Abstract
Molecular Dynamics (MD) simulations provide accurate descriptions of the motions of molecular systems, yet their computational demands pose significant challenges in applications in molecular biology and materials science. Given the success of deep learning methods in a wide range of fields, a timely question concerns whether these methods could be leveraged to improve the efficiency of MD simulations. To investigate this possibility, we introduce Molecular Dynamics Language Models (MDLMs), to enable the generation of MD trajectories. In the present implementation, an MDLM is trained on a short classical MD trajectory of a protein, where structural accuracy is maintained through kernel density estimations derived from extensive MD datasets. We illustrate the application of this MDLM in the case of the determination of the free energy landscape a small protein, showing that this approach makes it possible to discover conformational states undersampled in the training data. These results provide initial evidence for the use of language models for the efficient implementation of molecular dynamics.
Competing Interest Statement
The authors have declared no competing interest.
