Final Report Summary - ERIKLINDAHLERC2007 (Multiscale and Distributed Computing Algorithms for Biomolecular Simulation and Efficient Free Energy Calculations)
Molecular simulation an important research technique in several research areas where it provides an amazing virtual microscope, and as highlighted by the 2013 Nobel Prize in Chemistry, simulations are now used by experimental researchers in general, and biomolecular life science in particular. However, to advance the state-of-the to the point where simulations rapidly and reliably provide molecular explanations to disease and accurate drug design requires orders of magnitude higher performance and new techniques to accurately predict experimental observables.
The present project has advanced molecular dynamics significantly on three different levels. First, it has produced the state-of-the-art in accelerating simulations on modern coprocessors such as GPUs (graphics cards). In version 4.6 of our GROMACS simulation code (http://www.gromacs.org ), the GPU acceleration provides almost 40X higher performance than a single CPU core, and even when using all processors in a workstation the GPU-accelerated simulations are a factor 5 faster. To the best of our knowledge, this is the fastest implementation in any code to date, and with the help of the ERC project it is freely available to thousands of researchers. We have also improved parallelization, which means GROMACS now scales to the point where each core only has ~100 particles - a tenfold improvement since 2008.
Second, to provide accurate, efficient, and easily accessible methods for free energy calculations, we developed a new implementation of free energy calculations based on the “Bennett Acceptance Ratio” formalism where the analysis is performed on-the-fly, which improves accuracy and avoids the huge amounts of data previously required, so this type of calculations can be used efficiently in distributed and cloud computing environments. In addition, we have automated the creation of free energy input data and introduced proper statistical analysis based on sampling efficiency of the ensembles in simulations, which provides standard error estimates for all free energy calculations.
Third, single simulations will never scale to billions of processors on next-generation supercomputers. Instead, we combined the discoveries above with our experience from distributed computing and introduced the idea of “parallel adaptive molecular dynamics”. For this, we developed a new framework (Copernicus, http://www.copernicus-computing.org ) that automatically schedules thousands of tightly coupled simulations that each can use thousands of cores on supercomputers and improve sampling by orders-of-magnitude even for very small molecules such as 30-40 residue proteins. Not only can we now fold these in single days, but we get statistically sound estimates of folding rates. We have also introduced adaptive simulations for free energy calculations; these rely on a sequence of 20-40 simulations as a function of a coupling parameter, but by performing short series of pre-simulations we can optimize the statistical overlap and reduce the number of simulations required by a factor 2-3. Finally, we have created new tools tool that automatically generate simulation parameters for small molecules, which enables Copernicus to calculate solvation and binding free energies of large series of compounds in mere hours, at a cost of a fraction of a euro per compound.
These advances have helped us perform a number of important simulations of membrane proteins and other systems. We have published the first models of intermediate states of a voltage-gated ion channel (covering an entire gating cycle), it enabled us to calculate the free energy as a large segment of the voltage sensor moves in the electric field, and it helped show how this segment assumes a 3-10 helix conformation during gating. We can move the protein between intermediate and experimental states in the simulations, and the free energy calculations predict how experimental properties change during gating due to mutations in the channels. We have invested similar efforts in ligand-gated ion channels, where our simulations were the first results to show that allosteric modulators such as alcohol and anesthetics bind between subunits, and through a series of simulations we have developed a model where a binding site inside each subunit would decrease channel activity, while the site between subunits amplify it. In the very last months of the project, we have been able to provide very strong evidence for this dual-site allosteric modulation mechanism based on a combination of state-of-the-art binding free energy calculations combined with experimental collaborations, which could help design new pairs of on/off anesthetic drugs.
The present project has advanced molecular dynamics significantly on three different levels. First, it has produced the state-of-the-art in accelerating simulations on modern coprocessors such as GPUs (graphics cards). In version 4.6 of our GROMACS simulation code (http://www.gromacs.org ), the GPU acceleration provides almost 40X higher performance than a single CPU core, and even when using all processors in a workstation the GPU-accelerated simulations are a factor 5 faster. To the best of our knowledge, this is the fastest implementation in any code to date, and with the help of the ERC project it is freely available to thousands of researchers. We have also improved parallelization, which means GROMACS now scales to the point where each core only has ~100 particles - a tenfold improvement since 2008.
Second, to provide accurate, efficient, and easily accessible methods for free energy calculations, we developed a new implementation of free energy calculations based on the “Bennett Acceptance Ratio” formalism where the analysis is performed on-the-fly, which improves accuracy and avoids the huge amounts of data previously required, so this type of calculations can be used efficiently in distributed and cloud computing environments. In addition, we have automated the creation of free energy input data and introduced proper statistical analysis based on sampling efficiency of the ensembles in simulations, which provides standard error estimates for all free energy calculations.
Third, single simulations will never scale to billions of processors on next-generation supercomputers. Instead, we combined the discoveries above with our experience from distributed computing and introduced the idea of “parallel adaptive molecular dynamics”. For this, we developed a new framework (Copernicus, http://www.copernicus-computing.org ) that automatically schedules thousands of tightly coupled simulations that each can use thousands of cores on supercomputers and improve sampling by orders-of-magnitude even for very small molecules such as 30-40 residue proteins. Not only can we now fold these in single days, but we get statistically sound estimates of folding rates. We have also introduced adaptive simulations for free energy calculations; these rely on a sequence of 20-40 simulations as a function of a coupling parameter, but by performing short series of pre-simulations we can optimize the statistical overlap and reduce the number of simulations required by a factor 2-3. Finally, we have created new tools tool that automatically generate simulation parameters for small molecules, which enables Copernicus to calculate solvation and binding free energies of large series of compounds in mere hours, at a cost of a fraction of a euro per compound.
These advances have helped us perform a number of important simulations of membrane proteins and other systems. We have published the first models of intermediate states of a voltage-gated ion channel (covering an entire gating cycle), it enabled us to calculate the free energy as a large segment of the voltage sensor moves in the electric field, and it helped show how this segment assumes a 3-10 helix conformation during gating. We can move the protein between intermediate and experimental states in the simulations, and the free energy calculations predict how experimental properties change during gating due to mutations in the channels. We have invested similar efforts in ligand-gated ion channels, where our simulations were the first results to show that allosteric modulators such as alcohol and anesthetics bind between subunits, and through a series of simulations we have developed a model where a binding site inside each subunit would decrease channel activity, while the site between subunits amplify it. In the very last months of the project, we have been able to provide very strong evidence for this dual-site allosteric modulation mechanism based on a combination of state-of-the-art binding free energy calculations combined with experimental collaborations, which could help design new pairs of on/off anesthetic drugs.