Collaborative Research: ACED: Building Molecule Generative Models for Drug Development via Conditional Diffusion and Multi-Property Optimization

Drug development is a critical yet notoriously resource-intensive and time-consuming process, typically taking 10-15 years and costing between $1 to $1.6 billion to bring a successful drug to market. To expedite the process and enhance cost efficiency, significant research has focused on developing computational methods as alternatives/in parallel to conventional experiment-based approaches. Although promising, these methods rely heavily on trial and error within limited chemical subspaces (e.g., molecular libraries), resulting in suboptimal precision and outcomes dependent on the expertise of the researchers. This reliance also limits the scalability and automation of rapid drug design for new protein targets. To address these challenges, this project seeks to develop comprehensive generative AI methodologies and computational tools that expedite drug discovery, enhance cost efficiency, and improve success rates. By creating a holistic generative artificial intelligence (AI) framework capable of generating high-quality drug candidates with multiple desired properties, the project has the potential to transform pharmaceutical research. This initiative promotes advancements in healthcare by reducing the time and costs associated with drug development, ultimately benefiting public health. It also supports education and diversity by involving students from varied backgrounds, integrating AI into coursework, and conducting outreach to K-12 students, fostering broader societal engagement with STEM fields. Generating structured data, such as molecules, with multiple properties is technically challenging. This project will develop a conditional diffusion model for 3D molecule generation to enable both ligand-based drug design and structure-based drug design. The diffusion model employs an SE(3)-equivariant denoising component conditioned on given ligands, binding pockets, or both, and a classifier-free guidance mechanism to ensure that generated molecules closely align with specified conditions. Additionally, the project introduces the Direct Multi-Property Optimization framework, which optimizes drug-specific properties without requiring expensive model retraining. This framework leverages advanced optimization techniques, such as bi-level and multi-objective methods, to enhance the quality and adaptability of generated molecules. Research activities include three thrusts: (1) developing the conditional diffusion model for conditional 3D molecule generation, (2) creating the Direct Multi-Property Optimization framework for multi-property molecule generation, and (3) conducting rigorous evaluations and validations in silico and on other applications. These innovations aim to significantly reduce the time, cost, and resources required for drug discovery while increasing its success rates. 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.