From atoms to mechanisms - Artificial Intelligence augmented molecular simulations for mechanistic ligand design

虽然在帮助合理药物设计的计算方法方面取得了许多进展，但大多数 到目前为止，这些方法对药物靶标采取了静态的观点，忽略了与 构象转变的动力学和热力学。因此，迫切需要新的 用于大规模药物发现和开发的准确、易处理和自动化的计算方法 化学生物学研究解释了仿制药靶标性质的变化。建立在强大的基础上 理论和计算的初步结果，本程序试图理解和指导设计 新的酪氨酸激酶抑制剂和机械引导范例中的核糖开关模型。 我们的研究计划是由以下中心假设驱动的：(A)机械意识配基设计 战略可以超越仅由结构指导的传统战略，以及(B)人工智能(AI)- 集成分子动力学(MD)模拟方法可以帮助高通量地学习机制 时尚。我们的计划可以分为两个重要而又相互补充的主题领域。在第一个区域， 我们将在统计力学、MD模拟和人工智能的界面上开发抽样算法，以 探索罕见事件过程的机制，如药物解离和构象变化。 具体地说，我们将在前期工作的基础上，使用神经信息处理和 自然语言处理，并使其适用于学习反应坐标的高级采样方法， 随着模拟的进行，热力学和动力学在运行中。此外，我们还将互动 与其他领先的计算小组密切合作，将我们的抽样方法与他们的抽样方法相结合，并 为可极化力场的开发提供高效、准确的采样。在第二个领域，我们将 使用我们的算法指导机械驱动的Src、Abl激酶和PreQ1抑制剂的设计 核糖开关。我们将采取机械驱动的观点，在其中我们将制定出不同的 给定目标的构象，并了解现有配体如何与这些构象相互作用，然后提出 基于这一认识的配体修饰。我们将使用我们的人工智能增强MD方法来 在一个不间断的“原子到机制”的工作流程中理解配体的解离机制 只需最少的人为干预。我们所有的预测都将以不同的方式得到我们的 石溪大学和国家癌症研究所的实验合作者

While there have been numerous advances in computational methods aiding in rational drug design, most approaches so far take a static view of the drug target, ignoring the complexities associated with the dynamics and thermodynamics of conformational transitions. There is thus a pressing need for new computational methods that are accurate, tractable and automatable for large scale drug discovery and chemical biology studies that account for the changing nature of a generic drug target. Built on strong theoretical and computational preliminary results, this program seeks to understand and guide design of new inhibitors of tyrosine kinases and model riboswitches in a mechanistically guided paradigm. Our research program is driven by the central hypotheses that (a) mechanistically aware ligand design strategies can outperform traditional strategies guided only by structure, and (b) artificial intelligence (AI)- integrated molecular dynamics (MD) simulation methods can help learn mechanisms in a high-throughput fashion. Our program can be split into two overarching yet complementary thematic areas. In the first area, we will develop sampling algorithms at the interface of statistical mechanics, MD simulations and AI to probe mechanisms for rare event processes, such as drug unbinding and conformational change. Specifically we will build on our preliminary work in using ideas from neural information processing and natural language processing, and adapt them for advanced sampling methods that learn reaction coordinate, thermodynamics and kinetics on-the-fly as the simulation progresses. In addition, we will also be interacting closely with other leading computational groups to integrate our sampling methods with theirs and to facilitate efficient, accurate sampling for polarizable force-field development. In the second area, we will use our algorithms to guide mechanistically driven design of inhibitors of Src, Abl kinases and PreQ1 riboswitches. We will take a mechanistically driven perspective wherein we will map out the different conformations of a given target and understand how an existing ligand interacts with these, and then propose ligand modifications based upon this understanding. We will use our AI-augmented MD methods to understand the dissociation mechanisms of ligands in one uninterrupted “atoms to mechanism” workflow with minimal human intervention. All our predictions will be validated in different ways by our experimental collaborators at Stony Brook University and the National Cancer Institute