Discovery Of Potential RIPK1 Inhibitors By Machine Learning And Molecular Dynamics Simulations

Receptor interaction serine/threonine protein kinase 1(RIPK1)is a key regulator of inflammation and cell death and is also an important target associated with various diseases such as cancer,psoriasis,ulcerative colitis,rheumatoid arthritis,Alzheimer’s disease(AD),multiple sclerosis(MS)and amyotrophic lateral sclerosis(ALS),and inhibition of RIPK1 has been shown to be a promising treatment for these diseases.AD is one of the most common neurodegenerative diseases as well as the leading dementia disease in the elderly,it has become one of the top 10 causes of death in the world and studies have shown that AD is strongly associated with necroptosis death regulated by RIPK1.Furthermore,other recent study has shown that SARS-COV-2accelerates virus proliferation by promoting the activation of RIPK1 and that suppression of RIPK 1 reduces the viral load of SARS-Co V-2 in lung organs.Although inhibition of RIPK1 activity can treat a variety of related diseases,up to now,no RIPK1 inhibitors have been approved by the Food and Drug Administration(FDA).In this context,the continued search for high efficacy and low toxicity potential inhibitors are very necessary and of significant importance.Given this,this paper uses four machine learning algorithms,Random Forest(RF),Extreme Random Tree(ET),Extreme Gradient Boosting Tree(XGBoost)and Light Gradient Boosting Machine(Light GBM),combined with Extended Connectivity Fingerprinting(ECFP)to construct four QSAR classification models for predicting potential small molecule inhibitors of this protein in a known chemical space.Subsequently,the performance was checked by the relevant evaluation metrics,and all four models performed remarkably well as indicated by the seven evaluation criteria,so the full prediction results for all models were used.In order to improve the hit rate of active molecules as well as drug-like properties,some properties such as Log P,molecular weight(MW),rotatable bonds(ROTB)were calculated using RDKit for the preliminary filtering,which eventually yielded 9769 possible active molecules in a subset of about 130,000 in the ZINC database.Subsequently,further screening by molecular docking was performed and the complete docking results were given.Considering that simply selecting molecules that matched the pharmacological properties according to the scoring order might result in redundant analogues,a cluster analysis was performed.As indicated by the silhouette coefficients and total sum of squares,six groups were clustered and one molecule in each group were selected,for a total of six active molecules whose pharmacological properties also meet the corresponding requirements.Finally,the final hit molecules were subjected to dynamics simulations and a series of evaluation metrics were calculated to verify their binding stability.In addition,the Shapley Additive Explanation(SHAP)was employed to further illustrate how these molecular fingerprint fragments influence model decisions.According to this method 1855 bit was identified as the most important molecular fingerprint fragment,followed by 819,which provided the basis for fragment-based drug design(FBDD).The current data indicate that the six screened small molecules are very promising to target RIPK1 for the treatment of relevant diseases,and molecular dynamics simulations also further confirm their binding stability to the target protein.It is worth noting that the small molecule ZINC000085897746 is found in Musa acuminate,a member of the Banana family,which means that it is a natural product and therefore deserves high priority for research.