Harnessing the Power of Multi-GPU Acceleration into the Quantum Interaction Computational Kernel Program.

We report a new multi-GPU capable ab initio Hartree-Fock/density functional theory implementation integrated into the open source QUantum Interaction Computational Kernel (QUICK) program. Details on the load balancing algorithms for electron repulsion integrals and exchange correlation quadrature across multiple GPUs are described. Benchmarking studies carried out on up to four GPU nodes, each containing four NVIDIA V100-SXM2 type GPUs demonstrate that our implementation is capable of achieving excellent load balancing and high parallel efficiency. For representative medium to large size protein/organic molecular systems, the observed parallel efficiencies remained above 82% for the Kohn-Sham matrix formation and above 90% for nuclear gradient calculations. The accelerations on NVIDIA A100, P100, and K80 platforms also have realized parallel efficiencies higher than 68% in all tested cases, paving the way for large-scale ab initio electronic structure calculations with QUICK.

Parallel Implementation of Density Functional Theory Methods in the Quantum Interaction Computational Kernel Program.

We present the details of a graphics processing unit (GPU) capable exchange correlation (XC) scheme integrated into the open source QUantum Interaction Computational Kernel (QUICK) program. Our implementation features an octree based numerical grid point partitioning scheme, GPU enabled grid pruning and basis and primitive function prescreening, and fully GPU capable XC energy and gradient algorithms. Benchmarking against the CPU version demonstrated that the GPU implementation is capable of delivering an impressive performance while retaining excellent accuracy. For small to medium size protein/organic molecular systems, the realized speedups in double precision XC energy and gradient computation on a NVIDIA V100 GPU were 60-80-fold and 140-500-fold, respectively, as compared to the serial CPU implementation. The acceleration gained in density functional theory calculations from a single V100 GPU significantly exceeds that of a modern CPU with 40 cores running in parallel.

Mechanism of Formation of the Nonstandard Product in the Prenyltransferase Reaction of the G115T Mutant of FtmPT1: A Case of Reaction Dynamics Calling the Shots?

FtmPT1 is a fungal indole prenyltransferase that affords Tryprostatin B from Brevianamide F and dimethylallyl pyrophosphate; however, when a single residue in the active site is mutated (Gly115Thr), a novel five-membered ring compound is obtained as the major product with Tryprostatin B as the minor product. Herein, we describe detailed studies of the catalysis of the Gly115Thr mutant of FtmPT1 with a focus on the observed regioselectivity of the reaction. We employ one- and two-dimensional potential of mean force simulations to explore the catalytic mechanism, along with molecular dynamics simulations exploring the reaction dynamics of the prenyl transfer reaction. Single-point electronic structure calculations were also used to explore the performance of the self-consistent charge density functional tight-binding method to model specific reaction steps. Importantly, we observe that the two reaction pathways have comparable activation parameters and propose that the origin of the novel product is predicated, at least in part, on the topology of the potential energy surface as revealed by reaction dynamics studies.

Acceleration of High Angular Momentum Electron Repulsion Integrals and Integral Derivatives on Graphics Processing Units.

We present an efficient implementation of ab initio self-consistent field (SCF) energy and gradient calculations that run on Compute Unified Device Architecture (CUDA) enabled graphical processing units (GPUs) using recurrence relations. We first discuss the machine-generated code that calculates the electron-repulsion integrals (ERIs) for different ERI types. Next we describe the porting of the SCF gradient calculation to GPUs, which results in an acceleration of the computation of the first-order derivative of the ERIs. However, only s, p, and d ERIs and s and p derivatives could be executed simultaneously on GPUs using the current version of CUDA and generation of NVidia GPUs using a previously described algorithm [Miao and Merz J. Chem. Theory Comput. 2013, 9, 965-976.]. Hence, we developed an algorithm to compute f type ERIs and d type ERI derivatives on GPUs. Our benchmarks shows the performance GPU enable ERI and ERI derivative computation yielded speedups of 10-18 times relative to traditional CPU execution. An accuracy analysis using double-precision calculations demonstrates that the overall accuracy is satisfactory for most applications.

Conformational analysis and parallel QM/MM X-ray refinement of protein bound anti-Alzheimer drug donepezil.

The recognition and association of donepezil with acetylcholinesterase (AChE) has been extensively studied in the past several decades because of the former's use as a palliative treatment for mild Alzheimer disease. Herein we examine the conformational properties of donepezil and we re-examine the donepezil-AChE crystal structure using combined quantum mechanical/molecular mechanical (QM/MM) X-ray refinement tools. Donepezil's conformational energy surface was explored using the M06 suite of density functionals and with the MP2/complete basis set (CBS) method using the aug-cc-pVXZ (X = D and T) basis sets. The donepezil-AChE complex (PDB 1EVE) was also re-refined through a parallel QM/MM X-ray refinement approach based on an in-house ab initio code QUICK, which uses the message passing interface (MPI) in a distributed SCF algorithm to accelerate the calculation via parallelization. In the QM/MM re-refined donepezil structure, coordinate errors that previously existed in the PDB deposited geometry were improved leading to an improvement of the modeling of the interaction between donepezil and the aromatic side chains located in the AChE active site gorge. As a result of the re-refinement there was a 93% reduction in the donepezil conformational strain energy versus the original PDB structure. The results of the present effort offer further detailed structural and biochemical inhibitor-AChE information for the continued development of more effective and palliative treatments of Alzheimer disease.

Acceleration of Electron Repulsion Integral Evaluation on Graphics Processing Units via Use of Recurrence Relations.

Electron repulsion integral (ERI) calculation on graphical processing units (GPUs) can significantly accelerate quantum chemical calculations. Herein, the ab initio self-consistent-field (SCF) calculation is implemented on GPUs using recurrence relations, which is one of the fastest ERI evaluation algorithms currently available. A direct-SCF scheme to assemble the Fock matrix efficiently is presented, wherein ERIs are evaluated on-the-fly to avoid CPU-GPU data transfer, a well-known architectural bottleneck in GPU specific computation. Realized speedups on GPUs reach 10-100 times relative to traditional CPU nodes, with accuracies of better than 1 × 10(-7) for systems with more than 4000 basis functions.

Catalytic mechanism of aromatic prenylation by NphB.

NphB is an aromatic prenyltransferase that catalyzes the attachment of a 10-carbon geranyl group to aromatic substrates. Importantly, NphB exhibits a rich substrate selectivity and product regioselectivity. A systematic computational study has been conducted in order to address several question associated with NphB-catalyzed geranylation. The reaction mechanism of the prenylation step has been characterized as a S(N)1 type dissociative mechanism with a weakly stable carbocation intermediate. A novel π-chamber composed of Tyr121, Tyr216, and 1,6-DHN is found to be important in stabilizing the carbocation. The observed difference in the rates of product formation from 5- and 2-prenylation arises from the differing orientations of the aromatic substrate in the resting state. 4-Prenylation shares the same resting state with 5-prenylation, but the lower free energy barrier for carbocation formation makes the latter reaction more facile. The high free energy barrier associated with 7-prenylation is caused by the unfavorable orientation of 1,6-DHN in active site pocket, along with the difficulty of proton elimination after the prenylation step. A water-mediated proton transfer facilitates the loss of hydrogen at the prenylation site to form the final prenylated product. Interestingly, the same crystallographically observed water molecule has been found to be responsible for proton loss in all three experimentally identified products. After proton transfer, the relaxation of the final product from a sp(3) carbon center to a sp(2) center triggers a "spring-loaded" product release mechanism which pushes the final product out of the binding pocket toward the edge of the active site. The hydrogen bond interactions between the two hydroxyl groups of the aromatic product and the side chains of Ser214 and Tyr288 help to "steer" the movement of the product. In addition, mutagenesis studies identify these same two side chains as being responsible for the observed regioselectivity, particularly 2-prenylation. These observations provide valuable insights into NphB chemistry, offering an opportunity to better engineer the active site and to control the reactivity in order to obtain high yields of the desired product(s). Furthermore, the S(N)1 reaction mechanism observed for NphB differs from the prenylation reaction found in, for example, the farnesyltransferase, which proceeds via an S(N)2-like reaction pathway. The spring-loaded release mechanism highlighted herein also offers novel insights into how enzymes facilitate product release.