Next generation free energy perturbation (FEP) calculations--enabled by a novel integration of quantum mechanics (QM) with molecular dynamics allowing a large QM region and no sampling compromises

项目摘要 计算化学对商业药物发现的价值现在已经确立。虚拟筛选 （包括分子对接）现在启动了大多数发现工作。最近，GPU和 基于云的计算极大地增加了可用于药物治疗的实际计算吞吐量。 的发现反过来，这又激发了人们对相对自由能计算的浓厚兴趣（例如，自由能 扰动，FEP）用于药物铅增强。FEP已应用于药物发现的边缘， 几十年来，但在最近的过去大规模并行移动FEP中心舞台，和FEP 在这些计算可靠的情况下，帮助减少了数月或数年的发现工作。 问题是FEP计算并不总是可靠的。而对于某些系统，FEP结果中的误差是 远低于1千卡/摩尔--他们已经成功地引导了缓慢/昂贵的实验室工作-- 在其他系统中，预测不是很有用。即使有可能进行回顾性分析， 通常不太清楚为什么FEP计算对某些目标系统如此有利，而对其他系统则如此不利。大体上， FEP的局限性可以归结为三个问题：对能量学（力场）的描述不好; 采样不足;或对基础科学的误解（例如，蛋白质模型不正确，错误 结合位点、错误的质子化状态等）。人们普遍认为，许多问题产生于第一个问题， 这些-和改进的评价能量使用量子力学（QM）将是重点在这里。 有一个巨大的兴趣的方法，可以帮助解决现有的问题与FEP。在此，我们建议 FEP的一个新平台，它结合了分子相互作用的量子力学描述， 中央利益。传统的FEP力场是一个简化的解析表达式，带有拟合系数 分子力学（Molecular Mechanics，MM）MM是真实分子相互作用的简单近似， 可以用量子力学来描述。但是，直到最近，QM一直过于昂贵，无法在 FEP所需的分子动力学（MD）采样背景。 最后，我们已经确定了如何将QM集成到FEP范例中，使用精心编程的 分布式处理平台，适合在商品云计算机上使用，并通过集成 半经验的QM实现，提供的预测比MM的预测好得多，但在一个 成本远低于完整的DFT QM预测。我们的实现允许FEP计算与现实的QM 核心区域的数百个原子进行采样的规模与准确的FEP 计算和周转与现代药物发现相称。在这里，我们建议验证这一点， 平台对传统的基于MM的FEP，以显示它解决了该方法的许多问题。我们将 还识别附加的实现思想以进一步提高有效吞吐量和/或准确性。

Project Summary The value of computational chemistry to commercial drug discovery is now well-established. Virtual screening (including molecular docking) now jumpstarts most discovery efforts. More recently, a combination of GPU and cloud based computing has vastly increased the realistic computational throughput available for drug discovery. In turn, this has ignited substantial interest in relative free energy calculations (e.g. Free Energy Perturbation, FEP) for drug lead enhancement. FEP has been applied at the fringes of drug discovery for decades, but massive parallelism in the more recent past has moved FEP to center stage, and FEP has helped shave months or years off discovery efforts where these calculations are reliable. The catch is that FEP calculations are not always reliable. While for some systems, the error in a FEP result is much less than one kcal/mol--and they have successfully steered slow/expensive bench efforts--there are other systems where the predictions are not very useful. Even where retrospective analysis is possible, it is often not very clear why FEP calculations are so good for some target systems, and so bad for others. Broadly, the limitations of FEP can be distilled down to three problems: A poor description of the energetics (force field); insufficient sampling; or a misunderstanding of the fundamental science (e.g., incorrect protein model, wrong binding site, wrong protonation state, etc.). It is generally believed that many issues arise from the first of these—and improving the evaluation of energetics using quantum mechanics (QM) will be the focus here. There is a huge interest in methods that can help obviate the existing problems with FEP. Herein, we propose a new platform for FEP, which incorporates a quantum mechanical description of the molecular interaction of central interest. The traditional force field used with FEP is a simplified analytic expression with fit coefficients termed Molecular Mechanics (MM). MM is a simple approximation of the true molecular interactions that can be described using quantum mechanics. But QM has been, until recently, far too expensive to use in the context of the molecular dynamics (MD) sampling required for FEP. At long last, we have determined how to integrate QM into the FEP paradigm, using a carefully programmed distributed processing platform that lends itself to use on commodity cloud computers, and by integrating a semiempirical QM implementation that provides predictions that are much better than those from MM, but at a cost far less than for a full DFT QM prediction. Our implementation allows FEP calculations with a realistic QM core region of hundreds of atoms to be carried out with the scale of sampling associated with accurate FEP calculations and with turnaround commensurate with modern drug discovery. Here, we propose to validate this platform against traditional MM-based FEP, to show it addresses many of the issues of that approach. We will also identify additional implementation ideas to further improve effective throughput and/or accuracy.