Agentic Deep Generative Models for De Novo Structure-Based Drug Design

The cost of developing a single therapeutic from bench to bedside costs upward of US$2 billion, and typically takes 10-15 years. However, attrition rates are estimated at 90%, with 9 out of 10 drugs that reach clinical trials failing at Phase I, II or III. Analyses of these failures has revealed poor-drug like properties and a lack of efficacy as central causes. As costs increase, the need for increased efficiency in research & development is greater than ever. Structure-based drug design (SBDD) leverages the 3D structure of target proteins to rationally design small molecules that can modulate their function. Traditionally, this has involved virtual screening of large chemical databases, a process that has proven time-consuming, expensive, and limits exploration to previously established molecules. Recently, there has been a resurgence of interest in applying diffusion models to SBDD, promising to revolutionize the field by enabling more efficient generation of novel, target-specific compounds. Diffusion models show particular promise in SBDD and can represent biomolecules in three-dimensional space, enabling more effective and targeted drug design strategies. This project will build upon current diffusion models for molecular generation, by guiding the generative process with a multi-objective reward function, which will be optimized through reinforcement learning with a specific focus on synthetic accessibility. We will also explore new diffusion model architectures and expand the work by focusing on more complex protein-ligand systems than previously explored. We will make the resulting computational methods accessible to medicinal chemists without a computational chemistry background through the use of large language model (LLM) agents. This project will be carried out under the supervision from Prof. Garrett M. Morris, an expert in the field of SBDD, in the Oxford Protein Informatics Group at the University of Oxford; and in collaboration-via the EPSRC SABS R3 CDT-with specialists in the field of computational chemistry and virtual screening at Eli Lilly. By reducing the time, expenditure, and attrition rates in the pre-clinical phase of drug discovery, this project has the potential to significantly impact the pharmaceutical industry. Moreover, by making AI tools accessible to non-computational chemists through conversational AI, we aim to bridge the gap between traditional medicinal chemistry experts and cutting-edge computational methods, fostering innovation in drug design across the field. This project aligns with EPSRC's strategies in advancing computational methods for drug discovery and promoting interdisciplinary research. Specifically, this project falls within the EPSRC research areas of Chemical Biology and Biological Chemistry, Computational and Theoretical Chemistry, Artificial Intelligence Technologies, and Natural Language Processing.