Authors

N Harrison¹; 
¹ Evariste Technologies, UK

Discussion

Structure-based virtual screening has cemented itself as a core pillar of drug discovery and is a routine aspect of both hit identification and lead optimisation. However, due to the complexities associated with predicting ligand binding modes and quantifying the energy of such a mode, virtual screening is fraught with inaccuracies.


To overcome this we have developed an automated preprocessing pipeline that attempts to optimise the efficiency, reliability, and effectiveness of our virtual screens. The modular pipeline is able to specifically search and filter the protein data bank and carry out structure-based analysis to collect high quality, diverse representatives of the target under investigation. A processing module utilises various tools to protonate, repair and re-annotate the protein files which are then provided to a constraint generator. Key protein-ligand interactions and water molecules extracted through this generator are added to a large parameter set that contains various other iterable features, such as choice of PDB and scoring functions. Hyper-optimisation over this parameter space, with evaluation through various enrichment metrics, allows selection of an optimised set of constraints.


Combining our synthon-based docking algorithms with these optimised knowledge-based constraints we are able to efficiently search through ultra-large chemical libraries, implicitly scoring billions of compounds and rapidly enriching high-scoring molecules. Post-processing provides a final refinement stage in which only the most promising of compounds rise to the top.