Utility of a time-lagged autoencoder for calculating free energy by generating a large number of synthetic trajectories based on molecular dynamics
by Jamali, S.; Haque, F.; Moradi, M.
Molecular dynamics have been used for the examination of the properties of molecules. Depending on the size of the system and other relevant parameters, a regular molecular dynamics simulation can be time consuming and costly since it needs to solve Newtonian equations of motions for a large number of atoms and simulation time steps. As the system becomes more complex, the accuracy of the free energy calculations tends to diminish. Due to the large application of machine learning in different areas specifically in nonlinear complex models, in this study, the time-lagged autoencoder (TAE) algorithm, an unsupervised machine learning algorithm, is used for calculating the free energy with more accuracy based on generating a large number of synthetic trajectories based on molecular dynamics simulation. As opposed to the numerous calculations involved in traditional molecular dynamics, the time-lagged autoencoder algorithm offers significant time savings by reducing the dimensionality of the problem. We have examined the working of the algorithm by varying different parameters involved in the TAE algorithm such as the number of nodes and layers that have a significant impact on the performance of the algorithm. This proposed that novel use of the TAE algorithm provides a straightforward method of using deep learning in amplifying rare events and estimating accurate free energies in biomolecular systems.