Russian

Graphics processors have evolved over the last few years into highly parallel, multithreaded computational devices. Unlike mainstream processor architecture, graphics processing units (GPUs) devote the majority of their logic units to performing actual computations, rather than to cache memory and flow control. Consequently, the use of graphics processing units (GPUs) in the general purpose scientific applications offers a significant speed gain as compared to the similar computations but on the central processing units (CPUs). Traditionally, graphics cards have been designed for rendering of complex imagery where all the fragments of an image undergo identical mathematical transformation (translations, rotations, etc.) in real time. In molecular simulations, too, all particles in a system are subject to the covalent interactions (vibrations, bond angle bending) and to forces due to the non-covalent coupling (electrostatic interactions, van der Waals interactions, dihedral and improper angles), which are computed using the same energy function for all atoms (force-field). This makes a numerical solution of a biological N-body problem into a prime candidate for implementation on the GPU device.

In our research group, we develop GPU-based implementations of the molecular simulations, including all-atom Molecular Dynamics (MD) simulations in implicit solvent and in explicit solvent (water), and Langevin Dynamics (LD) simulations. We work on developing new algorithms, e.g., for the generation of (pseudo-)random numbers, numerical integration of equations of motion, and computations of forces on the GPU architecture. The developed methodology for GPU-based computations is fully utilized in our lab to energyze and speedup the computational studies of a range of biological systems and processes. We are now heavily involved in the development of numerical algorithms for MD simulations in implicit water. Unfortunately, these methods allow mostly for the theoretical exploration of equilibrium properties of biomolecules, while reaching the biologically important millisecond-to-second timescale is virtually impossible even for a small system. For these reasons, we also develop numerical algorithms for LD simulations of biomolecules which are based on simplified coarse-grained models. We have tested the GPU-based implementation of LD simulations using a C_α-based coarse-grained Self Organized Polymer (SOP) model of the protein chain. Combining SOP model with GPU-based computations (SOP-GPU package) enables us to carry out dynamic force measurements in silico for a range of biological systems using experimental pulling speeds. These efforts help us compare directly the results of experiments and simulations, and to interpret the experimental data.