Systems-based Drug Screening From Natural Products By The Prediction Of Drug Target Interaction

Natural products are known to be essential in maintaining and improving human health for thousands of years. The strategy to discover new drugs from natural products has proved to be very successful. However, due to the complexity of chemical components and mechanisms of action, a search and understanding of therapeutic molecules from natural products based on the traditional method is extremely difficult. In recent times, the confluence of spectacular advances in ADME/T prediction, systems biology and systems pharmacology has led to the discovery and development of numerous novel potential drugs and therapeutic agents for the treatment of a wide spectrum of diseases. The present work is an attempt to introduce how to integrate in silico ADME/T, pharmacokinetics, systems pharmacology, omics and systems biology for the discovery of natural products. Our study is not intended to be an exhaustive one of comprehensive approach for natural products, but rather is aimed at highlighting the invaluable role that natural products have played, and continue to play, in the drug discovery process and its future perspectives.After identifying a new drugable natural product, finding its underlining mechanisms of curing disease is important for safety and repositioning. As small chemicals/drugs play roles generally by targeting to its protein targets, drug target interaction(DTI) is being crucial in pharmacology. Recently, experimental determination of drug-target interactions remains challenging because of funding investment and difficulties of purifying proteins. The available computational models mainly focus on predicting indirect interactions or direct interactions on a small scale.The ligand and target dataset information with known binding affinity was collected from Psychoactive Drug Screening Program(PDSP) Ki database(http://pdsp.med.unc.edu/kidb.php, which is a unique resource in the public domain and provides information on the abilities of drugs to interact with an expanding number of molecular targets. In the process of building dataset, some drugs and targets were removed due to their chemical descriptors could not be evaluated. Besides, after analyzing the original Ki values, we found that there are hundreds of same Ki values in the database, such as 1000 nM and 10 000 nM, which are potential censored data. These repeat Ki values may strongly affect the accuracy of the following modeling. So we excluded the ligand-target-Ki entries with the repeats number of Ki more than 70. This threshold was selected according to the following two criteria:(1) maintain entries as many as possible;(2) exclude the censored data as many as possible. There are 2003 ligands and 209 targets without redundancy in this dataset, which were used as a benchmark dataset.DTI was quantified as binding affinity by the Ki constant value. Based on 1589 molecular descriptors and 1080 protein descriptors in 9948 ligand-protein pairs, we built in silico models by using machine learning methods, i.e., support vector(SV) machine and random forest(RF). The performances of these models were evaluated by five-fold cross-validation and compared by F test.The optimal models show good performance prediction of drug target interaction. The cross-validation coefficient of determination of R2 = 0.6079 and MSE = 7.0487 using SVM, R2 = 0.6267 and MSE = 6.5828 using RF for the test set, which indicates that both models provide a potent Ki predictability without overfitting, while the RF model performed vigorously and was more reliable as compared to SVM(F test, P<0.001). Further, by deep analyzing on the importance of descriptors provided by RF model, we discovered that 2D autocorrelation, topological charge indices and 3D-MoRSE descriptors for compound, and amphiphilic pseudoamino acid composition, autocorrelation descriptors and quasi-sequenceorder descriptors for protein are most important to Ki. Thus, RF model can be used as a potential predictor that provides a new opportunity for Ki prediction which will facilitate appropriate drug target discovery and drug toxicity evaluation in drug development.