FLF Next generation atomistic modelling for medicinal chemistry and biology

几十年来，药物研发的效率一直在下降，以至于每一种新分子到达消费者手中，估计研发成本超过20亿美元。药物发现过程通常包括设计一种能够与体内靶标结合并产生治疗效果的小有机分子。部分问题在于，在这个过程中，太多的分子是在实验室里合成的，花费巨大，结果却没有所需的结合亲和力。需要一种能够可靠地预测结构结合的计算模型。结合结构生物学方法，如x射线晶体学和低温电子显微镜，基于计算结构的生物分子模拟是解决方案的重要组成部分。在最简单的层面上，生物分子模拟可以通过寻找原子上的力，从而解决它们的动力学问题，来“激活”结构生物学解决的静态图像。但我们也可以超越这一点，使用严格的热力学来预测潜在药物分子结构的变化对其与目标结合的影响。在COVID-19大流行期间，世界各地的计算研究人员都采用了这种方法，以紧急提供对SARS-CoV-2病毒的新认识，并设计其功能抑制剂。虽然我们知道原子上的力应该用量子力学的方程来计算，但对于常规的生物学问题来说，这些计算成本太大了。相反，生物分子的动力学和相互作用通常是用一个简化的计算模型来计算的，这个模型被称为力场。力场将原子模拟为由弹簧结合，并通过静电和范德华力相互作用。这些相互作用的强度由数千个可调参数模拟，这些参数经过调整以重现实验数据。这些力场对于生物分子模拟科学家来说是一项重要的使能技术，他们预测的准确性取决于力场模型及其参数的真实性。传统上，力场模型经过几十年的发展。因为用新的设计规则重新训练模型是不可行的，所以在过程的早期所做出的设计决策就变成了“固有的”。开放力场计划是一个学术-工业合作伙伴关系，旨在推进构建下一代分子力学力场所需的科学和软件基础设施。在我们合作项目第一阶段工作的一个例子中，我们共同开发了一个灵活的框架，用自定义力场模型扩展开放力场软件堆栈。在原理验证中，我们能够在几周内训练和测试一个新的广义力场模型，而不是几年，与传统力场相比，精度有所提高。我对未来领袖奖学金的更新期的设想是部署这个软件基础设施，以快速地从新的假设转移到训练好的力场模型，明确地确定设计决策对模型准确性的影响。例如，我将测试在高级量子力学数据集上训练的机器学习模型是否产生准确的力场原子电荷，以及是否可以使用上面描述的新力场模型构建准确的蛋白质力场模型。力场模型显示必要的准确性将部署在分子设计工作流程。通过与制药行业的项目合作伙伴和开放科学抗病毒发现计划的合作，我将展示基于结构的生物分子模拟的准确性改进，这将转化为提高药物发现管道的效率。

The efficiency of drug discovery has been falling for decades such that, for each new molecule that reaches the consumer, estimated research and development costs are in excess of $2 billion. The drug discovery process involves the design of, usually, a small organic molecule that is capable of binding to its target in vivo with therapeutic benefit. Part of the problem is that during this pipeline too many molecules are synthesised in the lab, at great expense, that turn out not to have the required binding affinity. A computational model that is capable of reliably predicting binding from structure is needed.Allied with structural biology methods, such as x-ray crystallography and cryo-electron microscopy, computational structure-based biomolecular simulation is an important part of the solution. At the simplest level, biomolecular simulation can be used to 'animate' the static pictures solved by structural biology, by finding the forces on the atoms, and hence solving for their dynamics. But we can also move beyond this, and use rigorous thermodynamics to predict the effects that changes in structures of potential drug molecules will have on binding to their target. Such approaches were employed by computational researchers all over the world during the COVID-19 pandemic to urgently provide new understanding of the SARS-CoV-2 virus and to design inhibitors of its function.Although we know that the forces on the atoms should be calculated using the equations of quantum mechanics, these are much too computationally costly to solve routinely for biological problems. Instead the dynamics and interactions of biological molecules are typically computed using a simplified computational model, known as a force field. The force field models the atoms as bonded by springs, and interacting with each other through electrostatic and van der Waals forces. The strengths of these interactions are modelled by thousands of adjustable parameters, which have been tuned to reproduce experimental data. These force fields are an important enabling technology for biomolecular simulation scientists, and the accuracy of their predictions depends on the realism of the force field model and its parameters.Traditionally, force field models evolved over periods of many decades. Design decisions taken early in the process became 'baked in', since re-training the model with new design rules was infeasible. The Open Force Field Initiative is an academic-industrial partnership aiming to advance the science and software infrastructure required to build the next generation of molecular mechanics force fields. In one example of our work from the first period of the Fellowship, we have co-developed a flexible framework to extend the Open Force Field software stack with custom force field models. In a proof-of-principle, we were able to train and test a new generalised force field model in a matter of weeks, rather than years, with improvements in accuracy over traditional force fields.My vision for the renewal period of the Future Leaders Fellowship is to deploy this software infrastructure to rapidly move from new hypotheses to trained force field models, with unambiguous determination of the effects of design decisions on model accuracy. For example, I will test whether machine learning models trained on high-level quantum mechanical datasets yield accurate force field atomic charges, and whether accurate protein force field models can be built using the new force field models described above. Force field models that show requisite accuracy will be deployed in molecular design workflows. Through working with the project partners in the pharmaceutical industry and at an open science antiviral discovery initiative, I will showcase the accuracy improvements in structure-based biomolecular simulations that will translate to improved efficiency of the drug discovery pipeline.