We plan to combine the commercial Schrodinger software (https://www.schrodinger.com/) and the DrugFlow platform developed by our company (https://drugflow.com/) to conduct a hierarchical virtual screening on the Enamine database. The preliminary scheme is as follows: (1) prepare the protein and molecules using the Protein Preparation and Ligand Preparation modules embedded in DrugFlow. The preparation of protein includes adding hydrogen atoms, filling in missing residues, optimizing H-bond networks and minimizing the system using the AMBER FF14SB force field, while the preparation of molecules includes generating possible ionization states and tautomers at pH=7.0±2.0, stereoisomers and appropriate 3D conformations; (2) Adopt several negative design strategies (e.g. drug-like filters, PANIS rules, and REOS rules) to perform preliminary screening and eliminate the undesirable molecules; (3) Conduct docking-based screening (either using the classical Glide SP embedded in Schrodinger program or the deep learning-based CarsiDock embedded in DrugFlow) against above-remained molecules, and keep the molecules with the highest scores; (4) Rescore the retained molecules with the state-of-the-art deep learning-based scoring function (e.g., RTMScore or InteractionGraphNet developed in house) and retain the molecules with the highest scores; (5) Use the Inno-ADMET module in DrugFlow to evaluate the ADMET properties of remaining compounds, and removevmolecules with poor ADMET properties; (6) Cluster the remaining molecules, analyze their binding modes, and select some of them for purchase.

The core method employed here is RTMScore (J Med Chem, 2022, 65, 10691), an AI-based scoring function developed by introducing a tailored residue-based graph representation strategy and several graph transformer layers for the learning of protein and ligand representation, followed by a mix density network to obtain residue-atom distance likelihood potential. This approach was resolutely validated on the CASF-2016 benchmark, and the results indicate that it can outperform almost all of the other state-of- the-art methods in terms of both the docking and screening powers. Further evaluation confirms the robustness of the method that can not only retain its docking power on cross-docked poses but also achieve improved performance as a rescoring tool in larger-scale virtual screening.

Our approach has three key ingredients: (1) published and in-house developed artificial intelligence-based models (e.g., RTMScore and InteractionGraphNet) that achieve state-of-the-art results in many well-established benchmarks; (2) an AI-driven drug design platform DrugFlow, which allows us to conduct the aforementioned screening protocol and try other ideas very efficiently and professionally; (3) our proposed screening strategy has been repeatedly tested and prevailed in other drug discovery projects, some examples may be found in (Eur J Med Chem, 2022, 237,114382; Acta Pharmacol Sin, 2022, 43, 229).

Schrodinger (https://www.schrodinger.com/)

DrugFlow (https://drugflow.com/)

RDkit (http://www.rdkit.org/)

OpenBabel (https://github.com/openbabel/openbabel/)

MDAnalysis (https://github.com/MDAnalysis/mdanalysis/)

Prody (http://prody.csb.pitt.edu/downloads/)

PyTorch (https://pytorch.org/)

DGL (https://github.com/dmlc/dgl/)

Chao Shen, Xujun Zhang, Yafeng Deng, Junbo Gao, Dong Wang, Lei Xu, Peichen Pan, Tingjun Hou, Yu Kang. J Med Chem. 2022, 65(15), 10691-10706.

Dejun Jiang, Chang-Yu Hsieh, Zhenxing Wu, Yu Kang, Jike Wang, Ercheng Wang, Ben Liao, Chao Shen, Lei Xu, Jian Wu, Dongsheng Cao, Tingjun Hou. J Med Chem. 2021, 64(24), 18209-18232.