Reliable computational prediction of molecular assembly

正确理解和控制分子组装是现代物理科学研究的一个重要前沿。“具有目标属性的扩展结构的定向组装”的大挑战从不同的角度解决了这一前沿问题。我专注于化学计算的角度，它已经成为一个独立的信息源，与实验相辅相成。对分子间相互作用能进行准确、因而更可靠的预测，是一种日益迫切的需要。分子组合的结构和动力学敏感地依赖于最细微的能量变化。这就是为什么准确预测能源的科学挑战仍然像以往一样严峻。如果能正确地预测能量，那么其他一切都会随之而来：现实的结构、动力学和性质。我提出了一种新颖的方法，与当前的范式截然不同。环境条件下的物质由一个称为Schrödinger方程的主方程控制，该方程返回分子组装的相互作用能。以目前的计算机能力，精确地求解这个大分子聚集的主方程是非常昂贵的，甚至是不可能的。然而，力场可以提供这种相互作用能量，而且速度要快很多个数量级。力场是一个公式，它传递分子系统的能量，作为系统原子坐标的直接函数。这个公式包含许多参数，具体到手头的系统。挑战在于设计一个可靠的力场。最好也是唯一的长期策略是尽可能忠实地将力场映射到Schrödinger方程的解上，包括能量和波函数。随着百亿亿次计算机的出现和GPU技术最近超越cpu，投资于更现实的力场是关键和及时的。在这里，我的目标是提供生物分子建模一个全新的途径来设计力场，更真实的静电。我们提出了柔性分子的第一个（高阶）多极力场的完整构建，具有分子内和分子间的极化。这对于分子组装和识别，以及氢键的现实建模是至关重要的。在由离子引起的强而不均匀的电场存在下，分子系统也将首次被真实地建模。力场的真正预测能力取决于小分子（或分子簇）向大分子传递信息的可靠性。只有当这种可转移性很高时，力场才能做出可靠的预测。我们的力场（称为QCTFF）背后的主要思想是构建“有知识的”原子。这些原子是从小分子中提取出来的，通过相互作用来预测大分子的性质。它们是电子密度的三维碎片，具有有限的体积。这些原子有明显的边界，这使它们具有“延展性”。它们精确的形状对它们所处的分子的直接环境作出反应。然后，一种机器学习方法捕捉到这些原子是如何根据邻近原子的位置改变它们的多极矩的。我们已经成功地达到了这个新想法的概念验证阶段，现在我打算充分利用它。虽然QCTFF是通用的，但它的应用偏向于蛋白质、离子和水。只有奖学金才能实现雄心勃勃但可行的目标，即为生命科学应用创造这种变革性的使能技术。这项技术也将作为一个强大的平台，从中开发一个创新的新力场，研究溶液和酶中的反应，以及晶体成核。在QCTFF设计之初采取的激进和创新决策是其长期成功的最佳保证。

An important modern frontier of research in the physical sciences is the proper understanding and control over molecular assembly. The Grand Challenge of "Directed Assembly of Extended Structures with Targeted Properties" tackles this frontier from various angles. I focus on the angle of chemical computing, which has established itself as an independent source of information, complementary to experiment. There is a need, of ever increasing urgency, for accurate and hence more reliable prediction of interaction energies between molecules. The structure and dynamics of molecular assemblies sensitively depend on most subtle energy changes. This is why the scientific challenge of accurate energy prediction is still as acute as ever. If energy is correctly predicted then everything else follows: realistic structures, dynamics and properties. I propose a novel approach, drastically different to the current paradigm. Matter at ambient conditions is governed by a master equation called the Schrödinger equation, which returns the interaction energy of a molecular assembly. Solving this master equation accurately for sizeable molecular aggregates is very expensive or even impossible with current computer power. Force fields, however, can provide this interaction energy, and do so many orders of magnitude faster. A force field is a formula that delivers the energy of a molecular system as a direct function of the system's atomic coordinates. This formula contains many parameters, specific to the system at hand. The challenge is to design a force field that is reliable. The best and only long term strategy is to map the force field as faithfully as possible onto the solution of the Schrödinger equation, both in terms of energy and the wave function. With exascale computers around the corner and GPU technology recently overtaking CPUs, it is pivotal and timely to invest into more realistic force fields. Here I aim to offer biomolecular modelling a completely new route to designing a force field, with much more truthful electrostatics. We propose the completed construction of the first ever (high-rank) multipolar force field for flexible molecules, with both intra- and intermolecular polarisation. This is crucial for molecular assembly and recognition, as well as the realistic modelling of hydrogen bonding. Molecular systems, in the presence of the strong and inhomogeneous electric fields caused by ions, will also be modelled realistically for the first time. The true predictive power of a force field depends on the reliability of the information transfer of small molecules (or molecular clusters) to large molecules. Only if this transferability is high, a force field will make reliable predictions. The main idea behind our force field, called QCTFF, is to construct "knowledgeable" atoms. These atoms are drawn from small molecules and made to interact in order to predict properties of large molecules. They are 3D fragments of electron density, with a finite volume. These atoms have sharp boundaries, which endows them with a "malleable" character. Their precise shape responds to the immediate environment of the molecule they are part of. A machine learning method then captures how these atoms change their multipole moments in response to the positions of their neighbours. We have successfully reached the proof-of-concept stage of this novel idea and now I intend to fully exploit it.Although QCTFF is generic, its application is biased towards proteins, ions and water. Only a fellowship can deliver the ambitious but feasible goal of creating this transformative enabling technology towards life science applications. This technology will also serve as a robust platform from which to develop an innovative novel force field, to study reactions in solution and in enzymes, as well as crystal nucleation. The radical and innovating decisions taken at the outset of QCTFF's design are the best guarantee for its long lasting success.