Fragmenstein: Stitching Compounds Together
Matteo P. Ferla, Rubén Sánchez-García, Frank von Delft and Charlotte M. Deane
Oxford Protein Informatics Group (OPIG), Department of Statistics, University of Oxford
Fragment-based drug discovery (FBDD) is a powerful method used in drug discovery to converge on a lead candidate quickly and cheaply without empirically screening huge libraries. It relies on the ability of experimental structure solving methods, such as crystallography, to detect relatively weak binders and under the assumption that analogues bind in a similar manner, smaller fragment hits can be merged or expanded into larger more potent molecules.
Most current strategies for merging or expanding fragments tend to disregard the 3D protein-ligand conformation of the inspiring hit molecules. To address this, we introduce Fragmenstein (https://github.com/matteoferla/Fragmenstein), a tool that stitches the inspiration atoms together either to create a novel merged compound or to place the conformer for a given compound, and energy minimises these in place with heavy constraints. This allows follow-up compounds to be faithful to the inspiration hits, while still being physically plausible. Mergers or placements that yield a conformer that is unable to obey the position of the inspirations can be disregarded: on a benchmarking test, 44% of potential mergers were found to be strained, while on another test, modelling of both enantiomers of a covalent compound recapitulated the experimental results wherein only one enantiomer was able to react. The importance of obeying the inspiration hits can be seen in the Covid Moonshot dataset, wherein 69% of fragment-inspired follow-up compounds that soaked successfully had an RMSD under 2 Å relative to their inspiration hits. In this dataset, Fragmenstein was able to model 82% of these with an RMSD under 2 Å, whereas a pharmacophore-constrained docking approach modelled only 22% correctly (3% without pharmacophores).
In addition to these retrospective analyses, Fragmenstein has shown its potential in real-world screening campaigns. For example, it was utilised in hit-to-lead screening, yielding a submicromolar merger from inspiration hits in a single round.
