Collaborative Research: III: Advanced Diffusion Models for Molecular Graph Generation

Discovering new molecules is critical in many science and engineering fields related to medicine, materials, energy, and manufacturing, etc. However, the discovery process is challenging mainly due to the vast space of possible molecules and the lack of technologies to explore it efficiently. Generative artificial intelligence (AI) methods hold the potential of accelerating the process, but molecular generation poses unique challenges compared to images and texts. Specifically, the three-dimensional structures of molecules strongly determine their properties and activities. Current generative AI models often struggle to capture such structural features and meet diverse practical design constraints. These models can also be computationally expensive, and high-quality training data are limited, making resource efficiency critical. This project will develop new generative AI models for molecular generation under structural and practical constraints, enabling faster and reliable discovery of useful molecules. Our research outcomes are anticipated to advance multiple science and engineering disciplines, including drug discovery, materials design, sustainable energy, and manufacturing, etc. The project will also support education and workforce development in AI and life sciences by providing open research tools, course materials, mentoring activities, and workshops for students from high school to graduate school. This project will develop advanced diffusion models for molecular graph generation. A molecule will be represented as a graph whose nodes and edges correspond to atoms and chemical bonds, respectively. Additionally, the three-dimensional coordinates describe the molecular shape. The proposed research has three connected directions. First, the project will develop molecular structure-aware diffusion methods that can generate chemically meaningful atoms, bonds, and 3D geometric information, while remaining reliable for molecules of different sizes. Second, the project will design task-centric diffusion methods that can be adapted to meet different practical design goals and application settings. Third, the project will build resource-efficient diffusion methods that reduce computational and data demands, making molecular generation faster and more effective when high-quality training data are limited. The methods will be evaluated through molecular generation, conformer generation, and drug discovery tasks, with emphasis on chemical validity, stability, speed, and data efficiency. Practically, the project will also produce open-source software tools to be used by researchers and practitioners. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.