The use of molecular simulation to estimate the strength of macromolecular binding free energies is becoming increasingly widespread, with goals ranging from lead optimization and enrichment in drug discovery to personalizing or stratifying treatment regimes. In order to realize the potential of such approaches to predict new results, not merely to explain previous experimental findings, it is necessary that the methods used are reliable and accurate, and that their limitations are thoroughly understood. However, the computational cost of atomistic simulation techniques such as molecular dynamics (MD) has meant that until recently little work has focused on validating and verifying the available free energy methodologies, with the consequence that many of the results published in the literature are not reproducible. Here, we present a detailed analysis of two of the most popular approximate methods for calculating binding free energies from molecular simulations, molecular mechanics Poisson–Boltzmann surface area (MMPBSA) and molecular mechanics generalized Born surface area (MMGBSA), applied to the nine FDA-approved HIV-1 protease inhibitors. Our results show that the values obtained from replica simulations of the same protease–drug complex, differing only in initially assigned atom velocities, can vary by as much as 10 kcal mol–1, which is greater than the difference between the best and worst binding inhibitors under investigation. Despite this, analysis of ensembles of simulations producing 50 trajectories of 4 ns duration leads to well converged free energy estimates. For seven inhibitors, we find that with correctly converged normal mode estimates of the configurational entropy, we can correctly distinguish inhibitors in agreement with experimental data for both the MMPBSA and MMGBSA methods and thus have the ability to rank the efficacy of binding of this selection of drugs to the protease (no account is made for free energy penalties associated with protein distortion leading to the over estimation of the binding strength of the two largest inhibitors ritonavir and atazanavir). We obtain improved rankings and estimates of the relative binding strengths of the drugs by using a novel combination of MMPBSA/MMGBSA with normal mode entropy estimates and the free energy of association calculated directly from simulation trajectories. Our work provides a thorough assessment of what is required to produce converged and hence reliable free energies for protein–ligand binding.

Accurate calculation of important thermodynamic properties, such as macromolecular binding free energies, is one of the principal goals of molecular dynamics simulations. However, single long simulation frequently produces incorrectly converged quantitative results due to inadequate sampling of conformational space in a feasible wall-clock time. Multiple short (ensemble) simulations have been shown to explore conformational space more effectively than single long simulations, but the two methods have not yet been thermodynamically compared. Here we show that, for end-state binding free energy determination methods, ensemble simulations exhibit significantly enhanced thermodynamic sampling over single long simulations and result in accurate and converged relative binding free energies that are reproducible to within 0.5 kcal/mol. Completely correct ranking is obtained for six HIV-1 protease variants bound to lopinavir with a correlation coefficient of 0.89 and a mean relative deviation from experiment of 0.9 kcal/mol. Multidrug resistance to lopinavir is enthalpically driven and increases through a decrease in the protein-ligand van der Waals interaction, principally due to the V82A/I84V mutation, and an increase in net electrostatic repulsion due to water-mediated disruption of protein-ligand interactions in the catalytic region. Furthermore, we correctly rank, to within 1 kcal/mol of experiment, the substantially increased chemical potency of lopinavir binding to the wild-type protease compared to saquinavir and show that lopinavir takes advantage of a decreased net electrostatic repulsion to confer enhanced binding. Our approach is dependent on the combined use of petascale computing resources and on an automated simulation workflow to attain the required level of sampling and turn around time to obtain the results, which can be as little as three days. This level of performance promotes integration of such methodology with clinical decision support systems for the optimization of patient-specific therapy.

We describe computational science research that uses petascale resources to achieve scientific results at unprecedented scales and resolution. The applications span a wide range of domains, from investigation of fundamental problems in turbulence through computational materials science research to biomedical applications at the forefront of HIV/AIDS research and cerebrovascular haemodynamics. This work was mainly performed on the US TeraGrid 'petascale' resource, Ranger, at Texas Advanced Computing Center, in the first half of 2008 when it was the largest computing system in the world available for open scientific research. We have sought to use this petascale supercomputer optimally across application domains and scales, exploiting the excellent parallel scaling performance found on up to at least 32 768 cores for certain of our codes in the so-called 'capability computing' category as well as highthroughput intermediate-scale jobs for ensemble simulations in the 32-512 core range. Furthermore, this activity provides evidence that conventional parallel programming with MPI should be successful at the petascale in the short to medium term. We also report on the parallel performance of some of our codes on up to 65 636 cores on the IBM Blue Gene/P system at the Argonne Leadership Computing Facility, which has recently been named the fastest supercomputer in the world for open science.

Large scale supercomputing applications typically run on clusters using vendor message passing libraries, limiting the application to the availability of memory and CPU resources on that single machine. The ability to run inter-cluster parallel code is attractive since it allows the consolidation of multiple large scale resources for computational simulations not possible on a single machine, and it also allows the conglomeration of small subsets of CPU cores for rapid turnaround, for example, in the case of high-availability computing. MPIg is a grid-enabled implementation of the Message Passing Interface (MPI), extending the MPICH implementation of MPI to use Globus Toolkit services such as resource allocation and authentication. To achieve coavailability of resources, HARC, the Highly-Available Resource Co-allocator, is used. Here we examine two applications using MPIg: LAMMPS (Large-scale Atomic/ Molecular Massively Parallel Simulator), is used with a replica exchange molecular dynamics approach to enhance binding affinity calculations in HIV drug research, and HemeLB, which is a lattice-Boltzmann solver designed to address fluid flow in geometries such as the human cerebral vascular system. The cross-site scalability of both these applications is tested and compared to single-machine performance. In HemeLB, communication costs are hidden by effectively overlapping non-blocking communication with computation, essentially scaling linearly across multiple sites, and LAMMPS scales almost as well when run between two significantly geographically separated sites as it does at a single site.