Based On The Molecular Structure Prediction Methods Receptor May Be

Prediction of direct targets of small molecules is of great value in drug design,chemical biology and other research areas. With the knowledge of known ligands ofprotein targets archived in the databases of ChEMBL, BindingDB and PubChem,machine learning was used to predict the direct target candidates of small molecules.Raddeanin A, Erianin, Flavokawain A/B/C can inhibit the tumor cell growtheffectively. We screened molecules through molecular similarity searching andsubstructure searching for Raddeanin A, Erianin,Flavokawain A/B/C, and classed themto target classes according to their targeting target for screened molecules. Thenumber of molecules were counted for each target classes. The targets were to aspredicted target that having the most screened molecules.Most of the predicted targetsrelate to cancer in our research. We could predict the targets for query molecule veryeffectively through counting the screened molecule for each target class by structuresimilarity searching and molecular substructure searching.A na ve Bayesian classifier model which was built upon molecular fingerprints ofthe ligands and decoys was found to be effective in predicting the ligands of a target.Tests on the Directory of Useful Decoys (DUD) showed that the overall true positiveand negative prediction rates are better than90%for most targets. The list of possibletargets of a given molecule was produced by integrating the prediction results of allavailable target models. The na ve Bayesian classifier model that using all moleculesof PubChem database as background and all ligands of each target class in the DUD asligands of train set preformed better than using the DUD for train set.40na veBayesian classifier model which was built upon molecular fingerprints of the ligandsand decoys for the DUD and were saved. Tests using all ligands of DUD showed thatthe31in40target class predicted true rate better than80%.A na ve Bayesian classifier model which was built upon molecular fingerprints ofall molecules in PubChem database as background and all molecule targeting toAngiotensin-Converting Enzyme in ChEMBL database as active molecules for train set. The model could distinguish ligands and decoys of ACE target class in the DUD. Alltarget classes were classified according to their targeting targets. The molecules werescreened for each target classes and the target classes were deleted that having only fewmolecules. The na ve Bayesian classifier model for each target classes were builtmolecular fingerprints of all molecules in PubChem database as background and allscreened molecule of each target classes in ChEMBL database as active molecules fortrain set, and the models were saved.Tests on the Directory of Useful Decoys showedthat most molecule’s target were predicted correctly.