A reference set of clinically relevant adverse drug-drug interactions

The accurate and timely detection of adverse drug-drug interactions (DDIs) during the postmarketing phase is an important yet complex task with potentially major clinical implications. The development of data mining methodologies that scan healthcare databases for drug safety signals requires appropriate reference sets for performance evaluation. Methodologies for establishing DDI reference sets are limited in the literature, while there is no publicly available resource simultaneously focusing on clinical relevance of DDIs and individual behaviour of interacting drugs. By automatically extracting and aggregating information from multiple clinical resources, we provide a scalable approach for generating a reference set for DDIs that could support research in postmarketing safety surveillance. CRESCENDDI contains 10,286 positive and 4,544 negative controls, covering 454 drugs and 179 adverse events mapped to RxNorm and MedDRA concepts, respectively. It also includes single drug information for the included drugs (i.e., adverse drug reactions, indications, and negative drug-event associations). We demonstrate usability of the resource by scanning a spontaneous reporting system database for signals of DDIs using traditional signal detection algorithms.

www.nature.com/scientificdata www.nature.com/scientificdata/ The early and effective identification of drug safety signals is of paramount importance for the pharmaceutical industry and regulatory authorities. For adverse reactions caused by DDIs, there is an increasing need for improved SDAs, with the existing state-of-the-art being less mature compared to the well-established algorithms used for detecting signals of ADRs caused by a single drug [8][9][10] . Even after drug approval, detection of novel DDIs might be delayed and difficult, due to the inherent complexity of DDIs, dose-dependency (i.e., some interactions only become evident in elevated drug levels) and natural human inter-variability (as well as intra-variability, in some cases) that accounts for the onset of some DDIs (e.g., rapid and poor metabolizers, genetic subpopulations, etc.). The growing volume of real-world health data presents both a challenge and an opportunity for the pharmacovigilance community, making manual review impossible and requiring a higher level of automation in the methods that are routinely used for scanning such databases.
As the available evidence is not static, the lack of gold standards (i.e., definitive positive and negative controls) poses a challenge when it comes to defining appropriate reference sets in pharmacovigilance for performance evaluation. Also, much disagreement exists in terms of the choices and criteria under consideration (e.g., well-established versus emerging cases) [11][12][13] .
For single drugs, a number of reference sets exist that include drug-event pairs that are either well-established (e.g., OMOP reference set 14 , Harpaz 15 ), belong to recent product labelling changes 16 , or can be found in product labels (e.g., EU-ADR initiative 17 ). More recently, efforts to automate the generation of a reference set for single-drug ADRs by combining multiple sources of evidence identified a number of limitations, including: size; consideration of a single data source for extracting positive controls; availability (i.e., not being open access); and inclusion of only a limited number of drugs and adverse events (AEs) 18 . Our understanding is that a fair algorithm comparison requires testing on a large reference set to derive performance metrics that are likely to indicate performance in the context of novel signals.
To the best of our knowledge, there is no established reference set for DDIs coupled with information on the individual behaviour of interacting drugs. Were such a reference set to exist, it could enable the classification of positive controls based on the possible underlying mechanism causing the interaction. Initial efforts for detecting signals indicative of DDIs included test cases that were limited in either size 9 or variety of drugs and AEs 10 . A more advanced approach was implemented by Juhlin et al. 19 , although their reference set relied on a single clinical resource, which might not be a good idea since discordance among DDI compendia has been identified in the literature [20][21][22][23][24] .
Although a definitive reference standard including the complete set of DDIs cannot exist, the automatic extraction and aggregation of information from multiple clinical resources on DDIs and the individual behaviour of interacting drugs, along with scanning the scientific literature for negative controls, enabled us to construct, share, and advocate CRESCENDDI (Clinically-relevant REference Set CENtered around Drug-Drug Interactions), a dataset that can be used to facilitate research in SDAs and allow common ground for comparing methodologies. We propose a scalable approach for generating a normalized reference set that requires less manual effort for future updates, considering the dynamic nature of data and evidence availability, and for subset selection using design criteria. Figure 1 outlines the main steps for generating the reference set.

Methods
The processing pipeline for the construction of CRESCENDDI included the following steps: (1) web data extraction of information related to DDIs and single-drug ADRs from 4 different online resources for DDIs, single-drug ADRs, and drug indications; (2) mapping (normalization) of drug names appearing in the extracted data; (3) extraction of the intersection of the DDI online resources; (4) manual annotation and mapping of English language text descriptions for DDIs, single-drug ADRs, and drug indications to MedDRA concepts; (5) generation of positive controls for DDIs using the normalized intersection of the DDI online resources; (6) generation of negative controls using drugs and AEs from the positive control set combined with PubMed search; and (7) aggregation of information on single-drug ADRs and drug indications for the drugs appearing in the DDI reference set (i.e., positive and negative control sets) and generation of negative controls for single drugs with a process similar to the one followed in the previous step for DDI negative controls.
Web data extraction. DDI data were derived from the following online resources: the British National Formulary (BNF) website 25 , the French National Drug Safety Institute (ANSM) Portable Document Format (PDF) file (Thesaurus) 26 and the Micromedex platform 27 . For BNF and Micromedex, web data extraction tools in Python 3.6 28 enabled the extraction of the relevant fields into a Comma Separated Values (CSV) file (June-August 2018). For Thesaurus, the R package IMThesaurusANSM 29 was used and the resulting R dataframe from the 2019 update was converted to a CSV file.
For single-drug ADRs, the following sources were considered: the BNF website 25 and SIDER dataset 30 . For drug indications, SIDER was used. BNF ADR data were extracted in a similar way as previously with DDI data (automated web data extraction) into a CSV file, while SIDER data for ADRs and drug indications were already available in CSV files.
Drug name mapping. To facilitate usability and ensure compatibility, a standardization process was followed such that we could provide a resource with normalized concepts to standard terminologies for drugs and medical events. Specifically, the Observational Health Data Sciences and Informatics (OHDSI) Vocabulary version 5 was selected for mapping the drug names occurring in each of the DDI online resources into RxNorm and RxNorm Extension standard codes (at the Ingredient level) using OHDSI Usagi 31 .
We removed combination drugs (as DDIs of their constituent drug ingredients were separately mentioned), vaccines, vitamins, herbal medicines, food, beverages, supplements, tobacco, and lab tests. Also, generic drug b Extracted DDI data c Extracted single-drug data e DDI pair intersection www.nature.com/scientificdata www.nature.com/scientificdata/ classes (e.g., combined hormonal contraceptives, hormonal replacement therapy) appearing in the BNF were not mapped to their individual drug ingredients, as there was no table on the BNF website specifying the drugs belonging in each drug class. We mapped the remaining unique drug names occurring in the DDI resources to OHDSI standard vocabulary concept identifiers. For example, Metoprolol Tartrate was mapped to the RxNorm Ingredient concept metoprolol. For Thesaurus, a native French speaker (pharmacist) confirmed the drug mappings of French drug names to English language OHDSI concepts.
A similar process was followed for drugs that appear in the single-drug data (ADRs and indications).
Intersection of DDI online resources. By matching drug names to their mapped drug ingredients in the extracted DDI data tables, we obtained the set of common drug pairs across the tables and generated a new table that contains only the DDIs and associated information which could be found in each of the DDI online resources under consideration. Cases where the interacting drug mapping of D1 and D2 were swapped in the original data tables (i.e., (D1,D2) and (D2,D1)) were considered equivalent. The final table contained the following fields: • drug_1 concept name (e.g., metoprolol); • drug_2 concept name (e.g., lidocaine); • bnf description (e.g., Lidocaine is predicted to increase the risk of cardiovascular adverse effects when given with metoprolol. Manufacturer advises use with caution or avoid); • micromedex description (e.g., lidocaine toxicity (anxiety, myocardial depression, cardiac arrest)); • bnf severity (e.g., Severe); • ansm severity (e.g., Précautions d' emploi (Precautions for use)); • micromedex severity (e.g., major); • bnf evidence (e.g., Study); • micromedex evidence (e.g., probable).

adverse event and indication mappings. For DDI-related text descriptions in English from BNF and
Micromedex that could be found in the DDI intersection table, a drug name blinding process was performed by replacing the interacting drug names with a common token in all cases (i.e., 'X'). In this way, the number of unique descriptions was reduced, thus facilitating the mapping process that followed. For example, the descriptions: Both dexibuprofen and ibuprofen can increase the risk of nephrotoxicity. and Both polymyxins and streptomycin can increase the risk of nephrotoxicity. were both mapped to the following blinded description: Both X and X can increase the risk of nephrotoxicity.
The set of blinded text descriptions for BNF and Micromedex was extracted from the table and a semi-automated mapping process using OHDSI Usagi mapped them to MedDRA PT concepts. We explicitly focused on text descriptions that included clinical manifestations of DDIs, e.g., X may increase the risk of hypoglycaemia when taken with X. Text descriptions containing a potential mechanism of the interaction were left unmapped. In some cases, a single text description was linked to multiple concepts. For example: Interaction Effect: An increased risk of cardiotoxicity (QT prolongation, torsades de pointes, cardiac arrest).
Also, serotonin syndrome was not mapped to its corresponding MedDRA PT and was not further considered as an AE for inclusion in the reference set.
Text descriptions from the BNF regarding single-drug ADRs were mapped to MedDRA PTs (where possible), but only for the drug ingredients that could be found in the DDI pair intersection table. SIDER ADR and indication data for the same list of drugs were also mapped to OHDSI concepts; however, for this resource, MedDRA PT codes were already available.
Positive controls. The set of positive controls was derived from the DDI intersection table, using mappings of text descriptions to AEs that were generated in the previous step. It contained 10,286 drug-drug-event (DDE) triplets, 454 unique individual drug ingredients, and 179 unique AEs (as OHDSI concepts) in total.
Negative controls. The set of negative controls was generated by randomly pairing two drug ingredients from the 454 unique drug ingredients that can be found in the positive controls and, in case the random drug pair did not appear in any of the DDI online resources, then it was randomly paired with an AE from the 179 unique AEs present in the positive control set. The choice of generating negative controls with common drug ingredients and AEs as the ones appearing in positive controls aimed to ensure the generation of a balanced reference set that does not contain added biases by design. For each of the created DDE triplets, a customized query ("(DRUG_1_CONCEPT_NAME) AND (DRUG_2_CONCEPT_NAME) AND ((EVENT_CONCEPT_NAME) OR (interaction))", e.g., "(Oxazepam) AND (Naproxen) AND ((Hyperkalaemia) OR (interaction))") was submitted to PubMed in an automated fashion and, if the search returned no results, the triplet was added to the negative control set. This process aimed to provide more confidence, to the best of our ability, about the absence www.nature.com/scientificdata www.nature.com/scientificdata/ of literature evidence of a potential DDI for the triplet under consideration, rather than definitive evidence to support the lack of a potential association.
The Single-drug aDRs, indications, and negative controls. By replacing text descriptions from BNF (N = 1,538) and MedDRA PT codes to their corresponding mapped OHDSI concepts, a table with ADR and indication information related to the drug ingredients of the DDI reference set was generated. The table included: 438 unique drug ingredients, which could be found in at least one of the resources under consideration (i.e., BNF and SIDER), 3,492 AEs and 1,557 indication terms (as OHDSI concepts).
BNF and SIDER jointly contained 69,721 single-drug ADRs, with 12,318 common instances; this set could be utilized as a source for single-drug positive controls. This set covered 381 unique drugs and 835 unique AE concepts. Random pairing of those drugs and AE concepts followed by submission of a customized query ("(DRUG_CONCEPT_NAME) AND (EVENT_CONCEPT_NAME) AND ((adverse event) OR (adverse drug reaction))'", e.g., "(oxazepam) AND (myoclonus) AND ((adverse event) OR (adverse drug reaction))") to PubMed (to ensure absence of literature evidence of a potential ADR for the various drug-event associations) enabled the generation of a negative control set for single drugs (N = 12,141).

Data Records
CRESCENDDI is publicly available online through Figshare 32 in 5 Excel spreadsheets. The fields contained in each of the spreadsheets are outlined below. DRUG_CONCEPT_NAME: Name of the drug (active ingredient) that comprises a single drug information case (drug-event pair). EVENT_CONCEPT_NAME: Name of the normalized event (as a MedDRA PT) that comprises a single drug information case (drug-event pair).

Data
EVENT_TYPE: Type of association between the drug and event of the single drug information case (i.e., ' Adverse event' or'Indication').
SOURCE: Label indicating the resource of the single drug information case (i.e., 'BNF' or 'SIDER'). Tab 2 -Negative drug-event associations. DRUG_CONCEPT_NAME: Name of the drug (active ingredient) that comprises a single drug negative case (drug-event pair). EVENT_CONCEPT_NAME: Name of the normalized event (as a MedDRA PT) that comprises a single drug negative case (drug-event pair).

technical Validation
Given the inability to generate 'gold standards' in pharmacovigilance, the validation of the reference set included steps that supported the technical quality of the procedures followed to generate the dataset, rather than attempting to further ensure the validity of each control. The process of validating the reference set was two-way.
First, we verified the original online data as well as the unmapped extracted DDI data versus the curated reference set in order to validate the accuracy of the automated extraction and concept mapping processes. A random sample of 40 positive and 40 negative controls was manually checked in each of the DDI online resources (i.e., BNF, Thesaurus, and Micromedex), to ensure the presence or absence, respectively, of the DDE triplet in the information provided. Due to the time lag between data retrieval (June-August 2018) and validation (September 2021), there were issues most probably related to differing versions of the resources. More specifically, there were 6 positive controls missing from one of the resources (5 from the BNF and 1 from Micromedex), 2 of which due to the complete removal of the drug monograph. Also, one of the negative control samples had been added to the BNF. No issues were identified during the validation of the random sample of controls from the curated reference set against the unmapped extracted DDI data.
Second, the validation of text description mapping for AEs included independent annotation of a random sample of 100 text descriptions for AEs to ensure agreement on event mapping. No issues were identified.
Expert manual review of a number of cases from the reference set could be potentially performed as an additional layer of the technical validation procedure to confirm that the included DDIs are worthy of note. We have attempted to replace this need for expert manual review by combining information from multiple resources of clinical interest.  Table 2. Statistics related to the performance evaluation of three SDAs for single drugs using FAERS data. AUC, area under the receiver operating characteristic curve; Score+, median signal score for the set of positive controls; Score−, median signal score for the set of negative controls.
www.nature.com/scientificdata www.nature.com/scientificdata/ The next section outlines the application of the reference set for drug safety surveillance in an SRS database using SDAs that have been described in previous studies, in an effort to showcase its validity through practical implementation as well as prove its potential use in the real world.

Usage Notes
This publicly available resource aims to provide a common ground for evaluation of SDAs related to DDI signals, with a focus on assembling information from disparate sources that could provide support toward the clinical relevance of controls. The size of the reference set alongside the relatively large number of drugs and AEs considered enables a quantitative approach in algorithm performance evaluation in terms of SDAs developed for DDI signals.
Also, the supplementary single-drug positive and negative control sets could be used separately for signal detection in the context of adverse drug-event associations.
This reference set could be applied to a variety of data sources (e.g., electronic health record data, social media, literature). Here, we illustrate applicability in one such context.

FaERS screening for DDI signals.
In pharmacovigilance, signal detection is largely based on methodologies that use disproportionality analysis, meaning that the observed counts are compared to the expected ones, assuming that the drug(s) and the events occur independently. For DDIs, similar measures of disproportionality have been described in the literature. For our analysis, the following statistical measures were considered: (i) An observed-to-expected shrunk interaction measure (Omega) 8 (ii) The 'interaction coefficient' in a linear regression model with additive baseline (delta_add) 9 (iii) A measure based on an adapted version of Multi-Gamma Poisson Shrinker (MGPS) model, called Interaction Signal Score (IntSS) 10 A curated and standardized version of FAERS was used as the test data source. In FAERS, each spontaneous report contains information on administered drugs, AEs experienced, indications, and demographic information (i.e., sex, age, etc). Using a slight modification of the Adverse Event Open Learning through Universal Standardization (AEOLUS) process that was described in published work 33 , FAERS data files corresponding to the period 2004Q1-2018Q4 were standardized to the RxNorm Ingredient level terms and MedDRA PTs, so that compatibility of the reference set with the test data could be established. The total number of reports containing at least one drug and one AE was 9,203,239.
ROC analysis was performed using MATLAB function perfcurve. 95% confidence interval (CI) estimates for the AUC were calculated using leave-one-out cross validation. Table 1 shows statistics related to the performance of the SDAs in FAERS using our advocated reference standard. For each algorithm, the table provides the AUC scores with 95% CI estimates. As opposed to expectations, only two out of three SDAs for DDIs (i.e., Omega and IntSS) performed better than random guessing. Median signal scores for the set of positive controls were larger compared to those of negative controls in those two cases, which provides support for the validity of CRESCENDDI. Limited performance analyses in the studies describing the SDAs for DDIs as well as absence of other efforts in the literature for quantitative performance assessment across multiple thresholds did not manage to provide an appropriate benchmark for this study.
FaERS screening for single-drug aDRs. For the single-drug reference set, empirical support was provided by applying three well-established SDAs to FAERS data. Positive controls (at the MedDRA PT level) were sourced from either the intersection (N = 12,318) or the union (N = 52,637) of the resources under consideration (i.e., BNF and SIDER) and ROC analysis for performance evaluation similarly to the previous section (Table 2). Table 2 shows that EBGM outperformed PRR, which is in line with the results of other studies 14 . Also, improved performance was noticed in the case of resource intersection for all SDAs, which advocates the use of information that appears in multiple resources rather than relying on a single data source for deriving positive controls.
Flexibility of the resource. This reference set has been mapped to standard vocabularies using the OHDSI framework, thus enabling linking to other ontologies and/or vocabularies, based on research needs, test data resources format, etc.
Considering the changing nature of available evidence for both positive and negative controls, reproduction of the resource in the future would be recommended to ensure that it is up to date. Although specific products (e.g., vaccines) were excluded given our focus on generating a reference set relevant to DDIs, slight modifications on the drug mapping process can allow the inclusion of controls pertinent to any additional products of interest.
The incorporation of additional information (e.g., shared indication, single-drug AE under the same HLT/ HLGT MedDRA level, evidence levels) enables filtering and stratification of controls. The effect of stratification using those design criteria is outside the scope of this paper, but we aim to explore this aspect in future research.