, covalent ligands, macrocyles, etc. Finally, we demonstrate how this can be scaled up to 1000 perturbations a day on the Amazon EC2 cloud using the NVidia GPU Grid, which has allowed us, for the first time, to examine how reliable and consistent binding energies can be predicted across numerous protein targets. 

Unraveling the fumarate addition reaction in the glycyl radical enzyme Benzylsuccinate Synthase: A GPU enabled comprehensive computational study

Vivek S Bharadwaj, Anthony M Dean, and C. Mark Maupin. Colorado School of Mines, Department of Chemical and Biological Engineering, Golden, Colorado, 80401, United States

Benzylsuccinate Synthase (BSS) is a glycyl radical enzyme that catalyzes the fumarate addition reaction which enables anaerobic bio-degradation of hydrocarbons in bacterial cultures. The energetically challenging nature of the fumarate addition reaction and the fact that it is catalyzed by a free radical mechanism, has intrigued researchers in recent times. However, the extreme sensitivity of the glycyl radical to oxygen has precluded structural characterization and detailed experimental investigation of BSS at the molecular level. Here, we present a systematic and comprehensive computational approach involving homology modeling, docking, GPU enabled molecular dynamics (MD) simulations, molecular mechanical generalized Born surface area (MMGBSA) calculations and umbrella sampling techniques to unravel the molecular basis for the fumarate addition reaction in BSS. The active site topology of BSS was elucidated using homology modeling and docking techniques while the overall stability of the free enzyme and the dynamics of substrates (toluene and fumaric acid) bound at the active site was analyzed using MD simulations, performed on GPUs using the cuda version of pmemd in AMBER. The results demonstrate the experimentally observed syn addition of toluene to fumaric acid as well as pin-point specific protein-substrate interactions that stabilize the substrates at the active site. The umbrella sampling simulations also reveal that substrate binding favorably impacts the active site dynamics so as to enable feasible radical transfer pathways for the proposed fumarate addition reaction mechanism.

Protonated RNA Structures are Not Static: WExploring the pH-dependent Ensemble of RNA

Garrett B Goh,1 Alex Dickson,1 Kamali Sripathi,2 Nils G Walter,1,2,3 and Charles L Brooks, III.1,3 1. University of Michigan, Department of Chemistry, Ann Arbor, MI, 48109, United States, 2. University of Michigan, Department of Medicinal Chemistry, Ann Arbor, MI, 48109, United States, and 3. University of Michigan, Biophysics Program, Ann Arbor, MI, 48109, United States.

pH-regulation of RNA-mediated biological activity is an emerging discovery. Traditionally, the analysis of static structures solved near physiological pH is used to understand the structure-function relationship of such pH-regulated activity. While this approach may be appropriate for proteins, RNA have rugged energy landscapes, so conformational dynamics and other near-native structures cannot be ignored. One example is the U6 internal stem-loop (ISL) of the spliceosome, which undergoes a pH-dependent base flipping conformational change. Extracting such structural information is experimentally non-trivial, which drives the need to develop better tools that can describe the full extent of pH-dependent structural ensembles of RNA. In this work, we present an application of WExplore, a weighted ensemble sampling method enhanced with a hierarchical weight-balancing algorithm that has been implemented with GPU-accelerated CHARMM. We surveyed a range of protonated RNA structures that are implicated in pH-dependent biological activities, such as ribozymes, the spliceosome and other RNA pseudoknot structures. Using GPU-accelerated WExplore sampling, we identified other conformations under high and low pH conditions, and when combined with constant pH MD (CPHMD) simulations, we show how the RNA conformation modulates the microscopic pKa of each state. Interestingly, most of our results using “ground state” crystallographic or “lowest energy” NMR structures produced pKa values that were inconsistent with experimental pKa. However, when the structural data obtained from WExplore sampling was integrated in our CPHMD simulations, we were able to reproduce experimentally observed pKa values, which suggests that using a pH-dependent structural ensemble that encompasses low population minor states is a necessary condition for a fundamental understanding of protonated RNA structures. In conclusion, our work provides the first comprehensive validation that protonated RNA structures have multiple physiologically-relevant conformations, which may not be fully captured by experimental characterization.

Overcoming the Sampling Problem in Force Field Evaluation via GPU-accelerated Multi-dimensional Replica Exchange Molecular Dynamics

Christina Bergonzo,1 Niel M. Henriksen,1 Daniel R. Roe,1 Jason M. Swails,2 Adrian E. Roitberg,2 and Thomas E. Cheatham, III1 1. University of Utah, Department of Medicinal Chemistry, College of Pharmacy, Salt Lake City, UT, 84112, United States and 2. University of Florida, Department of Chemistry, Gainesville, FL, 32611, United States.

To assess and validate force field performance in biomolecular simulation, there is a need to exhaustively sample all accessible conformational space. In doing so, we decouple errors in the potential function parameters from errors due to incomplete sampling. This is challenging for large structures in general, and for nucleic acids in particular due to their highly charged state and high conformational variability. In this study we used RNA tetranucleotides which display considerable conformational variability, making it challenging yet tractable to generate the complete structural ensemble. Using multidimensional replica exchange methods (M-REMD), which are implemented in AMBER's GPU-accelerated code, and in concert with the large scale GPU resources NCSA Blue Waters and XSEDE Keeneland and Stampede, we have completely sampled the conformational ensemble of GACC. Using 192 replicas, each run on independent GPUs, we show M-REMD generates the converged ensemble within and between independent simulations. The advantage of using GPU accelerated code is that we can perform 50-80 μs aggregate sampling per run, allowing us to converge ensembles with different force fields and compare populations to experimental NMR data. The results of this work will be used to evaluate current force field reliability and propose modifications to better reproduce experimental data.