Over nearly five decades, biomolecular simulation techniques have become indispensable tools for uncovering fundamental principles governing life at molecular scale. These computational approaches enable quantitative characterization of complex biomolecular events including protein folding, conformational dynamics and protein‑protein interactions, and have greatly advanced both fundamental and applied biological research. In recent years, the integration of machine learning, especially deep‑learning algorithms, has sparked substantial innovation within this field.

Recently, a team from Wenfei Li and Wei Wang at Nanjing University published a perspective article titled "Advances and challenges in multiscale biomolecular simulations: artificial intelligence‐driven paradigm shift" in Quantitative Biology. The authors discuss state‑of‑the‑art technical advances of biomolecular simulations, and explore emerging applications, development trends and major open challenges in biomolecular dynamics simulations.

Biological processes at the molecular level arise from biomolecular assembly, motion and interactions. To decode biological function, we need both high‑resolution static 3D structures and mechanistic knowledge of molecular dynamics. Though tools like AlphaFold2 have revolutionized static protein‑structure prediction, experimental dynamic data is scarce, restricting AI‑based dynamics prediction. Molecular dynamics (MD) simulations fill this gap. Classical all‑atom MD computes atomic trajectories via empirical force fields, yet exhaustive conformational sampling is hampered by vast degrees of freedom and complex energy landscapes. Large‑scale events such as protein‑machinery assembly, amyloid aggregation and liquid‑liquid phase separation (LLPS) are computationally out‑of‑reach for standard atomistic simulations. While coarse‑grained (CG) models cut computational load by merging atoms into virtual particles, building reliable CG force fields remains challenging. This well‑known accuracy‑efficiency bottleneck may be relieved by AI algorithms, as illustrated in Figure 1, where Figure 1A outlines the multiscale hurdles spanning different spatial sizes for biomolecular systems and Figure 1B demonstrates the workflow combining AI models with molecular simulations to tackle these obstacles. This perspective reviews key AI‑driven research directions: quantum‑accurate machine‑learning force fields (MLFFs); bottom‑up and top‑down AI coarse‑grained modelling; integrative MD constrained by experimental and AI‑generated structural data; MD‑trained generative AI for protein conformational ensembles; and expanded applications covering glycosylation, proteome‑ and near‑cellular‑scale simulations. It also addresses community progress in FAIR‑aligned data sharing and open‑source simulation software. The authors highlight major open challenges: poor transferability of learned potentials, lack of high‑quality training data, difficulties simulating quantum‑driven reactions and protonation dynamics, inadequate CG sampling, and the need for more unified computational tool chains.
DOI
10.1002/qub2.70024