Recent advances in artificial intelligence-driven biomolecular dynamics simulations based on machine learning force...

Area of Science:

• Computational Chemistry
• Biophysics
• Materials Science

Background:

• Molecular dynamics (MD) simulations are vital for understanding biomolecular mechanisms.
• The accuracy, efficiency, and generalizability of force fields are critical for MD simulation success.
• Classical force fields are efficient but approximate; quantum mechanics is accurate but computationally expensive.

Purpose of the Study:

• To review various machine learning force fields (MLFFs) and evaluate their performance.
• To discuss the challenge of generalizability in MLFFs.
• To explore future directions for MLFFs in multiscale simulations.

Main Methods:

• Review of existing literature on MLFFs, from classically parametrized to end-to-end models.
• Evaluation of MLFF performance based on accuracy and efficiency metrics.
• Discussion of fragment-based methods and universal MLFFs for improved generalizability.

Main Results:

• MLFFs bridge the gap between classical and quantum mechanics for MD simulations.
• Generalizability remains a significant challenge for MLFFs, limiting extrapolation to new data.
• Universal MLFFs like AI²BMD and GEMS are being developed to enhance generalizability.

Conclusions:

• MLFFs show great promise for biomolecular simulations but require further development in generalizability.
• Future integration with virtual cell and coarse-grained models will enable large-scale, multiscale simulations.
• Addressing MLFF limitations and trade-offs is key for advancing computational biophysics.