Open data-driven infrastructure for building biomolecular force fields for predictive biophysics and drug design

项目概要/摘要 本项目摘要与原始 R01 相同。生物分子相互作用和设计的研究 新疗法的发展需要小分子和小分子之间原子相互作用的精确物理模型 生物大分子。近几十年来，分子力学力场已经证明了 物理模型在定量生物物理建模和预测分子设计方面具有潜力。然而， 我们在构建实现高精度、可系统化的力场的能力方面存在重大技术差距。 以统计上稳健的方式得到显着改进，扩展到化学的新领域，可以模拟翻译后 和共价修饰，能够量化预测中的系统误差，并且可以广泛应用于 一个高性能的软件包。在这个项目中，我们的目标是弥合这一技术差距，使新一代 化学生物学和药物的精确定量生物分子建模和（生物）分子设计的时代 发现。在目标 1 中，我们将建立一个现代化的、开放的基础设施，使从业者能够快速、交流 通过自动化机器学习轻松构建和使用准确且统计上稳健的物理力场 方法。在目标 2 中，我们将构建开放的、机器可读的实验和量子化学数据集， 将加速下一代力场的发展。在目标 3 中，我们将开发统计稳健的贝叶斯模型 推理技术能够自动构建类型分配方案，避免过度拟合 以及由数据统计证明合理的身体功能形式的选择。这种方法还将提供一个 由参数或函数形式的不确定性引起的预测特性的系统误差的估计 选择——通常是误差的主要来源——可以在几乎不增加成本的情况下进行量化。在目标 4 中，我们将 集成并应用该基础设施来产生开放的、可转移的、自洽的力场，从而实现高 模拟小分子与生物分子（包括非自然的相互作用）的准确性和广泛的覆盖范围 氨基酸或核酸以及有机分子的共价修饰），最终目标是覆盖所有 主要生物分子。 这项研究意义重大，因为该项目开发的技术有可能从根本上改变 通过提供高度精确的力场和极其广泛的化学物质来研究生物分子现象 通过有机（生物）分子完全一致的参数化进行覆盖。此外，我们将生产新工具 自动创建力场并针对特定问题领域进行定制，量化预测中的系统误差， 并识别新数据以提高力场精度。这将大大提高我们学习多元化的能力 分子水平的生物物理过程，并合理设计新的小分子、蛋白质和核酸 酸疗法。这种方法将为力场构建和应用领域带来统计的严谨性 通过提供一种做出数据驱动决策的方法，同时通过使其成为一种方法来增强可重复性 使用完全开放的基础设施和数据集进行严格且可重复的科学。

PROJECT SUMMARY/ABSTRACT This Project Summary is unchanged from the original R01. The study of biomolecular interactions and design of new therapeutics requires accurate physical models of the atomistic interactions between small molecules and biological macromolecules. Over the least few decades, molecular mechanics force fields have demonstrated the potential that physical models hold for quantitative biophysical modeling and predictive molecular design. However, a significant technology gap exists in our ability to build force fields that achieve high accuracy, can be systemati- cally improved in a statistically robust manner, be extended to new areas of chemistry, can model post-translational and covalent modifications, are able to quantify systematic errors in predictions, and can be broadly applied across a high-performance software packages. In this project, we aim to bridge this technology gap to enable new gen- erations of accurate quantitative biomolecular modeling and (bio)molecular design for chemical biology and drug discovery. In Aim 1, we will produce a modern, open infrastructure to enable practitioners to rapidly and conve- niently construct and employ accurate and statistically robust physical force fields via automated machine learning methods. In Aim 2, we will construct open, machine-readable experimental and quantum chemical datasets that will accelerate next-generation force field development. In Aim 3, we will develop statistically robust Bayesian inference techniques to enable the auto- mated construction of type assignment schemes that avoid overfitting and selection of physical functional forms statistically justified by the data. This approach will also provide an estimate of the systematic error in predicted properties arising from uncertainty in parameters or functional form choices—generally the dominant source of error—to be quantified with little added expense. In Aim 4, we will integrate and apply this infrastructure to produce open, transferable, self-consistent force fields that achieve high accuracy and broad coverage for modeling small molecule interactions with biomolecules (including unnatural amino or nucleic acids and covalent modifications by organic molecules), with the ultimate goal of covering all major biomolecules. This research is significant in that the technology developed in this project has the potential to radically transform the study of biomolecular phenomena by providing highly accurate force fields with exceptionally broad chemical coverage via fully consistent parameterization of organic (bio)molecules. In addition, we will produce new tools to automate force field creation and tailoring to specific problem domains, quantify the systematic error in predictions, and identify new data for improving force field accuracy. This will greatly improve our ability to study diverse biophysical processes at the molecular level, and to rationally design new small-molecule, protein, and nucleic acid therapeutics. This approach will bring statistical rigor to the field of force field construction and application by providing a means to make data-driven decisions, while enhancing reproducibility by enabling it to become a rigorous and reproducible science using a fully open infrastructure and datasets.