In Other Journals Science

The ultimate objective of biomolecular modeling is to provide molecular dynamics simulations with quantum-mechanical accuracy at system sizes and timescales relevant to realistic applications. Current approaches often face trade-offs among efficiency, accuracy, scalability, and transferability. Machine learning force fields have successfully bridged the gap between some of these factors, but achieving truly general molecular simulations remains elusive. Kabylda et al. report a pretrained neural network and universal pairwise force fields that demonstrated robust performance in nanosecond-long simulations of small biomolecular systems, high transferability throughout biochemical space, and scalability to hundreds of thousands of atoms. Although many biomolecular processes are currently beyond reach, the proposed method is a promising step toward the long-standing goal of accurate large-scale modeling across extended chemical space.