Accurate Molecular Mechanics Force Fields through Data-driven Parameter Type Definitions

分子过程很复杂，只有一小部分细节可以通过实验技术辨别出来。然而，在许多应用中，能够预测相关的分子细节是非常有兴趣的。在这种背景下，原子模拟对于探索(生物)分子的性质和相互作用变得越来越重要。尽管这些模拟在理论上是合理的，但它们并不一定准确，而误差的一个关键来源是潜在的分子力学力场，它将给定的分子结构与原子力联系起来。今天，改进力场的一个主要障碍是，用于将原子分成不同类别来分配力场参数的方案缺乏严密性。这些类别通常被称为参数类型，它们将相似的化学环境(即子结构)分组，并为这些化学环境中的原子分配一组共同的参数。为了避免过度拟合和促进参数优化，这些类型应该尽可能少，同时仍然能够在计算和参考(实验或高水平量子化学计算)分子性质之间保持良好的一致性。然而，参数类型在历史上基本上是以临时方式分配的。这防止了当新的参考数据可用时对力场参数的严格优化，以及直接将新的化学子结构引入现有的力场。在这里，我提出了一种新的方法，通过联合数据驱动的力场参数类型定义和力场参数值的优化来克服上述障碍。这种方法从根本上不同于现有的力场优化方法，即只对给定的力场进行参数调整或添加参数。在拟议的项目中，将使用贝叶斯推理和蒙特卡罗抽样算法对参数类型定义进行抽样，以获得高精度的力场，同时尽可能少的类型(从而尽可能简单)。在参数采样过程的任何给定步骤，现有参数类型要么被合并，要么被拆分成新的参数类型。由于可能的合并或分裂操作的数量很多，参数类型将通过量子级别的原子特征来表示，从而能够对给定的化学环境进行基于物理的可计算描述。所提出的工作的意义在于其基本的数据驱动和建立力场的严格方法，而不受力场的特定函数形式或应用领域的限制。此外，如果有新的参考数据可用，开发的方法将使力场很容易扩展--这是材料设计和药物发现的一个重要方面。最后，通过将开发的技术实现到开放源码的Python包中，研究的影响将最大化。

Molecular processes are complex and only a fraction of their details is discernable by experimental techniques. However, there are many applications in which it is of high interest to be able to predict the relevant molecular details. In this context, atomistic simulations have become increasingly important to probe the properties and interactions of (bio)molecules. Although these simulations can be theoretically sound, they are not necessarily accurate, and a key source of error is the underlying molecular mechanics force field, which relates a given molecular structure to atomic forces. Today, a major hurdle to improving force fields is the lack of rigor in the schemes used to cast atoms into categories for assignment of force field parameters. These categories, which are commonly termed parameter types, group similar chemical environments (i.e. substructures) and assign a common set of parameters to the atoms within these chemical environments. To avoid overfitting and facilitate parameter optimization, these types should be as few as possible while still enabling good agreement between computed and reference (experimental or high-level quantum chemistry calculation) molecular properties. However, parameter types have historically been assigned in a largely ad hoc manner. This prevents the rigorous optimization of force field parameters as new reference data becomes available and the straightforward introduction of new chemical substructures into existing force fields. Here, I propose a novel approach that overcomes the aforementioned obstacles through the combined data-driven optimization of force field parameter type definitions and force field parameter values. The approach is fundamentally different from existing force field optimization approaches that only tuned or added parameters to a given force field. In the proposed project, Bayesian inference and Monte Carlo sampling algorithms will be applied for the sampling of parameter type definitions in order to obtain force fields with high accuracy while at the same time having as few types as necessary (thus being as simple as possible). At any given step of the parameter sampling process, existing parameter types are either merged or split into new ones. Since the number of possible merging or splitting operations is vast, parameter types will be represented through quantum-level atomic features, thus enabling a computable physics-based description for a given chemical environment. The significance of the proposed work is its fundamentally data-driven and rigorous way to build force fields without the restriction to a particular functional form or application domain of the force field. Furthermore, the developed approach will make force fields easily extensible if new reference data becomes available- an important aspect in materials design and drug discovery. Finally, the impact of the research will be maximized by implementing the developed technology into an open source python package.