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

Molecular dynamics simulations are crucial for investigating biomolecular mechanisms. The success of these simulations hinges on the accuracy, efficiency, and generalizability of the underlying force field. While classical molecular force fields are efficient yet approximate and quantum mechanics is accurate but computationally prohibitive for large systems, machine learning force fields (MLFFs) have emerged to bridge this gap. We review various MLFFs-from classically parametrized to end-to-end models-evaluating their performance in accuracy and efficiency. However, a significant challenge for MLFFs is generalizability as models trained on specific data often fail to extrapolate to unseen molecules or conformations. To address this, universal MLFFs, such as fragment-based methods like AI²BMD designed by Wang et al. and GEMS designed by Unke et al., are being developed. Beyond recent progress, we also discuss the inherent limitations and trade-offs of MLFFs. Looking forward, the integration of MLFFs with virtual cell models and coarse-grained representations is poised to enable whole-cell multiscale simulations.