AmericanChemicalSociety.com

(1) Graphics Processing Units (GPUs) as performance accelerators: Molecular Dynamics (MD) based modeling of the hydrophobic interactions under the experimentally relevant conditions of the mechanical force application, used in the AFM-based dynamic force measurements, is a formidable task. To obtain a single 10-50ns trajectory of a C16H34–C16H34 bond rupture on a central processor (CPU) in 3-7 days of wall-clock time using all-atom MD simulations, requires 10³-10⁴ times faster force-loads (pulling speeds), compared to the experimental value (~1-10μm/s). Under these conditions the hydrophobic bonds are highly unstable, which lowers the statistical resolution of the rupture forces. This makes it difficult to analyze and model the simulation output and to make connection with the experiment. To overcome these limitations, we utilized Graphics Processing Units (GPUs) used as performance accelerators in a range of scientific applications. CUDA language (Compute Unified Device Architecture) provides a high level programming platform, which allows to perform parallel computations for all particles using many multiprocessors in a single GPU device. Benchmark tests showed that these efforts allow to speedup simulations 100-fold for the C16H34–C16H34 system (~5,000 atoms), and hence, to decrease the pulling speeds 10-100-fold down to ~100μm/s. This has enabled us to resolve the kinetics and thermodynamics of formation and rupture of the hydrophobic bonds. The computational methodology is presented in the paper published in Proteins: Struct. Funct. Bioinform., 78, 2984-2999 (2010).

(2) Resolving Thermodynamics using Replica Exchange methods: Parallel tempering algorithms, applied to formation/rupture of “hydrophobic bonds”, allow to obtain information about the thermodynamics of hydrophobic interactions. To implement these methods on a graphics processor requires a reliable source of pseudo-random numbers, generated on a GPU. While there are some stand-alone methods of generation of random numbers on a GPU, in an MD simulation run, a (pseudo)-random number generator (RNG) should be incorporated into the main simulation kernel (i) to generate the initial distribution of atomic velocities, and (ii) to use the Metropolis criterion to exchange replicas. This requires a large amount of random numbers generated on a GPU device. None of the existing programs for a GPU use fast RNGs of proven statistical quality. We have developed and tested several GPU-based implementations of RNGs, including Ran2, additive Lagged Fibonacci, Hybrid Taus, and several variants of Mersenne Twister algorithm. The results of simulations of Langevin dynamics of a system of N Brownian oscillators (Ornstein-Uhlenbeck process), fully implemented on the GPU, compare very well with the results obtained on the CPU. In addition, the developed GPU-based realizations of RNGs pass stringent statistical tests accumulated in the DIEHARD test suite and in the TestUO1 package. The computational methodology for generation of (pseudo)-random numbers on a GPU is described in the manuscript submitted to the J. Phys. Chem. B.

(3) All-atom Molecular Dynamics simulations of the hydrophobic interactions on the GPU: In the first and second year of this project, we have developed and tested a highly optimized, easy to use GPU-based CUDA software for all-atom MD simulations in explicit water. The optimized CUDA code is now being used to carry out the proposed research activities, described in Specific Aim 1 – Mapping the free energy landscape for the hydrophobic bond rupture, and Specific Aim 2 – Probing the dependence of hydration barrier on the hydrodynamic bond size and direction of stretching force. The results obtained are quite interesting: (i) Formation of “hydrophobic bonds” involves continuous reversible transitions between many intermediate states, rather than discrete states, during which the C16H34 molecules slide and wind around each other, forming overlaps of 4-10 carbon atoms. (ii) When subject to pulling force, the mechanical rupture of hydrophobic interactions is an “all-or-none” process. (iii) We observe the formation of a short-lived solvent separated state (intermediate state, SSM). In the absence of force, the system leaves the SSM state and returns back to the bound state (CM), i.e. SSM→CM; however, when the force is applied, the transition, SSM→U, to the globally unbound state (U) is more frequent than the reverse process (SSM→CM). (iv) The free energy landscape alters appreciably with size of the hydrophobic contact. These results are now being used to estimate the kinetic and thermodynamic parameters, and to resolve the free energy landscape underlying the hydrophobic interactions between short hydrocarbon chains. 

(4) Broader Impact: GPU-based calculations represent one of the major directions of development in petascale computing in theoretical and computational chemistry. Recent technological advances in the throughput-oriented hardware architecture of GPUs unleashed tremendous computational power that has been utilized in the calculations of electronic structure, ab initio quantum chemistry calculations, and quantum Monte Carlo simulations. Future directions might include GPU-based computing in the mixed-quantum classical description of chemical processes in condensed phases. The long-range goals of my career are (i) to create optimized implementations of molecular simulations on alternative computing systems, which might include GPU-based and/or combined CPU-GPU-based platforms. This will benefit the scientific community at large, as it (ii) will help reduce the time-to-solution of today’s challenging problems. As a graduate advisor and educator, I thrive (iii) to generate a synergy between mentoring, teaching, and research, where the study of fundamental processes motivates learning. Graduate and undergraduate students, working in my lab, design advanced computational models and develop novel theoretical and computational approaches to tackle scientific problems.

(5) Gradiate Students: During the second year of this project, Mr. Artem Zhmurov, a chemistry graduate (PhD) student working in my research group, has received advanced training in programming on graphics processors using CUDA language (dialect of C and C++), and in software testing and optimization. This far, Mr. Zhmurov has gained experience necessary to fully implement the all-atom MD simulations in explicit water running on the GPU device. Mr. Zhmurov is now working to generate a sufficient amount of the simulation data. In the meantime, these data are being analyzed by the PI to resolve the kinetics and thermodynamics of the hydrophobic bond rupture/formation, as proposed in the grant application.