Our team will recommend compounds predicted to bind to the RNA-binding site of the SARS-CoV-2 helicase NSP13 and with potential for subsequent medicinal chemistry optimization. To this end, we will first filter a set of commercially available molecules (including those suggested in the CACHE guidelines) to reduce potential safety liabilities and undesired chemical reactivity and maximize lead-likeness. This step will also considerably reduce the chemical space that needs to be considered. The resulting dataset (approximately 18 M compounds) will be docked against a target structure suggested by the CACHE organizers using a freely available variant of Autodock Vina. Besides the Vina scoring function, the docking results will also be re-scored using RF-Score-VS v2, a state-of-the-art machine-learning scoring function, and SCORCH, an improved deep learning-based scoring function developed this year by our group. These three different scores will be combined as part of a consensus score. We will also conduct additional docking and scoring of selected compounds using more exhaustive docking settings and utilizing a “consensus docking” scheme originally proposed by our team, and which has subsequently become state-of-the art in the field (see references). We will also dock and score the library against alternative conformations of the target protein. The selection of the 100 compounds for experimental testing will be made based on the following criteria: the consensus predicted affinity scores, druglikeness scores, chemical diversity, and manual inspection of the predicted binding poses. 

Our approach involves a unique combination of computational approaches that our group is highly experienced with and that have proved to deliver excellent performance in the past (see references). Our pipeline not only considers pose prediction accuracy and binding affinity predictions, but also physicochemical properties, reactivity, and druglikeness, which are often responsible for high rates of attrition. Furthermore, our approach makes use of a newly developed scoring function from our group. This scoring function, SCORCH, promises unprecedented performance in identifying binder compounds by having addressed key pitfalls in the application of machine learning to virtual screening.  

StarDrop

Autodock, PSOVina2, GWOVina, RF-Score-VS v2, SCORCH, Osiris DataWarrior, PDB2PQR, OpenBabel, RDKit

SCORCH: Improving structure-based virtual screening with machine learning classifiers, data augmentation, and uncertainty estimation. doi: 10.1016/j.jare.2022.07.001. 

Comparison of ATP-binding pockets and discovery of homologous recombination inhibitors. doi: 10.1016/j.bmc.2022.116923 

Consensus Docking: Improving the Reliability of Docking in a Virtual Screening Context. doi: 10.1021/ci300399w 

Consensus Docking in Drug Discovery. doi: 10.2174/1573407214666181023114820