Identifying Drug–Drug Interactions in Spontaneous Reports Utilizing Signal Detection and Biological Plausibility Aspects

Abstract

Translational approaches can benefit post-marketing drug safety surveillance through the growing availability of systems pharmacology data. Here, we propose a novel Bayesian framework for identifying drug–drug interaction (DDI) signals and differentiating between individual drug and drug combination signals. This framework is coupled with a systems pharmacology approach for automated biological plausibility assessment. Integrating statistical and biological evidence, our method achieves a 16.5% improvement (AUC: from 0.620 to 0.722) with drug-target-adverse event associations, 16.0% (AUC: from 0.580 to 0.673) with drug enzyme, and 15.0% (AUC: from 0.568 to 0.653) with drug transporter information. Applying this approach to detect potential DDI signals of QT prolongation and rhabdomyolysis within the FDA Adverse Event Reporting System (FAERS), we emphasize the significance of systems pharmacology in enhancing statistical signal detection in pharmacovigilance. Our study showcases the promise of data-driven biological plausibility assessment in the context of challenging post-marketing DDI surveillance.

Study Highlights

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

Recent efforts in pharmacovigilance integrated systems pharmacology into signal detection for single drugs, but not in the context of drug–drug interactions (DDIs).

• WHAT QUESTION DID THIS STUDY ADDRESS?

The study proposed a signal detection algorithm for adverse DDI surveillance coupled with a biological information network, aiming to distinguish signals arising from drug combinations rather than constituent drugs and automate their biological plausibility assessment.

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

This study provides a novel framework for detecting DDI signals using disproportionate reporting in the FDA Adverse Event Reporting System (FAERS) database combined with a biological information network. Also, two potential DDI signals related to QT interval prolongation were identified.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

This study provides a data-driven approach to support signal refinement in pharmacovigilance, specifically focusing on DDIs. By utilizing a network of biological information on drug targets, metabolizing enzymes, and transporters, the proposed framework aims to improve statistical signal detection and address the complex landscape of DDIs.

Polypharmacy rates have been increasing in the Western world over the last few decades, with important implications for patient safety and a higher risk of drug–drug interactions (DDIs). A DDI occurs when a drug's effect on the body changes in the presence of another drug and can be categorized as either pharmacodynamic (i.e., at the level of drug receptors or other drug targets) or pharmacokinetic (i.e., due to changes in the drug absorption, metabolism or excretion).¹ DDIs can be beneficial or adverse. Adverse DDIs can lead to adverse drug reactions (for example, bleeding with warfarin precipitated by starting amiodarone, a CYP2C9 enzyme inhibitor) or may reduce efficacy (for example, loss of contraceptive efficacy with the use of an enzyme inducer such as carbamazepine). For the remainder of this paper, we will focus on adverse DDIs.

When a DDI is clinically observable, the possible outcome is either a lack of drug efficacy or an adverse drug reaction (ADR) that cannot be attributed to the individual drugs separately. Pre-marketing clinical trials and drug interaction studies during drug development are inadequate to fully capture adverse effects linked to DDIs, which seem to contribute to a significant proportion of DDI-related adverse effects in clinical practice.^{2, 3} The prediction of clinically relevant DDIs is also a difficult task due to human inter-individual variability (which can be due to many factors including pharmacogenetic variation, concomitant disease, smoking, and diet, to name a few) and differences between how the human body acts compared with animal or in silico models. Thus, post-marketing surveillance is essential to identify complications due to interacting drug combinations.

Post-marketing safety surveillance databases, called spontaneous reporting system (SRS) databases, compile reports that contain information on suspected drug complications. Examples include VigiBase (maintained by the Uppsala Monitoring Centre), the US Food and Drug Administration Adverse Event Reporting System (FAERS) database, and Eudravigilance from the European Medicines Agency. Disproportionality analysis is a popular approach for signal detection in SRS data that aims to identify “unexpectedness” in reported data by comparing the observed rate of adverse event (AE) occurrence to the expected one using the background of the rest of the database.⁴ While these databases and methods remain at the forefront to identify drug safety issues in the post-marketing setting, there are multiple inherent challenges. Those include the inability to get an accurate estimate of the incidence or reporting rates, the potential presence of reporting biases due to the voluntary nature of reporting (i.e., either under- or over-reporting), stimulated reporting, duplicate reporting arising from multiple sources reporting the same adverse incident, data quality issues, data loss when moving from an unstructured to a structured format of information in SRS data, data incompleteness, etc.⁵

For DDI surveillance, existing signal detection algorithms (SDAs) test the degree of “unexpectedness” of associations between two drugs and an AE, assuming that the baseline model (i.e., the contributions of the individual effect of each drug in the absence of an interaction) is either multiplicative⁶ or additive.^{6, 7} Other approaches have compared the disproportionality measure for a drug combination (e.g., EB05 and EB95 scores) relative to the measures for the individual drugs for a specific AE to identify signals of potential DDIs.⁸ At the same time, the calculated measure does not allow us to directly infer whether, and which of, the individual drugs give rise to unexpected reporting on their own compared with their expected background rate. However, we would expect that, for example, a pharmacokinetic DDI would involve disproportionate reporting rates for the victim drug and the combination, while the reporting rate for the perpetrator drug would not differ from the background rate. In contrast, a pharmacodynamic interaction would involve both unexpected reporting rates for the two drugs individually, plus a departure from the baseline additive model when the two drugs are used concomitantly.

The drug safety signal lifecycle involves several steps following the detection of a signal, such as signal prioritization, signal evaluation and, if appropriate, risk communication and management interventions.⁹ The processes of signal evaluation rely heavily on in-depth manual clinical review to assess the biological plausibility of the identified signal. The well-known Bradford-Hill criteria are traditionally used at this stage.^{10, 11} Some previous efforts attempted to integrate some biological aspects into the signal generation process for DDIs by considering metabolic enzymes and transporters to assess the plausibility of a signal from a mechanistic perspective.¹² However, as DDIs, apart from pharmacokinetic, can also be pharmacodynamic, leveraging supporting information relevant to the biological mechanisms that might be involved in the generation of this signal in the first place could enable mechanism-based filtering of signals.¹³

Recent efforts aimed to integrate systems pharmacology aspects into pharmacovigilance signal detection have included evaluating connections between drug targets and AEs as putative AE mechanistic pathways to identify signals of single-drug side effects.^{14, 15} Other studies have attempted to predict drug safety profiles in the post-marketing setting based on target similarity with comparator drugs that have known safety profiles.^16-18 In the context of DDIs, the number of possible drug combinations in the clinical setting and the amount of data that are constantly accumulating indicate the potential of coupling data mining and biological plausibility aspects to incorporate a mechanistic understanding of the associations in data mining methodologies. However, target-AE associations have not been utilized in related studies to assess the performance of SDAs for DDI surveillance. This approach could potentially limit the number of spurious associations that are identified as signals and help uncover novel DDIs that cannot be solely detected in SRS data.

The aims of this study were: (a) to advocate an SDA for adverse DDIs that could produce a pharmacologically motivated output by detecting increasing reporting rates in SRS data while being able to distinguish signals that might arise from constituent drugs; and (b) to utilize a systems pharmacology framework of established associations between biological nodes (i.e., drug targets, drug ingredients, and AEs) to refine the signals of potential DDIs and automate an assessment of their biological plausibility.

METHODS

Data sources

CRESCENDDI reference set

We used CRESCENDDI, an open-access reference set for adverse DDIs, as the source of positive and negative controls for DDIs.¹⁹ All controls were additionally stratified using information from CRESCENDDI regarding their individual drug safety profile (i.e., single-drug ADRs). For example, the combination of paroxetine and ibuprofen can lead to an increased risk of bleeding.²⁰ At the same time, both drugs are individually associated with hemorrhagic events as adverse effects. Thus, it was possible to classify the control based on both the combined behavior (i.e., whether there is evidence that the drug combination interacts leading to a specific AE) and the individual ones (i.e., whether each of the drugs is separately associated with the AE).

For a DDI control to be classified for the behavior of either of its constituent drugs (Table 1), we checked whether the ADR list from the product information of the drug contained any closely related MedDRA Preferred Terms (PTs) to the DDI control's PT. These “closely related” PTs were found under the same SMQ groups as the PT under consideration. For example, if a DDI control was associated with Hypertension (PT), the product information ADR lists of the constituent drugs were checked for the presence of any PTs from the SMQ “Hypertension” (e.g., hypertensive crisis). This classification enabled a stratified analysis to assess the ability to detect disproportionate reporting for the different categories of individual drug behavior: (i) Both (i.e., both drugs are independently related to the AE), (ii) One (i.e., only one of the constituent drugs is linked to the AE), and (iii) None (i.e., neither of the drugs is related to the AE).

Table 1. Categories of individual drug behavior for control stratification

Category	Description
00,0	A DDI negative control where the adverse event (AE) is not a side effect for either of the drugs
00,1	A DDI positive control where the AE is not a side effect for either of the drugs
10,0	A DDI negative control where the AE is a side effect only for the first but not the second drug
10,1	A DDI positive control where the AE is a side effect only for the first but not the second drug
01,0	A DDI negative control where the AE is a side effect only for the second but not the first drug
01,1	A DDI positive control where the AE is a side effect only for the second but not the first drug
11,0	A DDI negative control where the AE is a side effect for both drugs
11,1	A DDI positive control where the AE is a side effect for both drugs

FAERS database

We curated and standardized the publicly available version of the FAERS database using the Adverse Event Open Learning through Universal Standardization (AEOLUS) process²¹ and considered spontaneous reports covering the period from the 1st quarter of 2004 through to the 4th quarter of 2018. Drug concepts were standardized to the RxNorm Ingredient level terms and AE concepts to MedDRA PTs, to ensure compatibility with the reference set. The curated dataset contained 9,203,239 reports that included at least one drug and one AE, with 3,973,749 (43.18%) reports mentioning more than one drug. Each drug was considered equivalent in the analysis irrespective of its reported role (i.e., primary suspect; secondary suspect; concomitant; and interacting).

Open Targets

Open Targets²² is a freely available resource that combines multiple public data sources regarding potential therapeutic drug targets and associated information, including evidence regarding target safety (i.e., associations between drug targets and potential unintended adverse consequences). After downloading the core annotations for drug molecules and targets as Parquet files (version 22.06), we extracted the following information:

• (i)

  Drug-target associations

• (ii)

  Target-AE associations representing target safety liabilities.

Drugs were mapped to RxNorm Ingredients, while target mappings were available as both Ensembl stable IDs (e.g., ENSG00000157764) (as the primary identifiers) and UniProtKB accession numbers (e.g., P15056). In Open Targets, AE terms were reported inconsistently in various ontologies (i.e. Experimental Factor Ontology (EFO), Human Phenotype Ontology (HPO), Gene Ontology (GO), Mondo Disease Ontology, Orphanet Rare Disease Ontology). Where possible, we retrieved relevant MedDRA synonym terms via querying the OLS WEBAPI (https://www.ebi.ac.uk/ols/docs/api) or using a manually curated mapping table for Human Phenotype Ontology (HPO)²³ entities (Table S1).

DrugBank

DrugBank²⁴ is an open-access knowledge base that contains information related to drugs and drug targets. We downloaded DrugBank version 5.1.9 (https://go.drugbank.com/releases/latest#full) as an XML file and extracted drug enzyme, transporter, and target data. Apart from therapeutic (i.e., primary) targets, DrugBank also included secondary ones that were considered to complement primary target information for drugs from Open Targets.

STRING

STRING 11.5²⁵ is an open-access database that contains protein–protein interactions (PPIs) in various organisms mined from multiple evidence channels. Each PPI is associated with a confidence score in the dataset, which is computed by combining the probabilities from the different channels and correcting for the probability of randomly observing an interaction. We only selected PPIs of higher confidence in humans (i.e., scores over 700 out of a maximum score of 1,000) for further consideration.

In Figure 1a, we present a diagram that illustrates the steps followed in this study, providing a visual representation of the proposed workflow, and delineating the specific utilization of information from each data source that is described in this section.

Figure 1

(a) Proposed workflow that illustrates the methodological steps (swim lane) and specifies the utilization of information from each data source (depicted in pink boxes) considered in this study. Orange boxes represent signal detection algorithms (SDAs) based on disproportionality analysis in spontaneous reporting systems data, and blue boxes represent network metrics (derived from the Biological Attribute Network) that encapsulate biological information. The logistic regression step (red box) combines input from SDA and network metrics (black dotted lines). The Receiver Operating Characteristics (ROC) Analysis (in yellow boxes) serves as the evaluation method. Black arrows denote operation processes, while red dotted arrows represent evaluation processes. (b) Illustrative presentation of the Biological Attribute Network.

Development of a pharmacology-driven signal detection algorithm for adverse DDIs

Considering a drug–drug-event (DDE) triplet, $D_1-D_2-\mathrm{AE}$, FAERS reports can be categorized based on the presence or absence of: (i) the first drug ($D_1$); (ii) the second drug ($D_2$); and (iii) the event ($\mathrm{AE}$), thus enabling the storage of the various category counts in a 4-by-2 contingency table (Figure S1a).

We developed a novel SDA for detecting signals of two-way DDIs in SRS data (Figure S1b). The focus of this novel approach is on detecting signals related to two-way DDIs in SRS data and distinguishing them from signals that arise due to individual complications of the constituent drugs. Given the presence or absence of either of the drugs, $D_1$ and $D_2$, there are four (potentially identical) rates of occurrence of the event (AE). There are eight distinct hypotheses for which of the rates differ (and which, if any, are identical): each hypothesis relates to whether each of the individual drugs gives rise to complications and whether a two-way DDI exists. By assuming a Beta-Binomial model, we assigned Beta prior distributions to the different rates in the context of each hypothesis. We defined the hyperparameters of these priors using an empirical Bayes method. For each DDE triplet, we calculated the posterior probability (and so log-likelihood) for a DDI being present by summing the probabilities related to all hypotheses that relate to the rate of the AE being reported when both drugs are present being different from the rate of the AE being reported in all cases other than when both drugs are present. The log-likelihood ratio and log posterior odds ratio were the chosen metrics to observe a changing or an increased probability of AE occurrence indicative of a DDI. A detailed mathematical description of the algorithm and the prior hyperparameter estimation method are available in the Supplementary Information.

Performance assessment—model validation

To assess the performance of the novel SDA, we implemented the following evaluations using Receiver Operating Characteristic (ROC) analysis: (1) a sensitivity analysis to check the impact of the hyperparameters of the prior distributions on the SDA performance (see Supplementary Information); (2) a comparison of the novel SDA with three other existing methodologies for DDI surveillance, namely Omega,⁷ the Interaction Signal Score⁸ (IntSS), and delta_add⁶; (3) PT-restricted analysis for specific AEs of interest with a sufficient number of controls (see Table S2): bradycardia; gastrointestinal (GI) hemorrhage; hemorrhage; hypertension; hypoglycemia; myopathy; QT prolongation; rhabdomyolysis; and torsade de pointes; and (4) stratified analysis based on the individual drug safety profile by utilizing the classification of controls from CRESCENDDI.

A systems pharmacology framework for signal refinement

We constructed the Biological Attribute Network (Figure 1b) by integrating multiple types of information pertinent to systems pharmacology. The network included the following types of nodes: (a) drug ingredients (RxNorm/RxNorm Extension concepts); (b) AEs (MedDRA PT concepts); (c) targets (Ensembl/UniProt/HGNC IDs); (d) enzymes; and (e) transporters (UniProt/HGNC IDs). The links between the nodes represented: (a) drug-target information from Open Targets and DrugBank; (b) target-AE associations from Open Targets; and (c) target-target associations (i.e., PPI) data from the STRING database.

We considered three different measures using the Biological Attribute Network for signal refinement of signals for DDIs. For a DDE triplet, $D_1-D_2-\mathrm{AE}$:

• (a)

  if the network contained all three nodes ($D_1,D_2,\mathrm{AE}$) and links connecting each node with any number of targets, we calculated the total number of nodes that were included in the union of the individual shortest paths between each drug and the AE under consideration (hereafter called shortest path measure);

• (b)

  if the network contained the drug nodes ($D_1,D_2$) and links connecting each drug node with any number of enzymes or transporters, we calculated the number of common enzymes and transporters between them (hereafter called enzyme measure and transporter measure, respectively, and collectively called PK measures).

We then estimated via ROC curve analysis the combined performance of each of the three types of measures with the log-likelihood ratio metric using logistic regression.

Novel signal evaluation – Case studies

We selected two AEs of interest (QT interval prolongation, rhabdomyolysis) to run a FAERS-wide screening of signals. For each AE, we extracted all drug pairs found in at least 5 FAERS reports and generated suitable contingency tables. We calculated the log-likelihood ratios, shortest path measures, and both PK measures to rank the DDE triplets. We removed DDE triplets that were present in any of three established DDI online resources (i.e., the British National Formulary,²⁶ IBM Micromedex,²⁷ and the French Medicines Agency Thesaurus for DDIs²⁸). We also did not consider drug pairs that were under the same Anatomic Therapeutic Class (ATC) 4th level category, as those belong to the same chemical/pharmacological class and are not taken concomitantly. We then extracted the top 20 associations: (a) when only applying the SDA (i.e., log-likelihood ratio); and (b) when applying a binary logistic regression model that considered the SDA and either of the signal refinement framework measures (i.e., shortest path measure for QT interval prolongation and PK measures for rhabdomyolysis) to estimate predicted probabilities by utilizing the logistic regression model coefficients derived from the reference set (see previous section). We compared the rankings of the top associations from either approach to assess the impact on the relative change of ranking of the drug pairs.

For rhabdomyolysis, we calculated the log-likelihood ratio scores related to drug pairs containing a statin (ATC code: C10AA), which is a drug class known to cause an increased risk of rhabdomyolysis due to interaction with other drugs (i.e., fibrates, macrolides, and fusidic acid) and compared them with the scores of ezetimibe (i.e., another lipid-lowering agent) with the same drugs.

RESULTS

Evaluation of the novel statistical signal detection algorithm for adverse DDIs

Before integrating the systems pharmacology framework for signal refinement, we assessed the novel statistical SDA performance in FAERS using 4,455 positive and 4,544 negative DDI controls from CRESCENDDI that involved 442 drug ingredients and 168 AEs as MedDRA PTs in total. The log-likelihood ratio, detecting a changing probability indicative of a DDI, performed slightly better compared with log posterior odds ratio, which monitored increases in that probability (AUC: 0.574 and 0.548, respectively). Thus, the log-likelihood ratio was then used for comparison with other SDAs. By taking the subset of controls from the evaluation set that were found in at least 5 FAERS reports (3,507 in total; 2,213 positive and 1,294 negative), the AUC score of the log-likelihood ratio increased to 0.614 (Figure 2a).

Figure 2

ROC curve analysis for: (a) log-likelihood ratio, using the whole evaluation set (red line) or the subset of controls that were found in at least 5 FAERS reports (blue line); (b) comparative assessment of SDAs for DDI surveillance: Omega (red line); Interaction Signal Score (green line); delta_add (blue line); and log-likelihood ratio (purple line).

The selection of beta prior's hyperparameters impacts the results produced by the novel SDA. We investigated the effect of the chosen prior hyperparameter set ($a_0$ and ${\mathrm{\beta }}_0$) on the SDA performance assessment using ROC curve analysis. The AUC scores varied between 0.502 and 0.599 (report range of values), with the lowest performance corresponding to an uninformative Beta prior for the null hypothesis ($a_0={\mathrm{\beta }}_0=1$) and the highest one to a Beta prior with hyperparameters estimated from the FAERS AE rate distribution (see Prior parameter estimation in Supplementary Information).

In terms of comparison with other SDAs for DDI surveillance, the log-likelihood ratio (AUC: 0.573) outperformed Omega, which showed the highest performance (AUC: 0.569) among the existing SDAs, followed by IntSS (AUC: 0.489) and delta_add (AUC: 0.417). Also, the log-likelihood ratio achieved better specificity for high-sensitivity values compared with Omega (Figure 2b).

From the selected AEs that were individually considered for ROC curve analysis, the novel approach produced the following AUC scores: hypoglycemia (AUC: 0.742), rhabdomyolysis (AUC: 0.674), bradycardia (AUC: 0.652), myopathy (AUC: 0.628) (Figure S2). The SDA performance was lower for hemorrhage (AUC: 0.528), GI hemorrhage (AUC: 0.542), hypertension (AUC: 0.546), QT prolongation (AUC: 0.568), and Torsade de pointes (AUC: 0.515).

Three stratified subsets of the reference set in terms of individual drug behavior were used: Both, One, and None. The “Both” subset contained 4,212 controls, the “One” subset contained 2,312 controls, and the “None” subset contained 1,365 controls. The overall log-likelihood ratio was used to calculate the AUC for the ROC curve for each subset. The AUC for the “Both” subset was 0.645, the AUC for the “One” subset was 0.578, and the AUC for “None” was 0.533 (Figure S3a). The relative performance among the three different subsets was similar in the case of log posterior odds ratio (Figure S3b).

Improving signal detection of DDIs using systems pharmacology aspects

The number of nodes and links from the individual data resources that were added to the Biological Attribute Network are provided in Table S3. In total, the network contained 1,311 drugs, 351 AEs, 16,814 targets, 325 enzymes and 204 transporters. There were 3,663 drug-target and 1,060 target-AE links from Open Targets, while the number of drug enzyme and drug transporter links was 4,966 and 2,108, respectively. The number of high-confidence STRING target-target associations was 505,968.

The application of the systems pharmacology framework for signal refinement produced ROC curves with higher AUC values when the log-likelihood ratio was combined with any of the three measures (shortest path, enzyme, and transporter measures) (Figure 3). The shortest path measure combined with the log-likelihood ratio produced an increase in the AUC score from 0.620 to 0.722 using the relevant controls with nodes that were present in the Biological Attribute Network. Similarly, for both PK measures, the AUC scores of the combined models (i.e., log-likelihood ratio + enzymes; log-likelihood ratio + transporters) were higher than the ones considering only the log-likelihood ratio (from 0.580 to 0.673 for consideration of enzymes; from 0.568 to 0.653 for consideration of transporters).

Figure 3

Receiver Operating Characteristic (ROC) analysis that considers the use of the log-likelihood ratio combined with: (a) the shortest path measure; (b) the enzyme measure; and (c) the transporter measure.

Exploring signal evaluation for selected AEs

For QT interval prolongation, the top 20 associations using only the log-likelihood ratio were consistently ranked lower using the combined approach with the shortest path measure (Figure 4). Only one positive control from CRESCENDDI was present in the top 20 associations using the log-likelihood ratio for ranking, while 6 drug pairs were found in at least one of the clinical resources that were considered in this study (i.e., BNF, Micromedex, or ANSM Thesaurus). On the contrary, the top 40 associations using the combined approach contained 5 positive controls from CRESCENDDI, while only 6 drug pairs out of 40 did not belong to any of the clinical resources (Figure 5). For the top 31 drug pairs when applying the combined approach, the differences in rankings ranged from 981 to 7,809. For the remainder of the associations (32–40), small ranking differences were observed. The combination of amlodipine with dofetilide has not been reported in any clinical resource but was ranked 14th when considering the shortest path measure (and moved 4,279 positions up). Similarly, the combination of clonazepam with acamprosate also moved 70 positions (36th) with the application of the combined approach.

Figure 4

Top 20 associations using the log-likelihood ratio approach and their respective rankings with the shortest path + log-likelihood ratio approach. Drug pairs in green represent the ones that were positive controls for QT interval prolongation in the CRESCENDDI reference set. Drug pairs in orange represent the ones that were mentioned as known to interact in any of the clinical resources that were considered in this study (i.e., BNF, Micromedex, or ANSM Thesaurus), although those drug pairs could be associated with other medical events apart from QT interval prolongation.

Figure 5

Top 40 associations using the shortest path + log-likelihood ratio approach and their respective rankings with the log-likelihood ratio approach. The difference between the two rankings for each drug pair is denoted in the label on the right side of the plot. Drug pairs in green represent the ones that were positive controls for QT interval prolongation in the CRESCENDDI reference set. Drug pairs in orange represent the ones that were mentioned as known to interact in any of the clinical resources that were considered in this study (i.e., BNF, Micromedex, or ANSM Thesaurus), although those drug pairs could be associated with other medical events apart from QT interval prolongation.

For rhabdomyolysis, we compared the rankings using the log-likelihood ratio scores of statins, which is a drug class known to cause an increased risk for rhabdomyolysis due to interaction with other drugs (i.e., fibrates, macrolides, and fusidic acid) as opposed to rankings of ezetimibe, which is another lipid-lowering agent, with the same drugs. In the ranking of the 16,799 filtered drug pairs (i.e., with at least 5 FAERS reports) that were screened for rhabdomyolysis, drug pairs that contain statins consistently populated the top places in the ranking table (e.g., simvastatin with gemfibrozil or clarithromycin, atorvastatin with fusidic acid), while drug pairs that contain the comparator drug (ezetimibe) were lower in rankings (below 1,000) (Table 2). In all four cases of interacting drugs, simvastatin and atorvastatin were placed above ezetimibe, with a very high log-likelihood ratio (values between 3.69 and 241.2).

Table 2. Drug pair log-likelihood ratio scores and rankings from screening rhabdomyolysis FAERS cases that contain: (i) a fibrate, macrolide, or fusidic acid (i.e., potentially interacting drug) (Drug 1) and (ii) a statin or ezetimibe (as a comparator drug) (Drug 2). Rankings were calculated out of the 16,799 eligible drug pairs (i.e., at least 5 FAERS reports) that were screened for rhabdomyolysis

Drug 1	Drug 2	Log-likelihood ratio	Drug pair ranking
GEMFIBROZIL	SIMVASTATIN	241.2192869	2
	CERIVASTATIN	15.45426136	1,315
	PRAVASTATIN	9.746103259	2,301
	ATORVASTATIN	3.687716196	4,656
	ROSUVASTATIN	−1.424158741	13,149
	LOVASTATIN	−1.444116569	13,200
	EZETIMIBE	−1.902040641	14,046
CLARITHROMYCIN	SIMVASTATIN	148.8437063	12
	ATORVASTATIN	11.26737274	1967
	EZETIMIBE	2.523279591	5,574
	PRAVASTATIN	−0.449481203	9,767
	ROSUVASTATIN	−0.767005855	10,653
FUSIDIC ACID	ATORVASTATIN	105.3754776	22
	SIMVASTATIN	88.46783535	34
	PRAVASTATIN	4.715557493	4,082
	EZETIMIBE	1.187202838	6,968
	ROSUVASTATIN	−0.301944238	9,388
FENOFIBRATE	SIMVASTATIN	37.277021	263
	ATORVASTATIN	26.35263041	518
	EZETIMIBE	17.91501155	1,054
	ROSUVASTATIN	−0.620035622	10,189
	PRAVASTATIN	−1.915656202	14,065

DISCUSSION

Signal detection in pharmacovigilance encounters multiple challenges due to the nature of the data and reporting, leading to methodologies either producing falsely generated signals or being incapable of spotting relevant ones. This issue is also augmented by the fact that SDAs mainly rely on disproportionality analysis and biological mechanisms, which can explain whether a signal is plausible from a pharmacological perspective, are considered later in the signal evaluation process. As this is particularly relevant in the case of DDIs, the main goal of this study was to assess a novel SDA for identifying novel DDIs using a Bayesian hypothesis testing framework and adding a signal refinement step utilizing systems pharmacology data. First, we performed a quantitative comparison of existing SDAs along with the novel one using a large and diversified publicly available reference set. The novel method outperformed all three existing ones in terms of AUC scores. We also noticed adequate or above-average algorithm performance for specific AEs of interest, especially for DDI surveillance, such as QT interval prolongation, rhabdomyolysis, bradycardia, and hypoglycemia. The novel SDA showed enhanced performance when combined with any of the three measures derived from the Biological Attribute Network (i.e., shortest path, enzyme, and transporter). Also, two case studies demonstrated the applicability of the novel approach for real-life signal detection purposes: the first one was related to signal prioritization for QT interval prolongation; the second one showed the relative magnitude of rhabdomyolysis signals of the novel SDA associated with statins and other lipid-lowering agents.

Systems pharmacology can support signal detection in pharmacovigilance to identify more signals with biological plausibility.^14-16 While this is important for single drugs, it is also particularly relevant for the detection of novel DDIs, considering the even larger number of potential drug combinations that can arise, many of which can be flagged as potential signals. The increasing availability of data makes it possible to automate the production of a biological attribute network. Such a network can capitalize on knowledge of the mode of action of drugs, their metabolic and elimination pathways, and the human protein interactome. The network then captures information pertinent to drug–drug interactions as the proximity of drugs and nodes associated with adverse events. This makes it possible for algorithms that automatically analyze the network to use this knowledge to inform statistical tests pertinent to novel drug–drug complications.

The strength of this study includes the use of a comprehensive and clinically relevant reference set.¹⁹ By having access to a large set of controls that also considers multiple AEs, a quantitative comparison of existing SDAs with the novel approach was possible. The novel SDA provides outputs that could be utilized in combination or separately to monitor the different probabilities that could provide a pharmacology-driven framework. Also, the signal detection framework could be extended to consider higher-order drug interactions. The use of open data (SRS database, reference set, systems pharmacology data) is another strength of this study. This study introduced the concept of incorporating biological plausibility aspects as a signal refinement step, which has been explored in other studies^{14, 15} but not in the scope of DDIs.

The SDA yielded reasonable results in terms of ranking drug pairs for signals of rhabdomyolysis based on existing pharmacological knowledge. These findings suggest that the novel SDA could be useful in screening SRS data in real-world applications. In terms of the signals of ezetimibe (that was used as a comparator drug), those were always below the respective signals from atorvastatin and simvastatin. Simvastatin and atorvastatin are predominantly metabolized by CYP3A4 and their levels increase significantly when co-administered with strong CYP3A4 inhibitors, such as clarithromycin.²⁹ Moreover, both statins are substrates of the organic anion-transporting polypeptide (OATP)1B1, which is responsible for their hepatic uptake and is also inhibited by clarithromycin. Therefore, the presence of the macrolide substantially affects the concentration of both statins.³⁰ Another example is the signal of cerivastatin with gemfibrozil, which ranked high in the analysis. Cerivastatin was withdrawn from the market worldwide in 2001 due to its association with rhabdomyolysis, with a higher risk observed when taken concurrently with gemfibrozil.³¹ Furthermore, although most statins generated relatively strong signals with gemfibrozil, the same was not observed with fenofibrate. In fact, according to clinical guidelines and literature, of the two fibrates, gemfibrozil has a higher risk of interacting with statins and leading to rhabdomyolysis.^{32, 33} A surprising finding was the rhabdomyolysis signal of fenofibrate with ezetimibe, which was in a higher ranking (1054th) in comparison to signals of other statins. Previous studies have examined the safety of this combination and have reported no clinically important elevations in creatine phosphokinase (CPK) (which are indicative of rhabdomyolysis) or additional risk of myopathy due to the combination therapy.^{34, 35}

This study also identified two potential signals of novel DDIs linked to QT interval prolongation (amlodipine – dofetilide and clonazepam – acamprosate), which are currently unknown but are supported by both statistical screening and biological information. More precisely, both clonazepam and acamprosate are positive modulators of the anion channel of the GABA-A receptor GABRG3.^{36, 37} In the Biological Attribute Network, GABRG3 is linked to the potassium voltage-gated channel KCNH2 that is associated with QT interval prolongation (target-AE association) via two nodes (GABARAP and AMK2). In terms of the combination of amlodipine and dofetilide, the associated nodes in the network are interconnected. Amlodipine blocks the voltage-gated L-type calcium channel (CACNA1C), while dofetilide blocks the voltage-activated potassium channel (KCNH2).^{38, 39} These two targets are directly linked and KCNH2 is also associated with QT interval prolongation.

A systematic evaluation of different SDAs for DDI surveillance was missing from the literature. A previous study used Stockley's as a source of positive controls, but the resulting reference set was not made available, hindering the reproducibility and extension of the study and the possibility to further dive into the nature of the controls.⁴⁰ Other efforts have used benchmarking only for a very limited number of AEs of interest to measure and compare SDA performance.^{6, 7, 41} In our study, the consideration of a large and diversified reference set enabled us to compare the performance of the novel SDA across multiple AEs. We noticed substantial differences in method performance depending on the AE. As an illustrative example, for common AEs, such as hemorrhage, the masking effect might have been responsible for lower performance.

Systems pharmacology has been incorporated into drug development. Multiple machine learning and artificial intelligence (AI) methodologies have also examined DDI prediction by integrating various data types and information sources as features, such as drug-target profiles (i.e., drug-protein interactions), metabolizing enzymes, and transporters.^{18, 42-44} However, systems pharmacology coupled with pharmacovigilance has only been recently considered in studies^{14, 15} and has not been explored in the case of adverse DDIs. For single drugs, we have seen some recent efforts to develop similar frameworks that, apart from main pharmacological targets, also consider off-targets to aid the detection of drug-related side effects.^{15, 45} The use of off-target data might be particularly relevant in the context of drug safety, as many drug complications leading to adverse drug reactions are related to secondary pathways and off-target activity of the drug molecule.

Limitations of the study

The focus of this study was on two-way DDIs, although high-order DDIs (i.e., involving more than two drugs) would be an area of focus for future studies. Some previous work has already attempted to explore this area.^{46, 47}

The masking effect that can be present in SRS data was not considered in this study in the data mining step. Revising the concept of masking in the case of contingency tables for two-way DDIs, which has been extensively described and explored in previous studies for single drugs,^48-51 would also be interesting. The modifications of the data mining step to minimize the bias resulting from masking in SRS data could potentially increase the performance measures of the different SDAs for DDI surveillance. Additionally, it could have an impact on the signal prioritization step, altering the drug pair rankings.

The existing evidence that currently appears in Open Targets regarding target safety liabilities is limited to a small number of targets and only validated associations; thus, it might correspond to well-known safety complications arising from drug combinations that appeared in the reference set that was utilized for performance evaluation.

This study considered the combination of each of the three different systems pharmacology measures with the SDA scores using binary logistic regression. However, combining multiple defined measures at the same time as well as non-linear approaches could be relevant. Also, apart from the shortest path, other centrality measures in network analysis (e.g., closeness centrality) offer potential for future research.

CONCLUSION

This study provides a novel framework for detecting DDI signals using disproportionate reporting in FAERS combined with a biological information network. With an increasing volume of systems pharmacology information now available, we show that this information has the potential to enhance signal detection in pharmacovigilance, with DDIs being an important and promising area of application. This study also identified two potential DDI signals related to QT interval prolongation. Further studies, including the consideration of additional SRS databases, real-world evidence, in vitro or in vivo experiments, are needed to validate the potential signals.

FUNDING

This study was jointly funded by EPSRC (grant number EP/R51231X/1) and AstraZeneca.

CONFLICT OF INTEREST

Elpida Kontsioti has received PhD studentship that was jointly funded by AstraZeneca and the EPSRC (grant number EP/R51231X/1). Isobel Anderson is an employee and shareholder of AstraZeneca. Munir Pirmohamed receives research funding from various organizations including the MRC and NIHR. He has also received partnership funding for the MRC Clinical Pharmacology Training Scheme (co-funded by MRC and Roche, UCB, Eli Lilly, and Novartis). He has developed an HLA genotyping panel with MC Diagnostics, but does not benefit financially from this. He is part of the IMI Consortium ARDAT (www.ardat.org). These funding sources were not utilized for this work. All other authors declared no competing interests for this work.

AUTHOR CONTRIBUTIONS

E.K., S.M., I.A., and M.P. wrote the manuscript. E.K., S.M., and M.P. designed the research; E.K. performed the research and analyzed the data.

Open Research

DATA AVAILABILITY STATEMENT

The code is available at: https://github.com/elpidakon/BANet/. This article was originally posted to the Research Square preprint server. (https://doi.org/10.21203/rs.3.rs-3478903/v1).