The Enamine REAL Database (5.5 billion compounds) will be used as target database for a deep learning-accelerated virtual screening campaign against the ADPr site of SARS-CoV-2 Nsp3 macrodomain (Mac1). First, we will remove molecules with a computed Tanimoto index of more than 0.6 from any available Mac1 ligand, in order to prioritize completely novel scaffolds. Second, we will perform benchmarking studies to select the most suitable target structure for virtual screening, by generating ligand-decoy sets and docking them to all the available Mac1 structures (previously optimized with Maestro) using Glide, Autodock-GPU and ICM, and then evaluating enrichment factors. We will then use our Deep Docking method to rapidly filter the chemical library for potential ligands with high binding affinities to the targets. Specifically, we will adopt a consensus strategy that has been proven successful in identifying potent and novel scaffolds from ultra-large libraries in previous studies: the initial library will be first screened using Deep Docking in combination with Glide; the identified molecules will be then screened using Autodock-GPU, and the same process will be then used for ICM. In this way, we will identify prospective consensus scoring molecules, providing a set of candidates that is highly enriched with active compounds. The resulting molecules will be then docked to Mac1 with the three programs, and subjected to two different, complementary selection procedures which we have successfully employed to identify distinct, non-overlapping active scaffolds from ultra-large libraries. A first selection will be made through consensus docking, which will be used to retain compounds that are consistently docked by multiple software, and consensus scoring will be used to prioritize candidate molecules to purchase. A second selection process will be performed by visually triaging the docking poses of top-ranked compounds separately for the three programs. The two resulting sets will be then merged to generate the final list of compounds to purchase.

Deep Docking leverages deep learning to enable efficient docking of billions of molecules, providing thus the advantages of structure-based virtual screening at a computational cost that is comparable to the one of ligand-based methods. In its consensus implementation, it represents the only structure-based method that has been successfully employed to screen more than 40 billion molecules, resulting in the discovery of hundreds of novel inhibitors of the SARS-CoV-2 main protease. The application domain of Deep Docking is virtually unlimited as it can be used in conjunction with any program, target and chemical library, and the code is entirely open-source.

Maestro, Glide, ICM

Deep Docking, Autodock-GPU

1. Gentile, F. et al. Deep Docking: A Deep Learning Platform for Augmentation of Structure Based Drug Discovery. ACS Cent Sci 6, 939–949 (2020).

2. Ton, A.-T. T., Gentile, F., Hsing, M., Ban, F. & Cherkasov, A. Rapid Identification of Potential Inhibitors of SARS- CoV-2 Main Protease by Deep Docking of 1.3 Billion Compounds. Mol Inform 39, 1–18 (2020).

3. Gentile, F. et al. Automated Discovery of Noncovalent Inhibitors of SARS-CoV-2 Main Protease by Consensus Deep Docking of 40 Billion Small Molecules. Chem Sci 12, 15960–15974 (2021).

4. Gentile, F. et al. Artificial intelligence–enabled virtual screening of ultra-large chemical libraries with deep docking. Nat Protoc 17, 672–697 (2022).

5. Radaeva, M. et al. Discovery of Novel Lin28 Inhibitors to Suppress Cancer Cell Stemness. Cancers 14, 5687 (2022).