Learning molecular dynamics with simple language model built upon long short-term memory neural network.

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Recurrent neural networks (RNNs) have advanced natural language processing and speech recognition.
• The application of RNNs to complex chemical and biophysical systems remains underexplored.

Purpose of the Study:

• To demonstrate that long short-term memory (LSTM) networks can effectively model the temporal evolution of chemical/biophysical trajectories.
• To establish LSTMs as a tool for understanding complex stochastic molecular systems.

Main Methods:

• Developed a character-level language model using LSTMs to learn probabilistic models of 1D stochastic trajectories.
• Trained the LSTM on benchmark systems and a force spectroscopy trajectory from a multi-state riboswitch.

Main Results:

• The LSTM model successfully captured Boltzmann statistics and reproduced kinetics across various timescales.
• Training the LSTM was shown to be equivalent to learning path entropy.
• The LSTM's embedding layer revealed nontrivial connectivity between metastable states in the physical system.

Conclusions:

• LSTMs are capable of capturing the dynamics of complex stochastic molecular systems.
• This work provides a novel application of RNNs in computational chemistry and biophysics.
• The findings represent a significant step towards utilizing machine learning for molecular dynamics analysis.