Time for a Step Change in Force Field Design

文献表明，生物分子模拟的核心存在一个严重的问题：它不能可靠地预测蛋白质在水溶液中的结构和动力学。因此，分子动力学模拟不能有效地补充实验，而研究阿尔茨海默病的发病。流行的力场不能可靠地复制本质上无序的淀粉样β肽的错误折叠和聚集。补救措施通常添加或删除定义不明确的能量项或重复重新参数化。由于这一战略并没有解决问题，因此需要对力场结构进行彻底改革。我们提出了FFLUX，一个真正新颖的力场，它更接近潜在的量子现实，并且“看到电子”。FFLUX利用了系统内原子的无参数定义。通过机器学习，FFLUX可以了解原子能量、电荷和多极矩如何随周围原子的几何形状而变化。因此，它在一个简化的方案中捕获了所有的极化和多体效应，以及电荷转移。该方法避免了微扰理论，从而受益于一个明确的治疗短程相互作用。此外，FFLUX打破了先进的可极化力场的刚体约束。FFLUX核心的定义明确的原子实现了基于物理的机器学习。它使用克里金法而不是神经网络，从而减少了训练数据的大小。我们仔细的工作计划植根于氨基酸和水簇，并通过一系列越来越相关的系统，在气相和水中扩大到溶剂化淀粉样β肽。我们将引入更复杂的机器学习，并在CPU、GPU和FPGA上实现最先进的并行化，从而为程序DL_FFLUX提供一个新的用户社区。在生物分子模拟中实现一步变化是一项艰巨的任务，即使是对于这种类型的资助，但我们小组的概念验证证明了这一点。

The literature shows that there is a serious problem at the heart of biomolecular simulation: it cannot reliably predict the structure and dynamics of proteins in aqueous solution. Thus, molecular dynamics simulation cannot usefully complement experiment while studying the onset of Alzheimer's disease. Popular force fields do not reliably reproduce the misfolding and aggregation of the intrinsically disordered amyloid beta peptide. Remedies usually add or take away ill-defined energy terms or repeatedly re-parameterise. Because this strategy has not solved the problem, force field architecture needs to beoverhauled. We propose FFLUX, a truly novel force field, that is much closer to the underlying quantum reality and one that "sees the electrons". FFLUX exploits a parameter-free definition of an atom inside a system. Using machine learning FFLUX learns how atomic energies, charges and multipole moments vary with the surrounding atoms' geometry. As such it captures all polarisation and many-body effects, as well as charge transfer, in one streamlined scheme. The approach avoids perturbation theory and thus benefits from a clear treatment of short-range interactions. Moreover, FFLUX breaks free from the rigid-body constraints of advanced polarisable force fields. The well-defined atom at the heart of FFLUX enables physics-based machine learning. It uses kriging instead of neural nets, thereby reducing the training data size. Our careful work plan is rooted in amino acids and water clusters, and scaled up to the solvated amyloid beta peptide via a sequence of increasingly relevant systems, both in gas-phase and in water. We willintroduce more sophisticated machine learning and implement state-of-art parallellisation on CPUs, GPUs and FPGAs, thereby offering a new user community the program DL_FFLUX. Bringing about a step change in biomolecular simulation is a huge task, even for this type of grant, but feasible as evidenced by proofs-of-concept from our group.