Overview
This project will systematically investigate how 3D-molecular novelty and complexity impacts synthesisability and will develop methods to mitigate this loss. Building on state-of-the-art 3D generative architectures (Irwin et al. 2024) and datasets (Axelrod et al. 2022 and Ramakrishnan et al. 2014), the research will quantify the synthesisability gap by integrating conditioning constraints from high-quality informatics sources such as the Cambridge Structural Database (CSD).
About this opportunity
The molecule design process is often hampered by high costs and lengthy development cycles. Recent advances in 3D-aware generative models offer a promising route to accelerate novel molecule discovery, yet these approaches frequently encounter two critical limitations:
- Limited extrapolation into novel spaces: Generative models often struggle to generate molecules exhibiting desired properties beyond existing chemical knowledge, restricting their utility for genuinely novel molecular design (Klarner et al., 2024; Ziv et al., 2024)
- 3D-Awareness and Synthesizability Constraints: Generated structures frequently lack practical synthetic accessibility (Cretu et al. 2024, Igashov et al. 2024) limiting their practical application in real-world scenarios.
Multiple levels of 3D complexity, e.g.,the incorporation of interaction field constraints from resources like Isostar, Superstar, and hotspot potentials—will be developed to understand their impact on synthesisability. Validation will be achieved through case studies targeting well-characterised systems (e.g., hERG and neglected tropical disease targets), ensuring that the outputs have direct relevance to molecule discovery pipelines. The project is positioned to bridge the gap between digital design innovation and practical synthesis, addressing a critical bottleneck in AI-driven materials chemistry.
Expected outcomes
The project will produce:
- A comprehensive analysis quantifying and characterizing the relationship between 3D-driven structural novelty and synthesizability in generative models
- A novel, integrated generative modelling framework, leveraging constraints from databases like CSD (IsoStar, SuperStar) and chemical informatics to reliably generate synthesizable, yet structurally novel molecular candidates – in collaboration with our industrial partner, CSD
- Ideally – experimentally validated molecular designs demonstrating practical relevance – in collaboration with our industrial partner, GSK.
This work will establish foundational guidelines for integrating synthetic constraints into generative models, bridging the current gap between novelty and chemical synthesis in 3D molecule generation.
This project is offered under the University of Liverpool EPSRC Centre for Doctoral Training in Digital and Automated Materials Chemistry along with other studentships for students from backgrounds spanning the physical and computer sciences to start in October 2025. These students will develop core expertise in robotic, digital, chemical and physical thinking, which they will apply in their domain-specific research in materials design, discovery and processing. By working with each other and benefiting from a tailored training programme they will become both leaders and fully participating team players, aware of the best practices in inclusive and diverse R&D environments.
Who is this opportunity for?
The candidate should possess a strong background in chemistry, physics, machine learning, or mathematics with a strong desire to learn the other components required for a successful project.
The student will work across the research groups of Dr Anthony Bradley (Chemistry) and Dr Gabriella Pizzuto (Computer Science).