Discovering Severe Adverse Reactions From Pharmacokinetic Drug–Drug Interactions Through Literature Analysis and Electronic Health Record Verification

Abstract

While drug–drug interactions (DDIs) and their pharmacokinetic (PK) mechanisms are well‐studied prior to drug approval, severe adverse drug reactions (SADRs) caused by DDIs often remain underrecognized due to limitations in pre‐marketing clinical trials. To address this gap, our study utilized a literature database, applied natural language processing (NLP) techniques, and conducted multi‐source electronic health record (EHR) validation to uncover underrecognized DDI‐SADR signals that warrant further investigation. PubMed abstracts related to DDIs from January 1962 to December 2023 were retrieved. We utilized PubTator Central for Named Entity Recognition (NER) to identify drugs and SADRs and employed SciFive for Relation Extraction (RE) to extract DDI‐SADR signals. The extracted signals were cross‐referenced with the DrugBank database and validated using logistic regression, considering risk factors including patient demographics, drug usage, and comorbidities, based on EHRs from Vanderbilt University Medical Center (VUMC) and the All of Us research program. From 160,321 abstracts, we identified 111 DDI‐SADR signals. Seventeen were statistically significant (13 by one EHR and 4 by both EHR databases), with 9 being previously not recorded in the DrugBank. These included methadone‐ciprofloxacin‐respiratory depression, oxycodone‐fluvoxamine‐clonus, tramadol‐fluconazole‐hallucination, simvastatin‐fluconazole‐rhabdomyolysis, ibrutinib‐amiodarone‐atrial fibrillation, fentanyl‐diltiazem‐delirium, clarithromycin‐voriconazole‐acute kidney injury, colchicine‐cyclosporine‐rhabdomyolysis, and methadone‐voriconazole‐arrhythmia (odds ratios (ORs) ranged from 1.9 to 35.83, with P‐values ranging from < 0.001 to 0.017). Utilizing NLP to extract DDI‐SADRs from Biomedical Literature and validating these findings through multiple‐source EHRs represents a pioneering approach in pharmacovigilance. This method uncovers clinically relevant SADRs resulting from DDIs that were not evident in pre‐marketing trials or the existing DDI knowledge base.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

Pharmacokinetic (PK) drug–drug interactions (DDIs) are a leading cause of severe adverse drug reactions (SADRs). However, many SADRs associated with known PK DDIs are not identified during pre‐marketing clinical trials and require clinical evidence from post‐marketing studies. While biomedical literature is a crucial resource for post‐marketing DDI‐SADR evidence, mining clinical evidence from this vast repository and validating the mined evidence for accuracy remains a significant challenge.

• WHAT QUESTION DID THIS STUDY ADDRESS?

Can we extract unrecognized SADRs with known PK mechanisms from biomedical literature using natural language processing tools and validate DDI‐SADR signals through multi‐source Electronic Health Record (EHR) databases?

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

This study introduces a translational informatics pipeline that includes DDI‐SADR extraction from biomedical literature and real‐world data validation. The aim is to detect unrecognized SADRs associated with PK DDIs, providing an efficient way to uncover evidence for PK DDI‐SADRs in the post‐marketing phase, thereby improving drug safety.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

Our approach provides evidence for post‐marketing clinical trials to confirm SADRs of PK DDIs that were not identified in pre‐marketing trials. Additionally, our validation results from real‐world data might help healthcare providers avoid prescribing drugs with known PK DDIs that are labeled as having minimal risk but have demonstrated SADRs in clinical practice based on real‐world data.

Drug–drug interactions (DDIs) are a significant concern in clinical healthcare, particularly in scenarios involving multiple medications and in vulnerable groups such as elderly patients. ¹ From the standpoint of pharmacokinetics (PK), DDIs occur when the presence of one pharmaceutical agent alters the absorption, distribution, metabolism, and/or excretion of another. ² A common form of PK DDI is the inhibition of Cytochrome P450 (CYP) enzymes by a perpetrator drug (or inhibitor), resulting in elevated levels of an objective drug (or substrate) metabolized by these enzymes. ³ A recent analysis of the FDA Adverse Events Reporting System (FAERS) identified 590 adverse drug events associated with 2,085 DDI pairs involving known CYP‐related PK mechanisms. ⁴ Some PK DDIs have been linked to serious adverse drug reactions (SADRs), underscoring their clinical significance. ⁵ For instance, drugs such as seldane, hismanal, and propulsid were removed from the U.S. market because their interaction with other drugs caused metabolic inhibition, leading to severe arrhythmias. ⁶ To prevent such incidents, clinical trials are conducted during the drug development phase prior to market release to not only detect PK interactions between the investigational drug and other drugs but also to determine the clinical significance of the observed or expected DDIs. ⁷ However, the efficacy of these trials can be limited by factors such as short duration and the limited number of participants, often only a few hundred to a few thousand. ⁸ Notably, clinical trials frequently exclude significant segments of the population, such as the elderly, children, pregnant women, and patients suffering from multiple diseases, implying that the trial participants may not accurately represent the broader population in which the drug will eventually be used. ⁹

In response to the challenges posed by pre‐market pharmaceutical investigations, regulatory authorities in the field of drug administration have established and operationalized pharmacovigilance (PV) systems to monitor drug safety following their introduction into the market. PV systems use data mining techniques to determine the clinical consequences of post‐marketing DDIs using various databases. Traditionally, PV analyses have relied on data from spontaneous reporting systems (SRS) and fundamental statistical methods such as the proportional reporting rate, ¹⁰ reporting odds ratio, ¹¹ multi‐item Gamma Poisson shrinker, ¹² and Bayesian confidence propagation neural network. ¹³ These models, originally designed to detect ADRs induced by single drugs and DDIs, have been instrumental in monitoring medication safety. Another valuable resource is the biomedical literature, a comprehensive repository of medical information ranging from preclinical research and clinical trials to observational studies and case reports. ¹⁴ The biomedical literature has distinct advantages, including a diverse range of studies and reports for a broader population with various comorbidities and concurrent medications, covering extended observation periods, accumulating data to identify rare events, and continuously updating with new studies for ongoing drug monitoring post‐approval. ¹⁵ Furthermore, this literature comprises peer‐reviewed studies, ensuring that the information presented has been rigorously scrutinized for accuracy and reliability. However, effectively mining clinical evidence from the literature presents two primary challenges. First, the overwhelming volume of literature on DDIs complicates the extraction and analysis of relevant data, requiring substantial time, financial resources, and specialized expertise. To address this, advanced Natural Language Processing (NLP) techniques, including Named Entity Recognition (NER) and Relation Extraction (RE) models, have been developed. ¹⁶ For instance, an NLP analysis predicted novel DDIs by mining CYP‐substrate and CYP‐inhibitor relationships from PubMed. ¹⁷ Subsequent NLP studies have generated new pharmacogenetics hypotheses from translational DDI evidence discovered in the literature. ¹⁸ These NLP techniques hold significant promise for efficiently extracting crucial DDI‐SADR signals from extensive literature, enabling the identification of potential SADRs that might otherwise be missed. Second, once potential DDI‐SADR signals are identified, they must be validated to ensure their accuracy and clinical relevance. This validation is crucial as it distinguishes true signals from spurious ones, thereby enhancing the reliability of the findings. EHRs provide a robust resource for this validation process, allowing researchers to cross‐reference and confirm the identified DDI‐SADRs in real‐world clinical settings. ¹⁷ , ¹⁹

In this study, we employed NLP approaches for initial discovery alongside multi‐EHR databases for validation to detect underrecognized DDI‐SADR signals from existing literature, thereby advancing the field of PV and improving patient safety.

METHODS AND MATERIALS

Candidate substrate‐inhibitor pairs and SADRs for investigations

This study focused on DDIs involving the inhibition of cytochrome P450 enzymes CYP1A2, CYP2C9, CYP2C19, CYP2D6, and CYP3A, which are responsible for metabolizing approximately 80% of pharmaceuticals. ²⁰ We gathered information on all possible known interactions between inhibitors and substrates based on the U.S. Food and Drug Administration's (FDA) guidelines for drug development and interaction ²¹ and the Drug Interactions Flockhart Table. ²² Substrates and inhibitors were represented using their RxNorm ingredient names. Our analysis included inhibitors with strong (≥ 5‐fold increase in the plasma Area Under the Curve [AUC] values), moderate (≥ 2 to < 5‐fold increase in the plasma AUC values), and low (≥ 1.25 to < 2‐fold increase in the plasma AUC values) inhibition strength. We excluded inhibitors with indeterminate potency levels, such as those whose strength levels are unknown or for which only in vitro evidence is available. Any interactions with withdrawn drugs were also excluded.

Of the various ADRs, SADRs are particularly noteworthy due to their potential to cause severe or life‐threatening medical situations. According to the International Council for Harmonization (ICH) criteria for seriousness, SADRs were defined as outcomes that can result in death, life‐threatening scenarios, hospitalization, persistent or significant disability/incapacity, or congenital anomalies/birth defects. For this analysis, 7,523 Medical Dictionary for Regulatory Activities (MedDRA) Preferred Terms (PT) from the Important Medical Events (IME) list, as updated on March 22, 2023, by the MedDRA Maintenance and Support Services Organization (MSSO), ²³ were considered as SADRs.

PubMed query

The process of DDI‐SADR evidence extraction is shown in Figure 1 . To extract scientific articles on DDIs, we initially used PubMed as our primary resource. We carefully designed a PubMed query, hoping to achieve a high level of sensitivity in our search. Our search strategy was designed to identify a comprehensive set of articles related to DDIs: “drug interactions”[Title/Abstract, TIAB] OR “drug interactions”[MeSH Terms] OR “drug interaction”[Text Word, TW] to find the majority of relevant articles. To keep the focus on DDIs, we used several exclusions: “NOT food‐drug interactions “[MeSH Terms] OR “NOT herb‐drug interactions”[MeSH Terms] OR “NOT Review”[PT] OR “NOT Systematic Review” [Publication Type, PT]. These exclusions helped filter out articles not focused on DDIs or those that were not original research.

Figure 1.

The process of extracting and validating DDI‐SADR signals.

NER model

To effectively search for evidence of DDI‐SADR signals within the scientific literature, it is crucial to first accurately identify and extract mentions of drug and SADR entities from the text. This task necessitates a specialized NER model, a subset of NLP technology. NER models are adept at recognizing and isolating specific entities within sentences, in this case, drug and SADR entities. For this purpose, the PubTator Central database ²⁴ is employed, which stands as a prominent web service in the realm of biomedical text mining. This database is not only comprehensive but is also regularly updated on a daily basis, ensuring that the information it contains is current. Furthermore, the NER model utilized by PubTator Central is continuously re‐trained, enhancing its accuracy and performance over time. PubTator Centrol employed the pre‐trained TaggerOne ²⁵ for annotating drug and ADR entities, demonstrating impressive performance with normalization F1 scores of 89.5 for chemicals and 83.7 for diseases. ²⁴ One notable limitation of PubTator Central is that it only provides MeSH IDs for drugs and SADR entities. To address this, we used the MRCONSO.RRF and RXNREL.RRF files from the UMLS Metathesaurus, ²⁶ mapping drug entity MeSH IDs to RxNorm ingredients and SADR entity MeSH IDs to the MedDRA PT. Finally, we selected sentences with at least two drug entities (at least one substrate and at least one inhibitor) and one SADR entity to conduct a more in‐depth analysis of potential DDI‐SADR signals.

RE model

We employed a two‐phase methodology to extract DDI‐SADR signals from sentences: initially identifying interactions between potential drug combinations, and then assessing if these interactions contribute to SADR. To accomplish the initial phase, We utilized a Relation Extraction (RE) model to detect DDI relationships within sentences, specifically employing the SciFive PubMed Large model ²⁷ —a specialized text‐to‐text transfer transformer (T5) model designed for biomedical applications, pre‐trained on PubMed abstracts to enhance its biomedical literature interpretation. We applied the pre‐trained weights of the SciFive PMC Large model distributed by the authors and further fine‐tuned its parameters specifically for DDI extraction tasks. The training corpus was sourced from the DDI Extraction 2013 challenge, featuring 792 DrugBank texts and 233 Medline papers, categorized into five DDI interaction types: no interaction, DDI‐mechanism (PK DDI), DDI‐effect (pharmacodynamics drug–drug interactions [PD DDI]), DDI‐advice (advice regarding a DDI), and DDI‐int (DDI with no additional information). The SciFive PubMed Large model is one of the best methods available for extracting DDI relations, as evidenced by its performance on the DDI extraction 2013 corpus (F1 score: 83.63). ²⁷ If a candidate DDI sentence contained more than two drug entities, all substrate‐inhibitor combinations were examined, implying that a single sentence could contain multiple DDIs. Any drug pair with an interaction type other than no interaction was deemed to have an interaction. Lastly, sentences containing DDI evidence were manually reviewed by two reviewers (one with an M.S. in biomedical informatics and one with a Ph.D. in computer science) to determine whether drug pairs were the cause of SADR. Both reviewers had 3 years of experience studying DDIs. Cohen's Kappa was used to measure the degree of agreement between the two reviewers.

Evaluation of DDI‐SADR signals

We used the DrugBank database (version 5.1.10, released on April 1, 2023) ²⁸ to see if the DDI‐SADR signals identified in our study were already cataloged in DrugBank's DDI knowledge base. The DDI signals in the DrugBank database are based on a variety of sources, such as FDA and NIH labels, as well as scientific and clinical research including systematic reviews, randomized controlled trials, and case reports.

EHR validation

We utilized anonymized EHR databases from two primary sources: the Synthetic Derivative (SD) maintained by Vanderbilt University Medical Center (VUMC) ²⁹ , ³⁰ and the All of Us Research Program, managed by the National Institutes of Health (NIH), a nationwide clinical database in the U.S. ³¹ The VUMC SD database includes longitudinal research data that is de‐identified, covering clinical details from healthcare interactions of over 3 million individuals over more than 15 years. The All of Us database, at the time of our study, provided anonymized EHR data for more than 230,000 participants. The study received approval from the Institutional Review Board (IRB) at VUMC under the approval identifier number #221125. To standardize drug terms in EHRs to the RxNorm Ingredient level terms, we used the RXNREL.RRF file from the UMLS Metathesaurus. ²⁶ To identify SADR instances in the EHR database, the study first aligned MedDRA PT levels with ICD‐9, ICD‐10, and SNOMED‐CT, using the mapping tables provided by MedDRA. Then, we manually reviewed and added any potentially overlooked diagnosis codes for each SADR, using information from the existing literature.

For each substrate in DDI‐SADR signals, we identified a group of patients who had been exposed to the substrate at least once. The observation period was then established, which included the duration of continuous substrate use and a subsequent washout period (Figure 2 ). The washout period was set at five half‐lives of the drug, as this interval is sufficient to eliminate approximately 94%–97% of the drug's effects, thereby removing any residual impact. Throughout the observation period, we looked for concurrent exposure to inhibitor and substrate drugs. If a SADR occurred while the patient was exposed to both substrate and inhibitor, it was assumed to be caused by DDI. To isolate the impact of the chosen inhibitor, we excluded patients who consumed other drugs that could inhibit the same CYP enzymes as the candidate substrate. This approach was intended to eliminate the influence of other inhibitors on the occurrence of SADRs, focusing solely on the selected inhibitor.

Figure 2.

Timeline illustration of patients experiencing SADRs following concurrent use of substrate and inhibitor.

To reduce the effects of confounding variables, adjustments were made for factors related to patients, drugs, and diseases. For patient‐related factors, we looked at variables such as age, sex, and race. We accounted for drug‐related factors by considering the total number of different drugs used during the observation period up to the occurrence of a SADR (or throughout the entire observation period if no SADR occurred) as well as the total duration of substrates. Disease‐related factors include a variety of conditions known to increase the risk of ADRs, such as cardiovascular issues, diabetes, cancer, depression, impaired liver and kidney function, dementia, elevated white blood cell counts, and high cholesterol. ³²

We utilized Firth logistic regression, ³³ a penalized likelihood‐based method, with a sample size criterion of 5 events per variable (EPV) to identify inhibitors that elevated the risk of substrate‐SADR signals. Traditional logistic regression can be imprecise and biased when the number of EPVs is low. ³⁴ The Firth logistic regression approach is increasingly popular for mitigating small‐sample bias in maximum likelihood estimations (MLEs) by incorporating a penalty term that eliminates the first‐order term in the asymptotic bias expansion of MLEs. To address multiple testing issues, we used a Bonferroni correction with a type I error rate of 0.05. A DDI‐SADR signal was deemed significant if the regression odds ratio (OR) exceeded 1 and the Bonferroni‐corrected P‐value was less than 0.05.

DDI‐SADR visualization

To depict the general structure of DDI‐SADR signals derived from literature, a bipartite network was created. This type of graph, commonly used to represent relationships between two different categories of entities, features nodes of two types. One set of nodes represents substrates‐inhibitor pairs, while the other represents SADRs. Importantly, links are only formed between nodes of different types, with no connections between nodes of the same type. Connections between these two node types were established if at least one sentence of DDI‐SADR evidence was found in scientific literature. The size of each node was proportional to its number of neighbors (or degree). The edge types distinguish DDI‐SADR signals between those recorded in Drugbank and those that are not. The width of the edges was calculated as the sum of the min‐max normalized values of two article‐level metrics: total citation count and total altmetric score. ³⁵ Citation counts, which traditionally accumulate over the years, indicate long‐term scholarly impact as they are referenced in subsequent studies. In contrast, an article's altmetric score—reflecting its presence on the internet and social media—measures short‐term, immediate engagement and dissemination in the digital era. The total citation counts were sourced from Google Scholar, and altmetric scores were assessed using Altmetric Explorer, both on January 10, 2024. The color of the edges represented the confirmation status of DDI‐SADR signals, which varied from unconfirmed to verified by one EHR database and confirmed by two EHR databases.

RESULTS

The DDI sentences extracted from PubMed

We retrieved 160,321 candidate PubMed articles published between January 1962 and December 2023 through a search query. Using Pubtator Central for NER, we extracted 3,145 sentences from 2,732 articles that contained at least two drug entities (at least one substrate and one inhibitor) as well as at least one SADR entity. The SciFive model was then used to identify 531 sentences that demonstrated a DDI relationship between the substrate and inhibitor from a pool of 3,145 sentences. Through careful manual examination of these 531 sentences indicating potential DDI‐SADR interactions, we discovered that 111 of them contained SADRs directly attributable to the interaction between the substrate and inhibitor. The two independent reviewers had high concordance, with a kappa value (κ) of 0.95 (P < 0.001).

Statistical characteristics of DDI‐SADR evidence from the literature

Table S1 presents all DDI‐SADR evidence gathered from the literature. 111 DDI‐SADR evidence included 70 unique drugs (38 as substrates, 23 as inhibitors, and 9 as both substrate and inhibitor) and 53 unique SADRs in 92 articles. The publication years for DDI‐SADR evidence ranged from 1983 to 2023. While each DDI‐SADR evidence was investigated with an average of 1.06 articles (SD: 0.348), the most investigated DDI‐SADR signal had 2 articles, and the article with the most DDI‐SADR contained 3 signals. Warfarin was the most investigated substrate (12 DDI‐SADR signals), whereas clarithromycin was the most investigated inhibitor (17 DDI‐SADR signals). Rhabdomyolysis was the most extensively studied SADR (20 DDI‐SADR signals). The Drugbank database included 21 of the 111 DDI‐SADR signals that we extracted.

DDI‐SADR network

There were 88 nodes representing drug pairs and 53 nodes representing SADRs (Figure 3 ). Orange circular nodes represent substrate‐inhibitor pairs, whereas blue squares represent SADRs. Solid edges represent DDI‐SADR signals documented in the Drugbank database, while sinewave edges represent DDI‐SADR signals that were not documented in Drugbank. Green edges represent DDI‐SADR signals confirmed by one EHR database, while pink edges represent DDI‐SADR signals confirmed by two EHR databases. The interaction between phenytoin and amiodarone, which caused hemorrhage, had the highest combined total of citation counts and altmetric scores. We discovered that each substrate‐inhibitor pair was typically associated with a single SADR, but the majority of SADRs were linked to multiple substrate‐inhibitor pairs. The interaction between colchicine and clarithromycin was linked to the highest number of SADRs (4 SADRs), while rhabdomyolysis, followed by bradycardia and serotonin syndrome, was linked to the highest number of substrate‐inhibitor pairs, with 20, 8, and 6 drug pairs, respectively.

Figure 3.

DDI‐SADR bipartite network. The orange round nodes represent substrate‐inhibitor pairs, while the blue square nodes represent SADRs. The size of each node corresponds to the degree of connectivity. The line type of an edge distinguishes DDI‐SADR signals between those recorded in Drugbank and those that are not. Edge width is determined by the combined min‐max normalized values of total citation counts and altmetric scores. The color of an edge signifies the confirmation status of DDI‐SADR signals.

DDI‐SADR signals validations using the EHR database

The VUMC SD database included 273,491,202 records of diagnoses and 868,741,933 records of drug prescriptions for 3,183,571 patients, covering the period from January 1989 to May 2023. In contrast, the All of Us database recorded 86,012,976 diagnoses and 79,531,217 drug prescriptions for 239,715 patients from January 1981 to July 2022.

Table 1 presents 17 DDI‐SADR signals confirmed by at least one EHR database. These confirmed signals include 3 associated with strong inhibitors, 8 with moderate inhibitors, and 6 with weak inhibitors. One signal indicated an interaction involving the inhibition of two CYP enzymes: the combination of methadone and voriconazole, which inhibits CYP2B6 and CYP3A4, leading to arrhythmia. Thirteen signals were confirmed by one EHR database, while four were validated by both EHR databases. Nine signals were not documented in Drugbank, whereas eight were recorded in Drugbank with the exact or similar SADRs, all classified as moderate in severity. For example, interactions between mexiletine and amiodarone, as well as between citalopram and amiodarone, were noted in DrugBank as potential causes of tachyarrhythmias and QTc prolongation, respectively. Our analysis revealed that both interactions led to torsade de pointes, a condition highly associated with tachyarrhythmias and QTc prolongation. All nine unrecorded signals were confirmed by one EHR database (eight from VUMC and one from All of Us).

Table 1.

DDI‐SADR signals confirmed by at least one EHR database.

Substrate	Inhibitor	SADR	CYP enzyme (inhibition level)	Drugbank (DDI severity)	VUMC (OR, adjusted P‐value)	All of Us (OR, adjusted P‐value)
Warfarin	Amiodarone	Haemorrhagic disorder	CYP2C9 (moderate)	Bleeding (moderate)	2.45 (< 0.001)	–
Propafenone	Amiodarone	Torsade de pointes	CYP2D6 (weak)	Hypotension, somnolence, bradycardia, intra‐atrial and intraventricular conduction disturbances, convulsions, high‐grade ventricular arrhythmias (moderate)	4 (< 0.001)	> 1,000 (1)
Citalopram	Amiodarone	Torsade de pointes	CYP2D6 (weak)	QTc prolongation (moderate)	5.47 (< 0.001)	–
Methadone	Ciprofloxacin	Torsade de pointes	CYP3A (moderate)	QTc prolongation, torsades de pointes (moderate)	2.32 (0.0017)	7 (0.13)
Methadone	Ciprofloxacin	Respiratory depression	CYP3A (moderate)	QTc prolongation, torsades de pointes (moderate)	2.16 (< 0.001)	0.88 (0.8)
Oxycodone	Fluvoxamine	Clonus	CYP2D6 (moderate)	Serotonin syndrome, migraine, automatic instability (moderate)	2.88 (0.017)	–
Tramadol	Fluconazole	Hallucination	CYP3A (moderate)	–	2.54 (< 0.001)	0.74 (0.77)
Simvastatin	Fluconazole	Rhabdomyolysis	CYP3A (moderate)	–	1.9 (< 0.001)	–
Ibrutinib	Amiodarone	Atrial fibrillation	CYP2D6 (weak)	–	35.83 (< 0.001)	> 1,000 (1)
Fentanyl	Diltiazem	Delirium	CYP3A (strong)	–	1.94 (< 0.001)	5.1 (0.26)
Clarithromycin	Voriconazole	Acute kidney injury	CYP3A (strong)	–	9.19 (0.001)	–
Colchicine	Cyclosporine	Rhabdomyolysis	CYP3A (moderate)	–	0.87 (0.78)	2.63 (< 0.001)
Methadone	Voriconazole	Arrhythmia	CYP2B6 (moderate), CYP3A (strong)	–	3.45 (< 0.001)	–
Warfarin	Amiodarone	Hemorrhage	CYP2C9 (moderate)	Bleeding (moderate)	1.43 (< 0.001)	3.8 (< 0.001)
Lidocaine	Amiodarone	Seizure	CYP2D6 (weak)	Generalized seizure, bradycardia (moderate)	6.37 (< 0.001)	13.48 (< 0.001)
Oxycodone	Escitalopram	Serotonin syndrome	CYP2D6 (weak)	Serotonin syndrome (moderate)	1.92 (< 0.001)	29.1 (0.004)
Mexiletine	Amiodarone	Torsade de pointes	CYP2D6 (weak)	Tachyarrhythmias, electrolyte disturbance, gastrointestinal upset, heart block (moderate)	7.89 (< 0.001)	320.43 (< 0.001)

CYP, cytochrome P450; DDI, drug–drug interaction; OR, odds ratio; SADR, severe adverse drug reaction; VUMC, Vanderbilt University Medical Center.

DISCUSSION

The majority of DDI‐SADR signals discovered in our study did not appear in the DrugBank database. While our approaches extracted 111 DDI‐SADR signals, only 21 were also found in DrugBank. Despite DrugBank's extensive use of scientific literature and product labels to build its DDI repository, it has a limited focus on clinical evidence of PK DDIs, resulting in a relatively small number of such entries. Although DrugBank contains over 2.7 million DDI signals, there are only 578 with clinical evidence of PK DDIs, of which 433 are PK DDI‐SADR signals, encompassing just 28 unique SADRs. These discrepancies between our findings and the DrugBank database suggest that our methods have the potential to significantly enrich DrugBank's existing PK DDI‐SADR knowledge base.

Eight DDI‐SADR signals confirmed by at least one EHR were found in the DrugBank database, demonstrating the validity of our approach. For example, the well‐known warfarin‐amiodarone interaction, associated with major bleeding and extensively studied in the medical community, was discovered in biomedical literature and confirmed by two EHRs. ³⁶ , ³⁷ This interaction is primarily PK, as amiodarone inhibits the metabolism of warfarin by inhibiting cytochrome P450 enzymes, particularly CYP2C9 and CYP3A4, leading to increased plasma concentrations of warfarin and an enhanced anticoagulant effect. ³⁶ Additionally, there is also a PD aspect, as both drugs independently affect coagulation pathways, ³⁷ and their combined use can potentiate the risk of bleeding beyond what would be expected from PK interactions alone. Another example is the interaction between oxycodone and escitalopram, which leads to serotonin syndrome. Oxycodone, primarily an opioid pain reliever, exhibits serotonergic activity, as demonstrated in several studies, increasing the risk when combined with serotonergic drugs like escitalopram. ³⁸ While Drugbank classifies the severity of all eight signals as moderate, indicating they may or may not result in significant changes for a patient, our analysis confirmed these ADRs through EHR validation and suggests that their severity might be considered major.

Beyond the known DDI‐SADR signals identified by our pipeline, we discovered several DDI‐SADRs that were not documented in Drugbank but were confirmed by a single EHR and require further investigation. For example, while Drugbank indicates that methadone and ciprofloxacin may cause QTc prolongation, our research points to respiratory depression as a consequence of their interaction. Methadone, as an opioid analgesic, inherently carries risks of respiratory depression, particularly at higher doses. Its pharmacodynamic properties, which include the suppression of the central respiratory centers, increase the vulnerability to additional respiratory depressants. The inhibition of metabolizing enzymes by ciprofloxacin can lead to increased concentrations of methadone, thereby potentiating its respiratory depressant effects. ³⁹ Moreover, there is a PD component to this DDI, as ciprofloxacin, a fluoroquinolone antibiotic, has been associated with central nervous system (CNS) effects like dizziness and confusion, which may further impair respiratory function. ⁴⁰ Another example is the interaction between oxycodone and fluvoxamine, which is known to cause serotonin syndrome according to Drugbank but was found in our study to potentially cause clonus, a specific neuromuscular condition marked by rhythmic, involuntary muscle contractions. Oxycodone primarily exerts its analgesic effects by binding to opioid receptors in the CNS, predominantly μ‐opioid receptors, thereby modulating pain perception and transmission. However, the activation of these receptors by oxycodone can induce CNS depressant effects such as sedation, respiratory depression, and euphoria. ⁴¹ Additionally, fluvoxamine inhibits serotonin reuptake, leading to increased synaptic concentrations of serotonin. ⁴² The overactivation of serotonin receptors, particularly 5‐HT₂A receptors, by the elevated serotonin levels from fluvoxamine can lower the seizure threshold and increase neuronal excitability. Simultaneously, the CNS depressant effects of oxycodone may disrupt normal neuromuscular function, further contributing to the development of clonus. ⁴³ Our findings also uncovered the interaction between colchicine and cyclosporine resulting in rhabdomyolysis. Colchicine, commonly used to manage gout and familial Mediterranean fever, functions by blocking microtubule assembly, thereby diminishing the action of neutrophils and other inflammatory agents. High colchicine concentrations can disrupt muscle cell microtubules, causing muscle fiber breakdown and the release of intracellular contents into the bloodstream, ultimately leading to rhabdomyolysis. ⁴⁴ Cyclosporine, an immunosuppressant used primarily to prevent organ transplant rejection, curtails T‐cell activity. This drug interaction is particularly dangerous due to cyclosporine's inhibition of both CYP3A4 and P‐glycoprotein, critical for metabolizing and eliminating various drugs, including colchicine. Since colchicine is known for its narrow therapeutic index, ⁴⁵ meaning that small increases in dose can lead to toxicity, inhibition by cyclosporine can elevate colchicine levels, thereby increasing the risk of severe complications like rhabdomyolysis. ⁴⁶ Additionally, cyclosporine may have direct nephrotoxic effects, ⁴⁷ exacerbating the consequences of rhabdomyolysis by impairing the renal clearance of myoglobin and other toxic substances released from muscle cells.

In our previous study using FAERS and EHRs, ¹⁹ we found no overlap in EHR‐confirmed DDI‐SADRs between signals identified in FAERS and those from this study. This lack of overlap highlights the unique contributions each database can offer. For future research aimed at uncovering novel DDI‐SADRs, we recommend a multi‐modal approach that integrates databases such as FAERS and literature sources. This integration could enhance signal detection by expanding the search space, potentially identifying signals overlooked when relying on a single database.

The validated DDI‐SADR signals in this study provide insights into potential drug‐gene interactions (DGI). Although DDIs may not consistently indicate DGIs, as drugs can be metabolized through multiple pathways and may affect various enzymes, research indicates that DDIs form a reliable basis for predicting DGIs. ⁴⁸ The interaction of propafenone and amiodarone, which has been linked to torsade de pointes, was validated in the VUMC database. Earlier studies have connected CYP2D6 activity scores with the pharmacokinetics of propafenone, ⁴⁹ and a retrospective pediatric study found that lower CYP2D6 activity scores were associated with a higher risk of ADRs, including changes in electrocardiogram (ECG) and drug discontinuation. ⁵⁰ This implies that DGIs can be forecasted using DDIs, allowing for a more personalized treatment strategy that improves both drug efficacy and safety.

This study, while insightful, has several limitations. First, it is important to acknowledge that the high‐performance NER and RE models we used for extracting potential DDI‐SADR signals are not infallible, potentially leading to missed evidence. For example, our focus was exclusively on sentence‐level DDI evidence, which means we might have overlooked DDI evidence present beyond this scope. Second, although we accounted for various confounding variables during the EHR validation process, the high rate of missing data limited our ability to consider other potential confounders like genetic variants, drug dosage, and smoking status which are found to be important risk factors of DDI‐SADRs. These factors could significantly impact the study's findings. Furthermore, due to the broad range of SADRs examined, we were unable to focus on risk factors specific to individual SADRs and instead only addressed general risk factors associated with SADRs. Third, although we used two EHR databases to validate DDI‐SADR signals, the low event rates observed may have inflated odds ratios, making it difficult to confirm many signals. Low event rates can lead to large confidence intervals and unstable estimates, which may exaggerate the strength of associations. For instance, new DDI‐SADRs identified in the VUMC database might not have replicated in the All of Us database because there were few or no cases for several of these signals. Lastly, while our study concentrated on SADRs, we may have overlooked numerous non‐life‐threatening DDIs that can significantly impact patients' quality of life. Given the vast number of such interactions, it is challenging to capture all of them and determine which ones should be prioritized. Although these non‐life‐threatening DDIs were not the focus of our analysis, we believe that our pipeline can be generalized to detect them as well. This extension could enhance patient care by identifying interactions that, while not life‐threatening, substantially affect patients' daily lives.

FUNDING

This research was supported by the National Library of Medicine of the National Institutes of Health under Award Number R01LM014199, the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Number UL1TR002243, and academic program support funds from the Department of Medicine at Vanderbilt University.

CONFLICT OF INTEREST

The authors declared no competing interests for this work.

AUTHOR CONTRIBUTIONS

E.J. and Y.C. wrote the manuscript; E.J., Y.S., L.L., and Y.C. designed the research; E.J. performed the research; E.J., Y.S., L.L., and Y.C. analyzed the data.

Ethics Statement

This research utilized de‐identified electronic health records (EHRs) to protect patient privacy and confidentiality. All personal identifiers were removed in accordance with the Health Insurance Portability and Accountability Act (HIPAA) and institutional guidelines. Since the data were de‐identified, informed consent was not required. The study protocol was reviewed and approved by the Institutional Review Board (IRB) at Vanderbilt University Medical Center (VUMC) under approval number #221125, ensuring compliance with ethical standards for the use of human subjects in research.

Supporting information

Table S1

DATA AVAILABILITY STATEMENT

The PubMed dataset and PubTator Central are publicly available and can be found here: https://pubmed.ncbi.nlm.nih.gov/ and https://www.ncbi.nlm.nih.gov/research/pubtator/, respectively. The VUMC SD dataset is not publicly available. The All of Us Research EHR database is owned by a third party, the All of Us Research Program. The source code and datasets for this study have been made available on a public GitHub repository (https://github.com/OHPENL/DDISADR). The code is freely accessible under the terms of the Apache License (http://www.apache.org/licenses/).