Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2017 Jun 8;18(1):44.
doi: 10.1186/s40360-017-0153-6.

Data-driven Prediction of Adverse Drug Reactions Induced by Drug-Drug Interactions

Affiliations
Free PMC article

Data-driven Prediction of Adverse Drug Reactions Induced by Drug-Drug Interactions

Ruifeng Liu et al. BMC Pharmacol Toxicol. .
Free PMC article

Abstract

Background: The expanded use of multiple drugs has increased the occurrence of adverse drug reactions (ADRs) induced by drug-drug interactions (DDIs). However, such reactions are typically not observed in clinical drug-development studies because most of them focus on single-drug therapies. ADR reporting systems collect information on adverse health effects caused by both single drugs and DDIs. A major challenge is to unambiguously identify the effects caused by DDIs and to attribute them to specific drug interactions. A computational method that provides prospective predictions of potential DDI-induced ADRs will help to identify and mitigate these adverse health effects.

Method: We hypothesize that drug-protein interactions can be used as independent variables in predicting ADRs. We constructed drug pair-protein interaction profiles for ~800 drugs using drug-protein interaction information in the public domain. We then constructed statistical models to score drug pairs for their potential to induce ADRs based on drug pair-protein interaction profiles.

Results: We used extensive clinical database information to construct categorical prediction models for drug pairs that are likely to induce ADRs via synergistic DDIs and showed that model performance deteriorated only slightly, with a moderate amount of false positives and false negatives in the training samples, as evaluated by our cross-validation analysis. The cross validation calculations showed an average prediction accuracy of 89% across 1,096 ADR models that captured the deleterious effects of synergistic DDIs. Because the models rely on drug-protein interactions, we made predictions for pairwise combinations of 764 drugs that are currently on the market and for which drug-protein interaction information is available. These predictions are publicly accessible at http://avoid-db.bhsai.org . We used the predictive models to analyze broader aspects of DDI-induced ADRs, showing that ~10% of all combinations have the potential to induce ADRs via DDIs. This allowed us to identify potential DDI-induced ADRs not yet clinically reported. The ability of the models to quantify adverse effects between drug classes also suggests that we may be able to select drug combinations that minimize the risk of ADRs.

Conclusion: Almost all information on DDI-induced ADRs is generated after drug approval. This situation poses significant health risks for vulnerable patient populations with comorbidities. To help mitigate the risks, we developed a robust probabilistic approach to prospectively predict DDI-induced ADRs. Based on this approach, we developed prediction models for 1,096 ADRs and used them to predict the propensity of all pairwise combinations of nearly 800 drugs to be associated with these ADRs via DDIs. We made the predictions publicly available via internet access.

Keywords: Adverse drug reactions; Drug-drug interactions; Drug-protein interactions; Pharmacovigilance; Synergistic drug-drug interactions.

Figures

Fig. 1
Fig. 1
Schematic illustration of the drug-protein interactions necessary for drug-drug interaction (DDI)-induced adverse drug reactions (ADRs). Drugs i and j interact with proteins α, β, γ, δ, ε, and ζ to induce both therapeutic effects as well as adverse effects Ψ, Ω, and Φ. In Case I, simultaneous drug interaction with both proteins α and β is necessary for a DDI to induce ADR Φ. Because drug i interacts with α but not β and drug j interacts with β but not α, no DDI occurs when the two drugs are administered individually. However, when the two drugs are co-administered, the requirement of simultaneous drug interaction with both α and β, and hence the condition for DDI-induced ADR Φ, are satisfied. In Case II, an existing adverse effect Ω caused by drug j is enhanced by drug i interacting with α, aggravating the adverse effect to a degree that is not possible by drug j alone
Fig. 2
Fig. 2
Binary bit string representation of genome-wide drug-protein interaction profiles of individual drugs and drug combinations. In the bit strings, the drug interaction with each protein is encoded by four bits representing binding (B), inhibition (I), activation (A), and catalysis (C). For any drug, if information for an interaction is present in the STITCH database with at least a medium confidence level, the corresponding bit in the string is turned on (assigned a value of 1); otherwise, it is turned off (assigned a value of 0). To generate a drug pair-protein interaction profile ĝ(d i d j) from the constituent drug-protein interaction profiles ĝ(d i) and ĝ(d j), we implemented the logical OR operation. Thus, we turned the bit off when neither drug interacts with a protein; otherwise, we turned it on when interactions for either or both drugs are present
Fig. 3
Fig. 3
Sample selection for training and evaluation of three types of model (I–III). For type I models, we used 10–90% of all drug pairs positive for an ADR as the positive class for model training, with half of the other drug pairs in the TWOSIDES database used as the baseline class. We used the remaining drug pairs as a test set for assessing model performance. Sample selection for type II models was the same as that for type I models except that we moved half of the positive samples in the test set into the baseline class of the training set to provide known false negatives. Sample selection for type III models was the same as that for type I models except that some randomly selected baseline samples in the test set were moved into the positive class in the training set to provide false positives. The number of false positives was equal to the number of true positives in the training set
Fig. 4
Fig. 4
Area under the receiver operating characteristic curve (AUC) derived from cross-validation studies by using three types of model (I–III) for the following adverse drug reactions (ADRs): rheumatic heart disease (Unified Medical Language System code C0035439), heat stroke (C0018843), spontaneous abortion (C0000786), and vestibular disorder (C0042594). These ADRs were correspondingly associated with 78, 331, 897, and 1022 drug pairs. For all panels, the horizontal axis represents the percentage of positive samples for an ADR used in the model training. The error bars show ±1 standard deviation from 50 simulations, using randomly selected training and test set samples
Fig. 5
Fig. 5
Weights of the drug-protein interactions contributing to the rheumatic heart disease (C0035439) model. The positions on the horizontal axis represent specific drug-protein interactions; the heights of the blue bars denote the weights of the drug-protein interactions in the ADR model. The figure demonstrates that the ADR model is not dominated by only one or a few drug-protein interactions
Fig. 6
Fig. 6
Area under receiver operating characteristic curve (AUC) for 1,096 models of adverse drug reactions (ADRs), estimated from 10-fold cross-validation studies. Each circle represents an ADR prediction model with its AUC value on the vertical axis and the number of drug pairs that induced this ADR on the horizontal axis. The ADR model with the most drug pairs was nausea (Unified Medical Language System code C0027497), with >16,000 constituent pairs
Fig. 7
Fig. 7
Impact of increasing the normalized DDI score threshold on average measures of prediction performance for 1,096 ADR models evaluated by 10-fold cross validation. The error bars indicate ± 1 standard deviation. The faint lines, which represent PPV traces of several ADRs, show that all initially increase with the DDI score threshold, but some reach a maximum and then fluctuate or drop off with further increases in the DDI score threshold
Fig. 8
Fig. 8
Screenshot of the AdVerse effects Of Interacting Drugs Database (AVOID-DB). The database can be queried with the name of a single drug to retrieve all predicted DDI-induced ADRs associated with the drug, or with two drug names to retrieve all DDI-induced ADRs associated with the two drugs. Similarly, the database can be queried with one or more specific ADRs to retrieve all drug pairs causing the ADRs via synergistic DDIs. The drug and ADR names, which constitute a controlled vocabulary, can be selected from among the available names in the database. When available, the drug and ADR names are cross-linked to DrugBank and National Library of Medicine resources for further information. The normalized DDI (Φ,d i d j)-scores are color-coded from dark to light red for visual guidance. The Web page is accessible at http://avoid-db.bhsai.org
Fig. 9
Fig. 9
Two-way clustering of predicted DDI-induced ADRs. We included all drugs on the market for which we could make predictions. Each element in the matrix represents the number of ADRs caused by each drug pair. The drugs roughly formed four clusters (C1–C4), where cluster C1 was associated with the most ADRs and cluster C4 with the least
Fig. 10
Fig. 10
Class effect between proton pump inhibitors and non-steroidal anti-inflammatory drugs. All non-steroidal anti-inflammatory drugs exhibited DDI-induced ADRs with the proton pump inhibitors lansoprazole, pantoprazole, and omeprazole. ATC, anatomical therapeutic chemical classification; ADR, adverse drug reaction; NSAIDs, non-steroidal anti-inflammatory drugs; PPIs, proton pump inhibitors
Fig. 11
Fig. 11
DDI-induced adverse drug reactions associated with specific drug classes. a Effects of jointly administering anti-infective (ATC code J) and anti-convulsant (ATC code N) drugs. b Effects of jointly administering anti-diabetics (ATC code A) and non-steroidal anti-inflammatory drugs (NSIADs, ATC codes M, N, and S) or proton pump inhibitors (ATC code A). ATC, anatomical therapeutic chemical classification; ADR, adverse drug reaction
Fig. 12
Fig. 12
Two-way clustering of spontaneous abortion DDI scores. We included all drugs on the market for which we could make predictions. Each element in the matrix represents the DDI(Φ,d i d j) score as calculated from equation (3) for causing spontaneous abortion (Unified Medical Language System C0000786) for each drug pair. The compounds roughly formed three drug-pair categories: N, those that do not cause spontaneous abortion; M, those with moderately high DDI(Φ,d i d j) scores; and S, those associated with the highest scores

Similar articles

See all similar articles

Cited by 6 articles

See all "Cited by" articles

References

    1. Lazarou J, Pomeranz BH, Corey PN. Incidence of adverse drug reactions in hospitalized patients: a meta-analysis of prospective studies. JAMA. 1998;279(15):1200–5. doi: 10.1001/jama.279.15.1200. - DOI - PubMed
    1. Berlin JA, Glasser SC, Ellenberg SS. Adverse event detection in drug development: recommendations and obligations beyond phase 3. Am J Public Health. 2008;98(8):1366–71. doi: 10.2105/AJPH.2007.124537. - DOI - PMC - PubMed
    1. Hanton G. Preclinical cardiac safety assessment of drugs. Drugs R D. 2007;8(4):213–28. doi: 10.2165/00126839-200708040-00002. - DOI - PubMed
    1. Magro L, Moretti U, Leone R. Epidemiology and characteristics of adverse drug reactions caused by drug-drug interactions. Expert Opin Drug Saf. 2012;11(1):83–94. doi: 10.1517/14740338.2012.631910. - DOI - PubMed
    1. Beijnen JH, Schellens JH. Drug interactions in oncology. Lancet Oncol. 2004;5(8):489–96. doi: 10.1016/S1470-2045(04)01528-1. - DOI - PubMed

Publication types

Feedback