Establishment And Application Of CB2 Ligand Prediction Model Based On Machine Learning

[Objective]Cannabinoid 2 receptor(CB2)is an important target for anti-epileptic,antiinflammatory,anti-fibrosis,and anti-osteoporosis,and due to small nervous system distribution,it can avoid spiritual side effects.Therefore,exploring new CB2 ligands has received extensive attention.However,due to the high cost and time-consuming of experimental screening strategies,we urgently need a more efficient and cost-effective method to speed up the screening and discovery of new CB2 ligands.This work aims to establish a high-precision machine learning prediction model,combined with molecular docking,molecular dynamics simulation,and other methods to predict the ligands with CB2 regulatory effects in the Chemdiv library.[Methods]1.Data set establishment;2.Calculation of molecular features(including calculation of molecular fingerprints,molecular descriptors,etc.);3.Training machine learning prediction model;4.Model performance evaluation;5.Application of SHAP analysis and interpretation model;6.Screening database build;7.Preliminary prediction and screening by machine learning models;8.Further filtering according to the physical and chemical properties of molecules:9.Further screening of compounds by molecular docking;10.Molecular dynamics simulation(analysis of the mechanism of action of compounds).[Results]1.Successfully established a high-precision machine learning prediction model.In the regression model(XGBoost+AARM),it demonstrated higher performances than the model developed by Mizra et al.(R2=0.667 vs R2=0.62).In the classification model(XGBoost+AMA),the AUC-ROC of its external verification set reached 0.917,which was significantly higher than the development by D-MPNN and Molmap,whose AUC-ROC were 0.898 and 0.912,respectively.The corresponding code is available on Github:https://github.com/ZhouHCB2/XGBoost_CB2.git.2.SHAP was applied to interpret the model,and explore CB2-active ligand characteristics.3.Established a new workflow of virtual screening using a machine learning model combined with SB VS.Through this procedure,our study successfully found 8 structurally novel ligands from the Chemdiv which could form stable complexes with CB2 protein for a long time.Additionally,after screening by molecular docking and MD simulation methods,we found that residues PHE87,SER90,HIS95,PHE183,SER285,etc.played an important role in CB2 activation.[Conclusion]A reliable machine learning prediction model was established,and this method of combining molecular fingerprints can provide new ideas for the development of other machine learning models.We also established a new workflow of virtual screening by combining machine learning model prediction with SB VS method,which can not only be used to identify CB2 ligands,but also provide a reference for the research of other similar targets.