In-silico prioritization of fragment hits from observed protein-ligand interactions

In the rapidly evolving field of drug discovery and cheminformatics, fragment-based drug discovery (FBDD) has become a key strategy, and one that is increasingly accessible, thanks to facilities like XChem at Diamond Light Source, which provides primary screening by protein crystallography to tens of academic and industrial projects annually. Nevertheless, ways of rapidly and economically progressing fragment hits to high-quality hits remains a both difficult and underserved problem. This project explores a computational approach for identifying the key protein-ligand interactions from sets of 3D-observed fragments in binding sites, the output from the XChem experiment, so that these can be prioritised in fragment follow-up design. Specifically, we will employ an approach called dynamic undocking (DUck) which leverages steered molecular dynamics (MD) simulations to evaluate the structural stability of a given fragment. This process calculates the work necessary to break critical protein-ligand contacts within the complex. Historically successful at assessing the robustness of hydrogen bonds, we look to expand DUck to also quantify other modes of interaction such as pi-stacking, salt bridges, and potentially hydrophobic interactions. This enhancement allows for the extraction of per-feature interaction scores for each element within the protein-ligand complex, thereby providing a more comprehensive assessment of fragment binding. This refined approach will be incorporated into an automated compound design pipeline, and we aim to containerize an open-source implementation of DUck, significantly enhancing its accessibility and usability for researchers and developers in the field. We aim to develop a nuanced understanding of observed interactions, discretised and then analysed as constellations of interactions, to identify "warhead" motifs. This presents a unique opportunity to additionally develop graph neural network models to produce global fragment scores and assist in the prioritisation of fragments in ongoing projects. These analyses and models can be validated against experimental measurements of congenic series being generated at XChem. Further investigations will focus on how to more generally map 3D-observed interactions onto the binding site and integrating this with generative models of compound design. We aim to revolutionize fragment-based drug discovery (FBDD) by integrating dynamic undocking (DUck) and graph neural networks, significantly accelerating the transition from fragment hits to chemical leads. By enhancing predictive accuracy and reducing experimental efforts, this methodology could set new industry standards, particularly benefiting rapid drug development in critical areas like antimicrobial resistance. In collaboration with Eli Lilly, this project harnesses both academic innovation and industry expertise. Eli Lilly's contribution, led by Lewis Vidler and Charlie Allerston, provides invaluable industry insight, ensuring research outputs are scientifically robust and commercially viable. This project falls within the EPSRC Antimicrobial resistance, Artificial intelligence technologies, Biological informatics, Biophysics and soft matter physics, Computational and theoretical chemistry, and Software engineering research areas.