Spring Shooting in OpenPathSampling¶
Authors: Sander Roet
This module implements the spring shooting method in OpenPathSampling
Transition path sampling is most efficient when paths are generated from the top of the free energy barrier. However, complex (biomolecular) activated processes, such as nucleation or protein binding/unbinding, can have asymmetric and peaked barriers. Using uniform selection on these type of processes will not be efficient, as it, on average, results in selected points that are not on the top of the barrier. Paths generated from these points have a low acceptance probability and accepted transition paths decorrelate slowly, resulting in a low overall efficiency. Spring shooting was developed to increase the efficiency of path sampling of these types of barriers, without any prior knowledge of the barrier shape. The spring shooting algorithm uses a shooting point selector that is biased with a spring potential. This bias pulls the selection of points towards the transition state at the top of the barrier. The paths that are generated from points selected by this biased selector therefore have an increased acceptance probability and decorrelation between accepted transition paths is also increased. This results in a higher overall efficiency. The spring shooting algorithm is described by Brotzakis and Bolhuis (http://dx.doi.org/10.1063/1.4965882).
In summary, the spring shooting selection algorithm is a selector for the one-way shooting method for transition path sampling, which uses a bias. This bias is of the shape . Where for forward shooting and for backward shooting, is the given spring constant and is the number of shifted frames of the new shooting point compared to the previous accepted shooting point . The choice of is limited to the interval . The shooting move is rejected if a is selected that is outside of the current path and is accepted if the trajectory satisfies the path ensemble.
The main difference of this module compared to the paper is that instead of using a rejection algorithm to sample from the correct distribution, the correct distribution is sampled directly.
The implementation introduces the following new classes:
ShootingPointSelectorand implements the shooting point selection, using a bias of the shape . At initialization it also takes an
initial_guessas the initial reference point. This will default to
floor(len(trajectory)/2). To correctly keep track of the history the selector has to be the same instance for the forward and backward mover. The
pickfunction has to be called with a direction in order to be able to use the correct value for . The selector then draws a random number in the range , multiplies it with the sum of the biases. It the sums the biases from to and returns the index when this sum is bigger than the random number. The
restart_from_stepfunction makes it possible to restart the selector at a specific step. It takes an
MCStepand reconstructs the correct history from the
EngineMoverand is the parent class for the
BackwardSpringMoverclasses. It calls the
self.directionto get the correct shooting point. It also build adds the needed
MoveDetailsin order to be able to restart the selector. If an invalid snapshot has been chosen it will not run dynamics but will build a
Samplewith an acceptance probability of 0.0.
SpecializedRandomChoiceMoverand behaves similar, except it takes the extra arguments:
initial_guess. These are then used to make the
SpringShootingSelectorand this selector is given to both the forward and backward mover.
SingleEnsembleMoveStrategyand will make the
SpringShootingMoverfor every ensemble, with the given values of
MoveSchemeand it will use the
SpringShootingStrategyto build all the necessary movers for the given
network, with the given values of
initial_guess. Also the functions
from_dicthave been adapted to save and load with all the data
This module builds on OpenPathSampling, a Python package for path sampling simulations. To learn more about OpenPathSampling, you might be interested in reading:
- OPS documentation: http://openpathsampling.org
- OPS source code: http://github.com/openpathsampling/openpathsampling
This module has been included in the OpenPathSampling core. Its tests can
be run by setting up a developer install of OpenPathSampling and running
py.test from the root directory of the repository.
- An example on how to set up a simulation using the
SpringShootingMoveSchemeand the comparison with the
UniformSelectorcan be found in
examples/miscdirectory (https://github.com/openpathsampling/openpathsampling/tree/master/examples/misc). Open it using
jupyter notebook spring_shooting_example.ipynb(see
Jupyter notebookdocumentation at http://jupyter.org/ for more details)
This module has been merged into OpenPathSampling. It was added in the following pull request: