Employing Artificial Neural Networks for Optimal Storage and Facile Sharing of Molecular Dynamics Simulation Trajectories

With the remarkable stride in computing power and advances
in Molecular
Dynamics (MD) simulation programs, the crucial challenge of storing
and sharing large biomolecular simulation data sets has emerged. By
leveraging AutoEncoders, a type of artificial neural network, we developed
a method to compress MD trajectories into significantly smaller latent
spaces. Our method can save up to 98% in disk space compared to xtc, a highly compressed trajectory format from the widely
used MD program package GROMACS, thus facilitating storage and sharing
of simulation trajectories. Atom coordinates are very accurately reconstructed
from compressed data. The method was tested across a diverse sets
of biomolecular systems, including folded proteins, intrinsically
disordered proteins, phospholipid bilayers, protein–ligand
complexes, large protein complexes and membrane-bound protein systems.
The reconstructed trajectories demonstrated consistent accuracy in
recovering key biophysically relevant properties for proteins, lipids
and composite systems. The compression efficiency was particularly
beneficial for larger systems. This approach enables the scientific
community to efficiently store and share large-scale biomolecular
simulation data, potentially enhancing collaborative research efforts.
The workflow, termed “compresstraj”, is implemented
in PyTorch and is publicly available at https://github.com/SerpentByte/compresstraj, offering a practical solution for handling the increasing volumes
of data generated in biomolecular simulation studies.