Finally, serum concentrations may not be representative of tissue levels

eminformatic methodologies, such as protein-structure-based and ligand-based methods, have been used in the detection of DDIs. Cheminformatics provides a useful approach through the use of 2D/3D QSAR , homology modeling and molecular docking. These methods can infer similarity between sets of drugs and study 71939-50-9 custom synthesis possible interactions with pharmacodynamics or pharmacokinetic targets. In previous work, we have leveraged cheminformatics to construct general models of DDIs. On the other hand, scientific literature and pharmacovigilance databases are additional sources with important implications in DDI discovery. Percha et al. mined the scientific literature to detect DDIs through the extraction of gene-drug relationships. Mining electronic health records or the FDA’s Adverse Event Reporting System is an alternative for the discovery of DDIs. In fact, Tatonetti et al. recently provided an important source of DDI candidates, the TWOSIDES database, through mining FAERS. However, analysis of pharmacovigilance data is still very challenging and rampant confounding leads to high false positive rates. Alternatively, cheminformatic methods can be applied to rank the DDI candidates extracted from a pharmacovigilance study. These methods offer the possibility to study the final candidates from the point of view of the molecular structure, pharmacological action or adverse effects comparison. Similarity-based methods were useful to rank drug candidates extracted from pharmacovigilance data mining that produce some adverse events, such as rhabdomyolysis and pancreatitis. In this paper, we systematically apply six different similarity-based techniques to evaluate drug interaction hypotheses mined from pharmacovigilance data. The objective of the current study is PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/19734939 to improve the detection of DDIs in the TWOSIDES database using methodologies we recently developed based on the application of similarity-based modeling. When applied to the TWOSIDES database a reference standard of DDIs that produce arrhythmia, we measured: 1) enrichment factor provided by TWOSIDES, and 2) performance when we rank the set of DDI candidates using proportional reporting ratio, p-values, and different similarity-based models. As is demonstrated by our results, the implementation of cheminformatic models in pharmacovigilance data is useful in DDI signal detection and decision making process. 2 / 17 Improving Detection of Drug-Drug Interactions in Pharmacovigilance Fig 1. Flowchart with the different steps implicated in the study. doi:10.1371/journal.pone.0129974.g001 Methods DDI reference standard We collected a reference standard with 149 DDIs present in the intersection of both DrugBank and Veterans Association Hospital database. The collected DDIs produced the effect of arrhythmias and related terms, such PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/19736622 as QT prolongation or increased heart rhythm. In our reference standard there are DDIs with different levels of documentation, from “well established through controlled studies” to “theoretical interactions but pharmacological reasons lead clinicians to recognize the possible interaction”. The 149 DDI pairs comprised 162 drugs and were included in a 162162 drug-drug matrix called M1. We codified the 149 reference standard DDIs in M1 with value 1 in each respective cell, and the non-DDIs with value 0. 3 / 17 Improving Detection of Drug-Drug Interactions in Pharmacovigilance Fig 2. Flowchart including the steps implicated in the calculation of different similarity measures. Drugs w