Geometric Deep Learning for Generative Modelling in Computational (Bio-)Chemistry

背景、新颖性和目标：几何深度学习是机器学习的一个快速发展的子领域，专注于开发深度学习体系结构，能够使用数据中发现的内在几何结构，包括图形、网格和其他拓扑结构，以提高模型性能和样本效率。与生成模型方面的最新进展相结合，几何深度学习已被证明非常适合于应对物理科学中的挑战，特别是计算(生物)化学中的挑战。一个值得注意的例子是AlphaFold在蛋白质结构预测方面取得的突破性进展。然而，尽管在计算(生物)化学中应用几何深度学习进行生成性建模的最新进展令人印象深刻，但一个持久的挑战仍然存在：对大分子系统的完全平衡分布进行建模。准确地模拟这些平衡分布是推导自由能差、结合亲和力和反应速率等量的关键步骤，这些量是药物发现和材料设计中的关键因素。因此，该项目的主要目标是基于几何深度学习的最新进展开发新的算法和技术，使平衡分布的建模适用于大规模分子系统，如蛋白质-蛋白质复合体。通过这样做，我们的目标是为计算化学领域做出切实的贡献。同时，我们打算利用计算化学领域的现有技术和见解来推动几何深度学习和生成性建模的进一步发展。EPSRC战略和研究领域：拟议的项目符合人工智能技术、化学反应动力学和机理以及计算和理论化学的研究领域

Context, Novelty and Objectives: Geometric Deep Learning is a rapidly growing subfield of machine learning that focusses on developing deep learning architectures capable of using the inherent geometric structure found in data, including graphs, meshes and other topological structures, to improve model performance and sample efficiency. When coupled with recent advances in Generative Modelling, Geometric Deep Learning has demonstrated to be remarkably well suited for addressing challenges in the physical sciences, particularly in Computational (Bio-)Chemistry. A notable example of this is AlphaFolds groundbreaking advancement in protein structure prediction. However, while recent progress in applying Geometric Deep Learning for Generative Modelling in Computational (Bio-)Chemistry is impressive, an enduring challenge remains: modelling full equilibrium distributions of large molecular systems. Preciselymodelling these equilibrium distributions is a critical step in deriving quantities such as free energy differences, binding affinities, and reaction rates-key factors in drug discovery and material design. The primary objective of this project is thus to develop new algorithms and techniques based on recent advances in Geometric Deep Learning to render the modelling of equilibrium distributions feasible for large scale molecular systems, such as protein-protein complexes. In doing so, we aim to contribute tangibly to the field of computational chemistry. Concurrently, we intend to leverage existing techniques and insights from the field of computational chemistry to drive further progression both Geometric Deep Learning and Generative Modelling. EPSRC Strategies and Research Areas: The proposed project aligns with the research areas Artificial Intelligence Technologies, Chemical reaction dynamics and mechanisms and Computational and Theoretical Chemistry