1. Protein folding/unfolding kinetics (~ 2006.02)
My past research in Korea was focused on finding important topological quantities from a protein structure that are highly related with folding/unfolding kinetics. Various topological quantities are extracted after making graphs corresponding to the protein structure. From the correlation between these quantities and folding/unfolding rates, it is observed that unfolding is different from the reverse process of folding. I introduced a topological quantity named “the impact of edge removal” that has a high correlation with the unfolding rate, revealing that protein unfolding process is highly dependent on the native protein topology rather than sequence information (J. Jung et al., Proteins, 58, 389-395 (2005)). More recently, another topological quantity compensating the previous investigation of protein folding kinetics has been suggested (J. Jung et al., J. Sol. Chem., 39, 943-958 (2010)
2. Development of QM/MM program (2006.03~2010.11)
My first research topic in Japan was the development QM/MM program. Molecular mechanics (MM) based on force field have limitations in describing chemical reactions . This feature is ameliorated by introducing a quantum mechanical (QM) treatment in the calculation. Many kinds of local problems (e.g. the binding of metal ions to a protein, enzyme catalysis by active site residue) can be understood by QM. However, a full QM calculation needs high computational cost, and thus a hybrid QM/MM method is a feasible way for large molecules. In particular, QM/MM is a powerful means to study chemical processes in biological systems with the inclusion of the protein or solvent environment.
The most important is the treatment of a boundary between QM and MM regions and their interactions. In the given period, I implemented the Generalized Hybrid Orbital (GHO) method in the quantum chemistry program GELLAN. In this method, a set of sp3 hybridized orbitals is placed on a boundary atom between the QM and MM regions. One of the hybridized orbitals participates in the QM region while the rest are frozen. With this method, more smooth connection between two regions is available. On the basis of the GHO method, I introduced several techniques: from single point energy to geometry optimization and minimum reaction paths. For fast evaluation of QM/MM, I also implemented MPI parallelization inside the QM program.
3. Molecular Dynamics (MD) development with efficient parallelization (2010.12~)
1) Development of MD on K supercomputer
After moving to RIKEN Advanced Institute for Computational Science (RIKEN AICS), I developed optimization of MD software, GENESIS, on K supercomputer. In particular, I implemented efficient parallelization and optimization of the time-comsing part. It includes lookup table instead of direct calculation, new hybrid (MPI+OpenMP) parallelization scheme, parallelization of fast Fourier transform (FFT), and so on. Programming was done from the scratch, and it showed world best parallel efficiency among existing MD programs. Due to the program and K supercomputer, a biological system consisting of 100 million atoms could be simulated.
2) Development of MD on heterogeneous computers
I also developed MD on heterogeneous computers, such as CPU with GPGPU by coworking with NVIDIA company. In this development, computation-intensive interactions are assigined on GPUs while communication-intensive oens are on CPUs. With the multiple time step integration development, it shows very good scalability on multiple CPU+GPU nodes. I also noticed that computational algorithm should be changed from hardware architecture to optimize the performance, and prepared multi-kernel scheme to choose the best algorithm from given hardware architecture.
3) Development of MD on Fugaku supercomputer
I was involved in the development of MD on Fugaku in recent a few years. Because Fugaku and K have many different characteristics, I optimized the MD program by coworking with Fujitsu company. I also developed several algorithms to increase the speed by introducing large time step method from accurate temperature and pressure evaluations. Recently, I’m developing parallelization for coarse-grained MD.