Peter L. Freddolino

Accelerating molecular modeling applications with graphics processors

Molecular mechanics simulations offer a computational approach to study the behavior of biomolecules at atomic detail, but such simulations are limited in size and timescale by the available computing resources. State‐of‐the‐art graphics processing units (GPUs) can perform over 500 billion arithmetic operations per second, a tremendous computational resource that can now be utilized for general purpose computing as a result of recent advances in GPU hardware and software architecture. In this article, an overview of recent advances in programmable GPUs is presented, with an emphasis on their application to molecular mechanics simulations and the programming techniques required to obtain optimal performance in these cases. We demonstrate the use of GPUs for the calculation of long‐range electrostatics and nonbonded forces for molecular dynamics simulations, where GPU‐based calculations are typically 10–100 times faster than heavily optimized CPU‐based implementations. The application of GPU acceleration to biomolecular simulation is also demonstrated through the use of GPU‐accelerated Coulomb‐based ion placement and calculation of time‐averaged potentials from molecular dynamics trajectories. A novel approximation to Coulomb potential calculation, the multilevel summation method, is introduced and compared with direct Coulomb summation. In light of the performance obtained for this set of calculations, future applications of graphics processors to molecular dynamics simulations are discussed.

Ten-Microsecond Molecular Dynamics Simulation of a Fast-Folding WW Domain

All-atom molecular dynamics (MD) simulations of protein folding allow analysis of the folding process at an unprecedented level of detail. Unfortunately, such simulations have not yet reached their full potential both due to difficulties in sufficiently sampling the microsecond timescales needed for folding, and because the force field used may yield neither the correct dynamical sequence of events nor the folded structure. The ongoing study of protein folding through computational methods thus requires both improvements in the performance of molecular dynamics programs to make longer timescales accessible, and testing of force fields in the context of folding simulations. We report a ten-microsecond simulation of an incipient downhill-folding WW domain mutant along with measurement of a molecular time and activated folding time of 1.5 microseconds and 13.3 microseconds, respectively. The protein simulated in explicit solvent exhibits several metastable states with incorrect topology and does not assume the native state during the present simulations.

Prediction of structure and function of G protein-coupled receptors

G protein-coupled receptors (GPCRs) mediate our sense of vision, smell, taste, and pain. They are also involved in cell recognition and communication processes, and hence have emerged as a prominent superfamily for drug targets. Unfortunately, the atomic-level structure is available for only one GPCR (bovine rhodopsin), making it difficult to use structure-based methods to design drugs and mutation experiments. We have recently developed first principles methods (MembStruk and HierDock) for predicting structure of GPCRs, and for predicting the ligand binding sites and relative binding affinities. Comparing to the one case with structural data, bovine rhodopsin, we find good accuracy in both the structure of the protein and of the bound ligand. We report here the application of MembStruk and HierDock to β1-adrenergic receptor, endothelial differential gene 6, mouse and rat I7 olfactory receptors, and human sweet receptor. We find that the predicted structure of β1-adrenergic receptor leads to a binding site for epinephrine that agrees well with the mutation experiments. Similarly the predicted binding sites and affinities for endothelial differential gene 6, mouse and rat I7 olfactory receptors, and human sweet receptor are consistent with the available experimental data. These predicted structures and binding sites allow the design of mutation experiments to validate and improve the structure and function prediction methods. As these structures are validated they can be used as targets for the design of new receptor-selective antagonists or agonists for GPCRs.

Force Field Bias in Protein Folding Simulations

Long timescale (>1 micros) molecular dynamics simulations of protein folding offer a powerful tool for understanding the atomic-scale interactions that determine a proteins folding pathway and stabilize its native state. Unfortunately, when the simulated protein fails to fold, it is often unclear whether the failure is due to a deficiency in the underlying force fields or simply a lack of sufficient simulation time. We examine one such case, the human Pin1 WW domain, using the recently developed deactivated morphing method to calculate free energy differences between misfolded and folded states. We find that the force field we used favors the misfolded states, explaining the failure of the folding simulations. Possible further applications of deactivated morphing and implications for force field development are discussed.

Common Structural Transitions in Explicit-Solvent Simulations of Villin Headpiece Folding

Molecular dynamics simulations of protein folding can provide very high-resolution data on the folding process; however, due to computational challenges most studies of protein folding have been limited to small peptides, or made use of approximations such as Gō potentials or implicit solvent models. We have performed a set of molecular dynamics simulations totaling >50 micros on the villin headpiece subdomain, one of the most stable and fastest-folding naturally occurring proteins, in explicit solvent. We find that the wild-type villin headpiece reliably folds to a native conformation on timescales similar to experimentally observed folding, but that a fast folding double-norleucine mutant shows significantly more heterogeneous behavior. Along with other recent simulation studies, we note the occurrence of nonnative structures intermediates, which may yield a nativelike signal in the fluorescence measurements typically used to study villin folding. Based on the wild-type simulations, we propose alternative approaches to measure the formation of the native state.

Molecular mechanism of ligand recognition by NR3 subtype glutamate receptors

NR3 subtype glutamate receptors have a unique developmental expression profile, but are the least well‐characterized members of the NMDA receptor gene family, which have key roles in synaptic plasticity and brain development. Using ligand binding assays, crystallographic analysis, and all atom MD simulations, we investigate mechanisms underlying the binding by NR3A and NR3B of glycine and D‐serine, which are candidate neurotransmitters for NMDA receptors containing NR3 subunits. The ligand binding domains of both NR3 subunits adopt a similar extent of domain closure as found in the corresponding NR1 complexes, but have a unique loop 1 structure distinct from that in all other glutamate receptor ion channels. Within their ligand binding pockets, NR3A and NR3B have strikingly different hydrogen bonding networks and solvent structures from those found in NR1, and fail to undergo a conformational rearrangement observed in NR1 upon binding the partial agonist ACPC. MD simulations revealed numerous interdomain contacts, which stabilize the agonist‐bound closed‐cleft conformation, and a novel twisting motion for the loop 1 helix that is unique in NR3 subunits.

The role of molecular modeling in bionanotechnology

Molecular modeling is advocated here as a key methodology for research and development in bionanotechnology. Molecular modeling provides nanoscale images at atomic and even electronic resolution, predicts the nanoscale interaction of unfamiliar combinations of biological and inorganic materials, and evaluates strategies for redesigning biopolymers for nanotechnological uses. The methodology is illustrated in this paper through reviewing three case studies. The first one involves the use of single-walled carbon nanotubes as biomedical sensors where a computationally efficient, yet accurate, description of the influence of biomolecules on nanotube electronic properties through nanotube-biomolecule interactions was developed; this development furnishes the ability to test nanotube electronic properties in realistic biological environments. The second case study involves the use of nanopores manufactured into electronic nanodevices based on silicon compounds for single molecule electrical recording, in particular, for DNA sequencing. Here, modeling combining classical molecular dynamics, material science and device physics, described the interaction of biopolymers, e.g., DNA, with silicon nitrate and silicon oxide pores, furnished accurate dynamic images of pore translocation processes, and predicted signals. The third case study involves the development of nanoscale lipid bilayers for the study of embedded membrane proteins and cholesterol. Molecular modeling tested scaffold proteins, redesigned apolipoproteins found in mammalian plasma that hold the discoidal membranes in the proper shape, and predicted the assembly as well as final structure of the nanodiscs. In entirely new technological areas such as bionanotechnology, qualitative concepts, pictures and suggestions are sorely needed; these three case studies document that molecular modeling can serve a critical role in this respect, even though it may still fall short on quantitative precision.

Simulations of a Protein Crystal with a High Resolution X-ray Structure: Evaluation of Force Fields and Water Models

We use classical molecular dynamics and 16 combinations of force fields and water models to simulate a protein crystal observed by room-temperature X-ray diffraction. The high resolution of the diffraction data (0.96 Å) and the simplicity of the crystallization solution (nearly pure water) make it possible to attribute any inconsistencies between the crystal structure and our simulations to artifacts of the models rather than inadequate representation of the crystal environment or uncertainty in the experiment. All simulations were extended for 100 ns of production dynamics, permitting some long-time scale artifacts of each model to emerge. The most noticeable effect of these artifacts is a model-dependent drift in the unit cell dimensions, which can become as large as 5% in certain force fields; the underlying cause is the replacement of native crystallographic contacts with non-native ones, which can occur with heterogeneity (loss of crystallographic symmetry) in simulations with some force fields. We find that the AMBER FF99SB force field maintains a lattice structure nearest that seen in the X-ray data, and produces the most realistic atomic fluctuations (by comparison to crystallographic B-factors) of all the models tested. We find that the choice of water model has a minor effect in comparison to the choice of protein model. We also identify a number of artifacts that occur throughout all of the simulations: excessive formation of hydrogen bonds or salt bridges between polar groups and loss of hydrophobic interactions. This study is intended as a foundation for future work that will identify individual parameters in each molecular model that can be modified to improve their representations of protein structure and thermodynamics.

Newly Identified Genetic Variations in Common Escherichia coli MG1655 Stock Cultures

We have recently identified seven mutations in commonly used stocks of the sequenced Escherichia coli strain MG1655 which do not appear in the reference sequence. The mutations are likely to cause loss of function of the glpR and crl genes, which may have serious implications for physiological experiments using the affected strains.

Going beyond clustering in MD trajectory analysis: an application to villin headpiece folding.

Recent advances in computing technology have enabled microsecond long all-atom molecular dynamics (MD) simulations of biological systems. Methods that can distill the salient features of such large trajectories are now urgently needed. Conventional clustering methods used to analyze MD trajectories suffer from various setbacks, namely (i) they are not data driven, (ii) they are unstable to noise and changes in cut-off parameters such as cluster radius and cluster number, and (iii) they do not reduce the dimensionality of the trajectories, and hence are unsuitable for finding collective coordinates. We advocate the application of principal component analysis (PCA) and a non-metric multidimensional scaling (nMDS) method to reduce MD trajectories and overcome the drawbacks of clustering. To illustrate the superiority of nMDS over other methods in reducing data and reproducing salient features, we analyze three complete villin headpiece folding trajectories. Our analysis suggests that the folding process of the villin headpiece is structurally heterogeneous.