Next-generation alchemical free energy methods and quantum/machine-learning models for drug discovery

新一代炼金术自由能方法和量子/机器学习模型 药物发现。 派：达林·M·约克，罗格斯大学，美国新泽西州皮斯卡塔韦，邮编：08854-8087。 炼金术自由能(AFE)模拟在药物发现的各个方面都是不可或缺的，因为它使 fi刚性和选择性的配基结合预测。取得进展的一个关键障碍是目前在预付款方面的限制 AFE模拟的精确度和准确性限制了它们的预测能力。目前的提案针对的是 这些障碍与新的AFE方法和模型将被集成到GPU加速的琥珀软件中- Ware套件在全球学术界、政府实验室和工业中使用(超过3万用户)。特别是fi，我们建议 目标：1.为稳健的高精度AFE模拟创造先进技术；2.开发精确的量子 机械/深度学习潜力(QDπ)强制fi领域用于药物发现和3.验证精确度和准确性 原子力学法和QDπ模型的合理性。在目标1中，我们将开发新的技术，以实现健壮和可重复使用 化合物文库配体-蛋白质结合自由能的计算。这些方法协同工作以实现 跨热力学图形网络的高精度、聚合AFE模拟。在目标2中，我们将制定一个 高精度和高计算精度的fi广义量子深势相互作用(QDπ)力fi场模型 用于药物研发。QDπ模型将作为机器学习势校正(∆-mlp)来表示 采用快速近似三阶密度泛函的量子力学/分子力学(QM/MM)能量 紧束缚QM模型和公认的琥珀MM力fi场和相容的水和离子模型。这个 ∆-mlp将利用我们最近开发的距离校正深度学习潜力(Dprc)来实现准确的内部学习。 以及分子间相互作用。在目标3中，我们将从以下方面对AFE方法进行深入验证研究 基于包括巨噬细胞迁移在内的系统基准系统的AIM 1和AIM 2的QDπ模型 抑制因子(MIF)、JAK2 JH2结构域、SARS-Cov2 MPRO以及Sigma 1和2受体。

Next-generation alchemical free energy methods and quantum/machine-learning models for drug discovery. PI: Darrin M. York, Rutgers University, Piscataway, NJ 08854-8087 USA. Alchemical free energy (AFE) simulations are indispensable in various aspects of drug discovery by enabling the prediction of ligand binding affinity and selectivity. A critical barrier to progress is the current limitation in pre- cision and accuracy of AFE simulations that restricts their predictive capability. The current proposal addresses these barriers with new AFE methods and models that will be integrated into the GPU-accelerated AMBER soft- ware suite used worldwide (over 30K users) in academia, government labs and industry. Specifically, we propose to: 1. Create advanced technology for robust high-precision AFE simulations; 2. Develop accurate quantum mechanical/deep-learning potential (QDπ) force fields for drug discovery and 3. Validate precision and accu- racy of AFE methods and QDπ model. In Aim 1, we will develop new technologies for robust and reproducible calculation of ligand-protein binding free energies of compound libraries. The methods work together to enable highly precise, converged AFE simulations across thermodynamic graph networks. In Aim 2, we will develop a highly accurate and computationally efficient general quantum deep-potential interaction (QDπ) force field model for drug discovery. The QDπ model will be formulated as a machine learning potential correction (∆-MLP) to the quantum mechanical/molecular mechanical (QM/MM) energy using fast, approximate 3rd-order density-functional tight-binding QM model and well-established AMBER MM force fields and compatible water and ions models. The ∆-MLP will leverage our recently developed range-corrected deep-learning potential (DPRc) for accurate intra- and intermolecular interactions. In Aim 3, we will conduct in depth validation studies of the AFE methods from Aim 1 and QDπ model of Aim 2 on a systematic set of benchmark systems, including macrophage migration inhibitory factor (MIF), JAK2 JH2 domain, SARS-Cov2 Mpro, and sigma 1 and 2 receptors.