Solving Molecular Conformation by Machine Learning and NMR

了解溶液中药物分子的构象对于优化药物与其靶蛋白结合的机制非常重要。这意味着计算分子模型是药物设计的关键一步。分配分子构象的最佳方法是使用核磁共振（通过解释耦合常数和核Overhauser效应，即NOEs）。核磁共振需要足够的、纯的化合物来分析，即使这样，也能得到该分子构象平均的信息。DFT是一种非常有用的计算分子或构象核磁共振参数的方法；然而，DFT需要很长时间，需要大量的计算能力，这是昂贵的和高能耗的（从而对环境产生负面影响）。机器学习是一种能源成本更低的替代方案。与DFT相比，机器学习非常快。经过训练后，它可以在几秒钟内做出使用DFT可能需要几天时间的预测。这些预测的质量完全取决于机器所训练的信息，在这种情况下，机器是根据dft计算的核磁共振参数进行训练的，从而得到一个可以预测核磁共振参数的模型，其精度接近dft水平。布里斯托大学的Butts小组已经发布了一个能够做到这一点的模型。目前，这些模型能够在非常高的水平上预测分子结构和连通性。然而，本博士的目标是将其扩展到预测构象信息。仅仅一个分子就可能有数百种不同的潜在构象，这意味着对一系列候选药物的每个构象进行DFT计算需要耗费大量的计算时间。该项目属于EPSRC人工智能和机器人研究领域，因为它涉及将机器学习应用于化学，具有潜在的重大影响。此外，它也属于EPSRC物理科学研究领域，因为它进一步了解了分子的构象行为。这项研究的应用可能非常重要，因为它可以提供一种更快、更便宜的方法来对可能开发成药物的化合物进行取样，从而促进药物的发现和开发。

Understanding the conformation of a drug molecule in solution is important to optimise the mechanism of a drug binding to its protein target. This means that computational molecular modelling is a crucial step in drug design. The best way to assign the conformation of a molecule is using NMR (by interpreting coupling constants and nuclear Overhauser effects, or NOEs). NMR requires sufficient, pure compound to analyse and even then, yields information of that molecule conformationally averaged. DFT is a very useful way to calculate NMR parameters for molecule or conformer of interest; however, DFT takes a long time and requires significant computational power which is expensive and highly energy consuming (thus negatively impacting the environment). Machine learning is a less energy-costly alternative. Machine learning is incredibly fast compared to DFT. Once trained, it can make predictions that might take days using DFT, in a matter of seconds. The quality of these predictions depends entirely on the information that the machine is trained on and in this case, the machine is trained on DFT-calculated NMR parameters, resulting in a model that can predict NMR parameters to near DFT-level accuracy. The Butts group at Bristol have published a model capable of doing just this. Currently, these models are able to predict molecular constitution and connectivity to a very high level. However, the objective of this PhD is to extend this into predicting conformational information. Just one molecule can have hundreds of different potential conformers which means that running DFT calculations on every conformer for a series of drug candidates uses a vast amount of computational time. This project falls within the EPSRC artificial intelligence and robotics research area since it involves the application of machine learning to chemistry with the potential to significant impact. Furthermore, it also falls within the EPSRC physical sciences field of research since it furthers the understanding of how molecules behave conformationally. The applications of the research are potentially very important since it could contribute a faster and cheaper way to sample compounds that could be developed into drugs, enhancing drug discovery and development.