Project

GromEx: A combined lambda-dynamics/fast multipole method for enhanced molecular simulations in GROMACS

Goal: Life is sustained by large biomolecular nanomachines. A common feature of these systems is that their mechanism often involves uptake and release of ligands, most prominently protons and electrons, by chemically variable sites. The most ubiquitous example are protonatable aminoacid sidechains followed by a plethora of biological cofactors that can occur in different protonation, tautomeric or redox forms. Also drug molecules ,that influence biomolecular function, often occur in multiple, chemically distinct forms.

We thus have to account for this chemical variability in our simulations. However, current simulation software often does not allow to include this physical detail at all or only in a limited or in a difficult to use way.
Here, we are implementing a very general variant of lambda-dynamics to add the missing physical detail for physically sound simulations on the thermodynamics of biomolecular function. An envisioned, additional important application is computational drug design for automated screening of many ligands in parallel as already demonstrated by Charles Brooks' and coworkers. Local, alternative topologies of the ligands that amalgamate traditional single- and multi-topology setups will enhance
versatility. Our implementation will also make available multiple different
lambda-dynamics variants in a single free open source software for direct and transparent comparability.

Besides chemical variability, a second common feature of biomacromolecular machines is their large size that ranges from a nanometer to many micrometers resulting in simulation system sizes that can easily reach 100,000s to many 1,000,000s of particles with a few, over 100s to 10,0000s of chemically variable sites. The computation of long-range interactions, first and foremost electrostatic
interactions is the computationally most demanding part of a molecular simulation. This demand is exacerbated by adding chemically variable sites, because the interactions between all pairs of forms of distinct sites need to be computed in addition to the usual interactions.
To address this issues, we are implementing a fast multipole method (FMM) with support for lambda-dynamics and highly parallel high performance computation. The FMM also alleviates the so-called communication wall that limits the number of compute devices that can be fruitfully used in a simulation with the currently most popular particle-mesh-Ewald method (PME) because it requires communication
between all pairs of compute devices. In effect, the reduction in computation time upon adding more compute devices to a simulation diminishes more and more with growing resource allocation size up to
the point where the time required to exchange data between the compute nodes becomes larger than the time spent in the actual computation. Consequently, the computational efficiency approaches zero and adding even more resources just heats the environment and burns taxpayers money.
Harnessing the full power of future exascale supercomputers, that are predicted to offer 10,000s to millions of CPU cores and likely accelerators like GPUs, will thus require reducing the communication to a minimum. The FMM allows for this reduction and is well parallelizable.
Our parallelization approach is reasonably fine grained task parallelism.
This approach enables automatic balancing of the compute load between all available devices.
This feature will be indispensable for making efficient use of an entire non-homogeneous cluster in a single simulation. Such heterogeneous clusters grow over time from different hardware generations in many
molecular simulation labs, but may also become more frequently employed in compute centers.

The attached video (see project log) shows hyper-sphere lambda dynamics, a variant developed in our project that can be used for any number of site forms. Other variants, including Nexp lambda-dynamics developed by Brooks and coworkers is also implemented but more difficult to visualize. Here, we place three protonation forms of a carboxylic acid sidechain on a sphere. Each site form has a conjugate lambda variable which ranges from zero, where the form is fully absent to one, where it is fully present. The regions of the sphere surface that correspond to the different site forms are indicated by the colored
isocontours of the lambda values. Moving 90 degrees on the sphere surface carries the system to the next site form. The current position of the lambda-particle is indicated by the circle with the crosshair. As the particle moves on the sphere, the structures of the forms on the right appear and disappear according to their lambda value and the color of the lambda-particle changes accordingly if one of the forms becomes dominant. The center shows a 2D-projection of the sphere surface. Angles are just like longitude and latitude on a world map, left connects to right but here also top to bottom. The dark shading indicates the equilibrium probability, which concentrates at the physical end states, where all lambda variables approach values of zero or one. The black arrows indicate transitions between pairs of forms, which are favored relative to transitions involving three (or more) forms, as you can see from the trajectory. Population of the regions close to the physical forms and the preference for the transitions between pairs of forms over transitions involving more forms is achieved through a tailored bias potential. The main innovation of the bias potential is that it directly penalizes the number of effectively populated states Omega. K denotes the barrier height K between two site forms. The shape parameter s, tunes the relative width of minima and barriers These two parameters allow us to finely tune our simulation for a balance between physical purity and sampling efficiency.

Updates
0 new
1
Recommendations
0 new
0
Followers
0 new
15
Reads
0 new
183

Project log

R. Thomas Ullmann
added an update
R. Thomas Ullmann
added a project goal
Life is sustained by large biomolecular nanomachines. A common feature of these systems is that their mechanism often involves uptake and release of ligands, most prominently protons and electrons, by chemically variable sites. The most ubiquitous example are protonatable aminoacid sidechains followed by a plethora of biological cofactors that can occur in different protonation, tautomeric or redox forms. Also drug molecules ,that influence biomolecular function, often occur in multiple, chemically distinct forms.
We thus have to account for this chemical variability in our simulations. However, current simulation software often does not allow to include this physical detail at all or only in a limited or in a difficult to use way.
Here, we are implementing a very general variant of lambda-dynamics to add the missing physical detail for physically sound simulations on the thermodynamics of biomolecular function. An envisioned, additional important application is computational drug design for automated screening of many ligands in parallel as already demonstrated by Charles Brooks' and coworkers. Local, alternative topologies of the ligands that amalgamate traditional single- and multi-topology setups will enhance
versatility. Our implementation will also make available multiple different
lambda-dynamics variants in a single free open source software for direct and transparent comparability.
Besides chemical variability, a second common feature of biomacromolecular machines is their large size that ranges from a nanometer to many micrometers resulting in simulation system sizes that can easily reach 100,000s to many 1,000,000s of particles with a few, over 100s to 10,0000s of chemically variable sites. The computation of long-range interactions, first and foremost electrostatic
interactions is the computationally most demanding part of a molecular simulation. This demand is exacerbated by adding chemically variable sites, because the interactions between all pairs of forms of distinct sites need to be computed in addition to the usual interactions.
To address this issues, we are implementing a fast multipole method (FMM) with support for lambda-dynamics and highly parallel high performance computation. The FMM also alleviates the so-called communication wall that limits the number of compute devices that can be fruitfully used in a simulation with the currently most popular particle-mesh-Ewald method (PME) because it requires communication
between all pairs of compute devices. In effect, the reduction in computation time upon adding more compute devices to a simulation diminishes more and more with growing resource allocation size up to
the point where the time required to exchange data between the compute nodes becomes larger than the time spent in the actual computation. Consequently, the computational efficiency approaches zero and adding even more resources just heats the environment and burns taxpayers money.
Harnessing the full power of future exascale supercomputers, that are predicted to offer 10,000s to millions of CPU cores and likely accelerators like GPUs, will thus require reducing the communication to a minimum. The FMM allows for this reduction and is well parallelizable.
Our parallelization approach is reasonably fine grained task parallelism.
This approach enables automatic balancing of the compute load between all available devices.
This feature will be indispensable for making efficient use of an entire non-homogeneous cluster in a single simulation. Such heterogeneous clusters grow over time from different hardware generations in many
molecular simulation labs, but may also become more frequently employed in compute centers.
The attached video (see project log) shows hyper-sphere lambda dynamics, a variant developed in our project that can be used for any number of site forms. Other variants, including Nexp lambda-dynamics developed by Brooks and coworkers is also implemented but more difficult to visualize. Here, we place three protonation forms of a carboxylic acid sidechain on a sphere. Each site form has a conjugate lambda variable which ranges from zero, where the form is fully absent to one, where it is fully present. The regions of the sphere surface that correspond to the different site forms are indicated by the colored
isocontours of the lambda values. Moving 90 degrees on the sphere surface carries the system to the next site form. The current position of the lambda-particle is indicated by the circle with the crosshair. As the particle moves on the sphere, the structures of the forms on the right appear and disappear according to their lambda value and the color of the lambda-particle changes accordingly if one of the forms becomes dominant. The center shows a 2D-projection of the sphere surface. Angles are just like longitude and latitude on a world map, left connects to right but here also top to bottom. The dark shading indicates the equilibrium probability, which concentrates at the physical end states, where all lambda variables approach values of zero or one. The black arrows indicate transitions between pairs of forms, which are favored relative to transitions involving three (or more) forms, as you can see from the trajectory. Population of the regions close to the physical forms and the preference for the transitions between pairs of forms over transitions involving more forms is achieved through a tailored bias potential. The main innovation of the bias potential is that it directly penalizes the number of effectively populated states Omega. K denotes the barrier height K between two site forms. The shape parameter s, tunes the relative width of minima and barriers These two parameters allow us to finely tune our simulation for a balance between physical purity and sampling efficiency.