the reference standard to the new candidate is correct, in many cases this AZD-0530 web information is valuable to assess the etiology and the importance of the DDI candidate. As we have shown previously, information provided by the different similarity scores can be implemented in the development of more complex models. Although the information is complementary, the different scoring measures showed PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/19704093 some correlation. 12 / 17 Improving Detection of Drug-Drug Interactions in Pharmacovigilance Fig 6. Examples of different pairs of similar drugs with different pharmacological profile detected by our models. Panel: methadone is similar to amitriptyline and predicted to interact with fluconazole. Panel: amitriptyline is similar to disopyramide and predicted by the 3D model to interact with gatifloxacin. Panel: citalopram, was found to be similar to disopyramide and hence, to interact with ranolazine. Panel: diltiazem was found to be similar to fluconazole and predicted to interact with imipramine. doi:10.1371/journal.pone.0129974.g006 The test using Drugdex as a reference standard showed poor performance in sets 2 and 3, where only probable and well-established DDIs were considered positives. The similarity measures, capturing chemical and pharmacological features can detect with better precision DDIs deemed as theoretical by Drugdex. The predictors are still useful pointing out possible dangerous drug combinations associated with severe outcomes. The test based on Drugs.com showed that the similarity models performed better than PRR and p-value scorings. In this set we used a DDI system 13 
/ 17 Improving Detection of Drug-Drug Interactions in Pharmacovigilance doi:10.1371/journal.pone.0129974.t003 classification based on clinical significance, related to the severity of the possible adverse events produced by the interactions. Drug phenotypic, therapeutic, structural and genomic similarity modeling have also been applied to predict DDIs based on machine learning methods. On the other hand, different types of similarity models were also previously published by our research group to predict different adverse effects or new potential DDIs of different etiology. In this study we showed the applicability of the similarity based models to improve the detection of DDIs that cause arrhythmias in pharmacovigilance data. Combining pharmacovigilance data with similarity modeling showed potential to facilitate the detection of new DDIs. In our study we integrated the data through an straightforward and simple approach that allows to obtain good PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/19705070 performance values but also assists the researcher in the decision making process. The method allows the calculation and evaluation of new drugs in an external test set. Drugs in the test can be added to the matrix M2 providing similarity information between drugs in the test and drugs in the reference standard. The method will generate for the new drug-drug candidates in the test a score based on the maximum similarity against the set of DDIs in the reference standard. However, a limitation is that our method only predicts interactions between our 162 reference standard drugs and drugs in the test. No DDIs can be generated when both drugs implicated in the interaction are different from our 162 reference standard drugs. This fact limits the applicability of the developed models. We applied similarity-based modeling to the DDI signals detected in FAERS when PRR>1 and p<.05. However, application of similarity modeling to all