Applications of Deep Learning for Binding Affinity Prediction

背景：通过将计算方法应用于生物化学并对生物系统有更深入的了解，可以提高药物发现过程的效率，从而提高患者的治疗效果。计算化学生物学的一个核心问题是预测小分子是否会与蛋白质目标结合，再现实验结合分析中观察到的准确性。近年来，机器学习为蛋白质-配体结合预测提供了实质性的改进；然而，目前的预测仍有很多不足之处。新的人工智能方法，如深度学习，有可能通过解决绑定亲和力预测等问题来改变医疗保健。目的和目标：提高基于机器学习的计算化学工具的通用性，用于预测蛋白质-配体结合亲和力。探索各种分子表征，目的是理解如何最好地解释生物系统的计算，特别是蛋白质-配体结合。探讨新型人工智能方法在蛋白质靶点（如hERG）药物发现中的适用性。研究方法的新颖性：我们建议研究深度学习的新应用，包括图神经网络、原子中心对称函数和大型语言模型，以解决蛋白质-配体结合亲和力预测的基本问题。我们将评估模型学习基础物理的能力，包括研究结合相互作用的改进基准，并探索可泛化性和不可加性的关键方面。模型应用将包括预测与小分子药物（如hERG和细胞色素p450）的吸收、分布、代谢和排泄有关的蛋白质的脱靶结合，以及伙伴感兴趣的其他靶标。各种分子表征和方法将被研究，包括蛋白质化学测量的1D（小分子的smile；蛋白质的氨基酸序列），2D（配体，蛋白质及其相互作用的图）以及3D（等变神经网络）。与EPSRC的战略和研究领域保持一致：这个多学科项目与EPSRC的战略交付计划保持一致，通过培养英国境内的工业和大学合作伙伴关系进行有效的研究；在物理和数学科学领域进行发现研究；以使命为导向的研究为目标，用人工智能改变健康和医疗保健。这个跨学科的生物信息学项目跨越了EPSRC职权范围内的多个研究领域。本项目属于EPSRC“人工智能技术”、“生物信息学”、“化学生物学与生物化学”、“计算与理论化学”研究领域。

Background:By applying computational methods to biochemistry and developing a deeper understanding of biological systems, the efficiency of the drug discovery process, and consequently patient outcomes, could be improved. A core issue in computational chemical biology is predicting whether small molecules will bind to a protein target, reproducing the accuracy observed in experimental binding assays. In recent years, machine learning has provided substantial improvements in protein-ligand binding prediction; however, current predictors still leave much to be desired. Novel artificial intelligence methods, such as deep learning, have the potential to transform healthcare by solving issues like binding affinity prediction.Aims and objectives:To improve the generalisability of machine learning-based computational chemistry tools for predicting protein-ligand binding affinity.To explore various molecular representations with the aim of understanding how best to interpret biological systems computationally, in particular protein-ligand binding.To investigate the applicability of novel AI methods to drug discovery on protein targets, such as hERG.Novelty of the research methods:We propose investigating novel applications of deep learning, including Graph Neural Networks, Atom Centred Symmetry Functions, and Large Language Models to the fundamental problem of protein-ligand binding affinity prediction. We will assess the ability of models to learn the underlying physics, including investigating improved benchmarks, of binding interactions and explore key aspects of generalisability and non-additivity. Model applications will include predicting off-target binding to proteins involved in the absorption, distribution, metabolism, and excretion of small molecule drugs, such as hERG and cytochrome P450s, as well as other targets of interest to the partner(s). Various molecular representations and approaches will be investigated, including proteo-chemometric 1D (SMILES for small molecules; amino acid sequences for proteins), 2D (graphs for ligands, proteins, and their interactions), as well as 3D (equivariant neural networks).Alignment to EPSRC's strategies and research areas:This multidisciplinary project aligns with EPSRC's strategic delivery plan by nurturing industry and university partnerships for effective research within the UK; pursuing discovery research within the Physical and Mathematical Sciences; and targeting mission-inspired research to transform Health and Healthcare with AI. This interdisciplinary bioinformatics project spans multiple research areas within the EPSRC remit. This project falls within the EPSRC "Artificial intelligence technologies", "Biological informatics", "Chemical biology and biological chemistry", "Computational and Theoretical Chemistry" research areas.