Computer simulations, particularly molecular dynamics (MD) simulations, are effective tools for studying microscopic and dynamic details of complex systems governing various physicochemical and biological processes. Recent advancements in computer hardware and algorithms have significantly improved the efficiency of MD methods and expanded their applicability to the treatment of large macromolecular assemblies and the understanding of intricate mechanisms of biomolecular processes. The MD method relies on the integration of the equations of motion to explore the conformational space and eventually for the evaluation of free energy surfaces. The evaluation of the forces acting on the various particles composing the system represents the main computational bottleneck in MD simulations with an increase of computational cost in the order of N², where N is the number of particles in the system. For this reason, the explicit representation of all atoms of a system constrains the application of MD simulations to the modest system sizes of ∼10³–10 (7) atoms. Systems of this size are not representative of large biomolecular assemblies. Moreover, these systems with atomistic resolution cannot be simulated for sufficiently long simulation time required to investigate a biophysical process such as large conformational changes of a protein, signal transduction through membrane proteins, and macromolecular diffusion. The system size and sampling time issues can be lessened by representing the system in a low-resolution “Coarse-Grained” (CG) model which significantly reduces the number of particles in the system. Over the years, a plethora of CG methods have been developed and applied to simulate systems from simple fluids to complex nanobiomaterials. (1−8) The CG methods can be categorized chiefly into two types; (i) the first one is based on mapping all-atoms (AA) to CG beads descriptions and constructing effective potential models that describe interactions between the CG beads (multiscale approach) (2,9−12) and (ii) physical property-based mapping which involves calculating a system’s equilibrium structural and physical properties from the CG model and comparing them with the experimental data. (1,3,4) In recent years, attention has been shifted to machine learning (ML)-based multiscale modeling. (13−18) The major driving forces are the intricate complexity inherent in the mapping of AA and CG representations as well as the development of accurate interaction potential models. These ML methods with various flavors are evolving at an unprecedented speed propelled by the increasing computational resources including modern GPUs. While some ML-CG methods have shown proficiency in dealing with complex organic and biomolecules, others are still in their nascent stages. Along this line, two crucial aspects that need special attention are the reproducibility of structural and dynamical details of a system and the transferability of the CG models across systems and length scales. Optimized ML frameworks and methods addressing these issues are the demand of the hour. In light of the rapid advancements of ML-CG methods and their future implications, the Journal of Chemical Information and Modeling (JCIM) invites researchers across the globe to submit contributions on the development and application of new ML-based methods for designing CG models of molecules, including large macromolecules, as part of the upcoming virtual special issue (VSI) on “Enhancing Coarse-Grained Models through Machine Learning”. We welcome all types of manuscripts published by JCIM, such as articles, perspectives, viewpoints, reviews, letters, and application notes. For more information on manuscript types and how to submit, please visit the journal’s Web site. Submissions will be received through March 15, 2025. All articles submitted under this VSI will be peer-reviewed prior to acceptance, to ensure they fit the scope of the Virtual Special Issue and meet the high scientific publishing standards of the Journal of Chemical Information and Modeling. If accepted, publications will go online as soon as possible and be published in the next available issue. Publications on this topic will be gathered into a Virtual Special Issue and widely promoted thereafter. T.A.S. acknowledges the support from FAPESP (2021/04283-3) and the RCN through the CoE-Hylleraas Centre for Quantum Molecular Sciences (Grant No. 262695). This article references 18 other publications. This article has not yet been cited by other publications. This article references 18 other publications.

中文翻译：

---

通过机器学习增强粗粒度模型

计算机模拟，特别是分子动力学（MD）模拟，是研究控制各种物理化学和生物过程的复杂系统的微观和动态细节的有效工具。计算机硬件和算法的最新进展显着提高了 MD 方法的效率，并将其适用性扩展到大型大分子组装体的处理和对生物分子过程复杂机制的理解。 MD 方法依靠运动方程的积分来探索构象空间并最终评估自由能表面。对作用在构成系统的各种粒子上的力的评估代表了 MD 模拟中的主要计算瓶颈，计算成本以N ²的量级增加，其中N是系统中粒子的数量。因此，系统所有原子的显式表示限制了 MD 模拟在~10 ^{3的适度系统尺寸上的应用。}–10 (7) 个原子。这种尺寸的系统不能代表大型生物分子组装体。此外，这些具有原子分辨率的系统无法模拟足够长的模拟时间来研究生物物理过程，例如蛋白质的大构象变化、通过膜蛋白的信号转导和大分子扩散。通过以低分辨率“粗粒度”(CG) 模型表示系统，可以减少系统尺寸和采样时间问题，从而显着减少系统中的粒子数量。多年来，已经开发并应用了大量的 CG 方法来模拟从简单流体到复杂纳米生物材料的系统。 (1−8) CG方法主要可分为两类： (i) 第一个基于将全原子 (AA) 映射到 CG 珠子描述，并构建有效的潜在模型来描述 CG 珠子之间的相互作用（多尺度方法）(2,9−12) 和 (ii) 物理性质 -基于映射，涉及从 CG 模型计算系统的平衡结构和物理特性，并将其与实验数据进行比较。 (1,3,4) 近年来，人们的注意力已转向基于机器学习 (ML) 的多尺度建模。 (13−18) 主要驱动力是 AA 和 CG 表示映射固有的错综复杂性以及精确的相互作用势模型的开发。在包括现代 GPU 在内的不断增加的计算资源的推动下，这些具有不同风格的 ML 方法正在以前所未有的速度发展。虽然一些 ML-CG 方法已显示出处理复杂有机和生物分子的能力，但其他方法仍处于初级阶段。沿着这条线，需要特别注意的两个关键方面是系统结构和动力学细节的可重复性以及跨系统和长度尺度的 CG 模型的可转移性。解决这些问题的优化机器学习框架和方法是当前的需求。鉴于 ML-CG 方法的快速发展及其未来影响，《化学信息与建模杂志》 (JCIM) 邀请全球研究人员就基于 ML 的新方法的开发和应用提交贡献，以设计 CG 模型分子，包括大分子，作为即将推出的“通过机器学习增强粗粒度模型”的虚拟特刊（VSI）的一部分。欢迎JCIM发表的各类稿件，例如文章、观点、观点、评论、信件和应用说明。有关稿件类型以及如何提交的更多信息，请访问该期刊的网站。提交内容的接收截止日期为 2025 年 3 月 15 日。在此 VSI 下提交的所有文章在接受之前都将经过同行评审，以确保它们符合虚拟特刊的范围并符合《化学信息杂志》和《化学信息杂志》的高科学出版标准。造型。如果被接受，出版物将尽快上线并在下一期出版。有关该主题的出版物将收集成虚拟特刊，并随后广泛推广。 TAS 感谢 FAPESP (2021/04283-3) 和 RCN 通过 CoE-Hylleraas 量子分子科学中心提供的支持（授权号 262695）。本文引用了其他 18 篇出版物。这篇文章尚未被其他出版物引用。本文引用了其他 18 篇出版物。