Computational protocol: Systems Pharmacology Dissecting Holistic Medicine for Treatment of Complex Diseases: An Example Using Cardiocerebrovascular Diseases Treated by TCM

Similar protocols

Protocol publication

[…] To obtain cerebrovascular diseases (CBVDs) and CVDs target profiles, a comprehensive drug targeting approach integrating the text mining, database search, and chemometric analysis was applied. First, a full-text data mining was carried out in TCMSP database ( to derive the molecular target information. Second, a chemical fingerprint-based Similarity Ensemble Approach was applied to obtain the potential targets ( Third, an in-house LTC [] was further introduced for expanding the target pool. The targets from different sources were connected to database UniProt ( for target name standardization, which were further subjected to PharmGkb [], Therapeutic Target Database [], and the Comparative Toxicogenomics Database [] to delete noise and errors and to ensure the quality of target database. [...] In an effort to stringently assess the relationships between compounds and corresponding targets, for the first time, a new PreAM model was built in this work. The data sets of drug structures and protein sequences for drug-target interactions (DTIs) with known action modes were retrieved from DrugBank database (, accessed on October 1, 2013). Here, we separated these DTIs into two categories based on their action modes: (1) activated DTIs, where the description label covers “agonist,” “activator,” “inducer,” “stimulator,” and “partial agonist”; (2) inhibited DTIs, where the action label contains any of the keywords “inhibitor,” “antagonist,” “inactivator,” “negative modulator,” “partial antagonist,” “suppressor,” and “reducer actions.” In total, 6,006 DTIs (including 1,251 activated DTIs and 4,755 inhibited DTIs) were used as benchmark data (Table S3).To characterize the interactions of drug and protein, drug structures and protein sequences were converted into numerical descriptors by employing DRAGON program ( and PROFEAT WEBSEVER (, respectively (see details in Table S3). The multiple DTIs were represented by concatenating these chemical and protein descriptors and the minimal-redundancy-maximal-relevance (mRMR) was applied as a variable selection strategy to recognize the best combination descriptors that are most relevant to obtain the models with the highest predictive power. The first 100 descriptors were used in subsequent study.The random forests (RF) algorithm ( was trained to generate a nonlinear classifier tailored to DTIs with known action modes. The accuracies to overall, activation, inhibition were used to measure the performance of the model. The derived model shows impressive performance of prediction for drug-target interactions, with an overall accuracy of 97.3%, an activated prediction accuracy of 87.7%, an inhibited prediction accuracy of 99.8%. (see details in Figure S2). […]

Pipeline specifications