Topologically and Geometrically Inspired Machine Learning for Drug Design

药物设计是非常耗时和昂贵的。最大的制药公司每年花费数十亿美元用于新药的研究和开发。这类研究的一个关键领域是对具有生物活性的小分子的研究。各种计算机辅助药物设计方法已经出现，以提高寻找这些活性小分子的效率。特别是，基于形状的配体方法，它使用配体本身的三维形状来预测活性，是一种有吸引力的方法，因为配体的形状对于配体与蛋白质的结合很重要。最近的拓扑数据分析（TDA）领域使用代数拓扑工具来分析高维数据集，并获得了许多有希望的最新结果。来自TDA的技术非常适合于分析点云的形状，这使其成为基于形状的药物发现的自然候选。该项目旨在从计算拓扑和几何中获得灵感，为计算机辅助药物设计创造新的基于形状的方法。这个项目将从分子的点云生成简单的复合体，这些复合体将被用来制作拓扑和几何特征。然后，这些特征将作为输入输入到机器学习管道中，该管道将被设计用于分类活性和非活性分子。任何对分子活性进行分类的机器学习管道在三维特殊欧几里得群SE(3)下都应该是不变的，因为分子活性不受分子平移和旋转的影响。因此，该项目将利用几何深度学习新领域的技术来创建一个不变或等变的分类管道。将几何特征与化学描述符相结合的数据融合也将作为提高分类方法预测能力的潜在方法进行研究。流行的ddu - e和最新的LIT-PCBA虚拟筛选基准测试数据集将用于验证所创建的方法。这项研究与EPSRC的以下研究领域保持一致：人工智能技术、几何和拓扑、统计学和应用概率。这项工作将与牛津药物设计公司合作完成。

Drug design is immensely time-consuming and expensive. The largest pharmaceutical companies annually spend billions on the research and development of new drugs. A key area of such research is the investigation of biologically active small molecules. Various computer-aided drug design methods have emerged to improve the efficiency of the search for these active small molecules. In particular, shape-based ligand methods, which use the three-dimensional shape of the ligands themselves for predicting activity, are an attractive approach to the task because the shape of a ligand is important for ligand-protein binding. The recent field of topological data analysis (TDA) uses tools from algebraic topology to analyse high-dimensional datasets with many promising recent results. Techniques from TDA are well suited to analysing the shape of point clouds which makes it a natural candidate for shape-based drug discovery. This project aims to take inspiration from computational topology and geometry to create novel shape-based methods for computer-aided drug design. This project will generate simplicial complexes from the point clouds of molecules which will be used to make topological and geometric features. These features will then be fed as inputs to a machine learning pipeline which will be designed to classify active and inactive molecules. Any machine learning pipeline that classifies molecular activity should be invariant under the three-dimensional special Euclidean group, SE(3), since molecular activity is unaffected by the translations and rotations of a molecule. Therefore, this project will take advantage of techniques from the new field of geometric deep learning to create an invariant or equivariant classification pipeline. Data fusion, combining geometric features with chemical descriptors, will also be investigated as a potential way to improve the predictive power of the classification method. The popular DUD-E and the more recent LIT-PCBA virtual screening benchmarking data sets will be used to validate the created methods.This research is aligned with the following EPSRC research areas: artificial intelligence technologies, geometry and topology, statistics and applied probability.This work will be done in collaboration with Oxford Drug Design.