Drug-drug interactions (DDIs) present a significant challenge in pharmacotherapy, a situation that is becoming more complex due to the aging global population and the increasing rates of chronic multimorbidity. DDIs arise when two or more drugs taken together influence each other’s pharmacokinetic or pharmacodynamic properties. This interaction can lead to a decrease in therapeutic effectiveness, unexpected side effects, or even severe, life-threatening consequences. The issue is particularly pronounced with the rise of polypharmacy, especially in elderly individuals and hospitalized patients, which has drawn increased attention from clinicians, researchers, and regulatory agencies focused on understanding and managing these interactions effectively.

Traditionally, identifying DDIs has depended on retrospective methods, including clinical observations, post-marketing surveillance, and spontaneous reporting of adverse events. Although these approaches provide valuable insights, they often suffer from fragmentation and lack the sensitivity and timeliness needed for proactive pharmacovigilance. Consequently, many DDIs go undetected, leading to preventable adverse drug events and adding strain to healthcare systems. Alarmingly, around 30% of adverse drug reactions (ADRs) are associated with DDIs, with a considerable number of these interactions remaining unrecognized in clinical practice.

In recent years, significant advancements in systems biology, pharmacokinetics, and molecular pharmacology, along with the emergence of artificial intelligence (AI), machine learning (ML), and network pharmacology, have transformed DDIs research. These innovative technologies facilitate the large-scale prediction and mechanistic investigation of potential DDIs, frequently uncovering risks before they become apparent in clinical settings.

Despite the promising advancements in the field, there are still notable gaps that need to be addressed. Many current reviews tend to overlook recent developments in computational methods, as well as the valuable real-world data derived from electronic health records (EHRs). Additionally, they often fail to consider the specific DDIs risks that vulnerable populations, such as the elderly and critically ill patients, face. This highlights the urgent need for a comprehensive and up-to-date synthesis that tackles these challenges effectively. While tools like the STOPP/START criteria are increasingly being utilized to minimize potentially inappropriate medications (PIMs) and DDIs events in geriatric care, further research is essential to fully understand and address the wide range of DDIs risks present in these populations.

This review presents a comprehensive synthesis of the current understanding of DDIs, covering epidemiological trends, mechanistic insights, predictive methodologies, and regulatory perspectives. It emphasizes the impact of emerging technologies, particularly artificial intelligence (AI) and pharmacogenomics, on the detection and prevention of DDIs. Unlike earlier reviews that typically focus on specific mechanistic or computational elements, this work combines AI, pharmacogenomics, and regulatory science into a cohesive framework, providing a more expansive view of DDIs research.

To assist readers in navigating this dynamic landscape, we present a comprehensive framework (Figure 1) that encompasses five essential components: epidemiological patterns, mechanistic classifications, AI-driven prediction methodologies, risk factors affecting vulnerable populations, and regulatory strategies. Central to this framework is artificial intelligence (AI), which serves as a crucial link connecting fundamental scientific research, clinical applications, and health policy.

This integrative approach highlights the crucial role of combining pharmacological knowledge with data-driven innovation to effectively shape the future of DDIs management.

Certain populations experience a higher incidence of DDIs. A retrospective cohort study published in Frontiers in Pharmacology by Schneider et al. (2021) revealed that older adults suffering from comorbid conditions such as hypertension, diabetes, and depression were prescribed an average of eight medications each day. This polypharmacy resulted in 3 to 5 potentially dangerous drug-drug interactions (pDDIs) per patient. Alarmingly, more than 60% of these interactions were deemed clinically significant, posing risks such as bleeding events or reduced therapeutic effectiveness. Thus, polypharmacy presents a complex challenge in geriatric care, where the goal of effective treatment must be carefully weighed against the potential for harmful interactions (Swinglehurst et al., 2023).

The prevalence of potentially dangerous drug interactions (pDDIs) is notably higher in intensive care units (ICUs) and long-term care facilities. A population-based study published in the Journal of the American Medical Directors Association by Anrys et al. (2021) found that 77.3% of elderly nursing home residents experienced at least one pDDIs, with 18.5% categorized as high-risk. Such interactions can lead to serious consequences, including delirium, hemorrhage, and acute kidney injury.

Real-world evidence has consistently highlighted certain drug combinations as particularly dangerous. Among the notable high-risk DDIs are:

• Warfarin + NSAIDs: Elevates bleeding risk through both pharmacokinetic (CYP inhibition) and pharmacodynamic (antiplatelet) mechanisms.

These interactions often appear gradually and can manifest in various ways, including elevated liver enzymes, subtle changes in neurological function, or a diminished response to treatment. This makes it challenging for clinicians to detect these issues without the use of proactive screening tools.

Figure 2 illustrates the average DDIs risk scores, measured on a scale from 3 to 5, across various organ systems, categorized by clinical risk levels: “Medium-High,” “High,” and “Very High.” The use of darker shades indicates higher average risk scores. It is particularly noteworthy that the renal and cardiovascular systems are consistently associated with very high-risk interactions, highlighting the critical need for careful monitoring in these areas.

Figure 3 presents a two-dimensional mapping of 30 high-risk DDIs pairs. In this visualization, drug combinations are plotted along the X-axis, which represents the affected physiological system, and the Y-axis, which denotes the identity of the drugs involved. The color coding indicates the clinical risk level associated with each interaction, with blue representing medium-high risk, gold indicating high risk, and red signifying very high risk. Additionally, the shape and size of the points on the graph correspond to the risk score of each interaction, where circles represent a score of 3, squares indicate a score of 4, and diamonds denote a score of 5. This graphical representation allows for the quick identification of priority interactions across various physiological systems, enhancing the ability to assess and manage potential risks in clinical settings.

Large-scale pharmacovigilance platforms, including the FDA Adverse Event Reporting System (FAERS), EudraVigilance, and institutional electronic health records (EHRs), offer crucial epidemiological insights into DDIs. Analyses of data from FAERS indicate that approximately 12%–15% of serious adverse drug reactions (ADRs) could be linked to either known or previously unrecognized DDIs. However, several challenges persist in this area, such as underreporting of adverse events and the absence of standardized taxonomies for categorizing DDIs.

Recent efforts utilizing artificial intelligence (AI) and natural language processing (NLP) to analyze real-world data (RWD) have uncovered the often-overlooked burden of DDIs. These advanced tools are capable of extracting structured signals of interactions from unstructured clinical notes, which frequently leads to the discovery of clinically significant interactions that are not recorded in traditional reference databases. Additionally, studies based on electronic health records (EHR) have emphasized the existence of population-specific DDIs profiles, particularly within pediatric, pregnant, and geriatric groups.

Despite growing awareness, many DDIs continue to go unrecognized, often due to factors such as delayed onset, nonspecific clinical symptoms, or a lack of familiarity among clinicians. For example, interactions involving warfarin may first appear as minor bruising, which can later escalate to more severe complications like gastrointestinal or intracranial hemorrhage. In a similar vein, interactions affecting the central nervous system (CNS) frequently present with symptoms like fatigue or cognitive changes, which may be mistakenly attributed to the natural progression of an underlying disease rather than the effects of medications.

The consequences of missed DDIs can be quite serious. Research indicates that as many as 30% of hospitalizations related to adverse drug reactions (ADRs) could be avoided with proper monitoring of DDIs. When these interactions go undetected, it can lead to ineffective treatment, higher rates of illness, and, in certain instances, even death.

Given the increasing frequency and seriousness of DDIs, it is essential to adopt proactive measures for their detection. Incorporating DDIs screening into clinical practices can significantly reduce their public health implications. This can be achieved by utilizing clinical decision support systems (CDSS), conducting pharmacist-led medication reviews, and providing ongoing medical education for healthcare professionals. These strategies work together to enhance awareness and management of DDIs, ultimately improving patient safety and outcomes.

Risk stratification that incorporates demographic and clinical variables can significantly enhance precision pharmacovigilance. For instance, elderly patients who are on multiple medications, known as polypharmacy, could be automatically identified as high-risk for dangerous drug combinations through the use of machine learning-enabled CDSS. Additionally, it is essential to prioritize the creation of region-specific drug-drug interaction (DDI) surveillance tools that are customized to reflect local prescribing habits and genetic profiles. This approach will improve risk mitigation strategies by ensuring they are context-sensitive and relevant to the populations being served.

Pharmacokinetic DDIs occur when one medication influences the absorption, distribution, metabolism, or excretion (ADME) of another, leading to alterations in plasma drug levels and, as a result, changes in pharmacological effects. A notable example of this is lithium, which necessitates careful therapeutic monitoring because of its significant potential for interactions (Fiorillo et al., 2023).

Pharmacodynamic DDIs arise when two medications affect the same or interconnected physiological pathways, resulting in effects that can be additive, synergistic, or antagonistic.

Various classification systems have been proposed to stratify DDIs based on severity, molecular mechanism, and clinical management recommendations:

• Severity-based grading (e.g., minor, moderate, major): Reflects the clinical consequences, ranging from negligible to life-threatening.

Emerging tools like molecular docking, enzyme phenotyping, and quantitative systems pharmacology (QSP) models offer robust frameworks for predicting DDIs at a mechanistic level. For instance, CYP phenotyping panels and probe cocktails are instrumental in assessing metabolic pathways, allowing researchers to understand how different drugs may interact within the body. Additionally, in silico tools such as SimCYP and GastroPlus play a crucial role by simulating drug behavior under various physiological conditions, providing insights into how drugs might perform in real-world scenarios. When these advanced technologies are combined with clinical pharmacogenomics, they significantly enhance our capacity to predict and prevent clinically significant DDIs, ultimately improving patient safety and treatment outcomes.

CDSS, which are usually integrated with electronic health records (EHRs), depend on carefully curated DDIs knowledge bases to produce alerts for healthcare providers. However, a significant challenge to the effective use of these systems is the absence of a standardized way to represent DDIs data, which hinders interoperability and the ability to reuse this information across different platforms (Hochheiser et al., 2021). These rule-based systems work by comparing prescribed drug pairs with established lists of interactions. Some prominent commercial databases that provide such DDIs information include:

• Micromedex–Provides interaction severity, clinical effects, and management strategies.

Despite their usefulness, rule-based systems often produce a high number of false positives, which can lead to clinician desensitization, commonly referred to as “alert fatigue.” Furthermore, these systems typically do not consider patient-specific factors such as age, organ function, or pharmacogenomics, highlighting the necessity for more flexible and data-driven solutions. CDSS implemented in community pharmacies have demonstrated effectiveness in minimizing adverse drug reactions by offering real-time alerts for potential drug interactions. A scoping review indicates that CDSS tools in these settings can significantly enhance drug safety and mitigate medication-related issues (Moon et al., 2024). Electronic prescription platforms also contribute by delivering immediate feedback on drug interactions, enabling clinicians to make well-informed decisions and avert possible adverse drug interactions. Research has shown that these platforms markedly enhance the safety and effectiveness of clinical decisions (Grammatikopoulou et al., 2024a). Additionally, medication safety initiatives like the EQUIPPED program have proven to significantly decrease adverse drug reactions by broadening the detection of drug interactions across healthcare systems, utilizing both traditional and hub-and-spoke models (Vandenberg et al., 2024).

With the rapid growth of biomedical data, computational models have become essential for predicting DDIs, especially those that have not yet been observed in clinical settings. These models combine various data sources, including chemical structures, biological pathways, electronic health records (EHRs), scientific literature, and gene expression data, to estimate the likelihood of interactions. For instance, the dual graph neural network model developed by Ma and Lei in 2023 effectively integrates molecular structures and drug interaction information, significantly improving the accuracy of DDIs predictions (Ma and Lei, 2023). Similarly, the feature extraction model based on graph neural networks introduced by Al-Rabeah and Lakizadeh in 2022 also enhances the precision of predicting DDIs (Al-Rabeah and Lakizadeh, 2022). To further increase the accuracy and clinical relevance of these predictions, it is crucial to develop more personalized and context-aware models. Muylle et al., in 2021 demonstrated that optimizing the integration of context-specific data can greatly enhance DDIs screening and management through their context-aware clinical decision support system (Muylle et al., 2021). This is further supported by Zhu et al. in 2021, who presented an attribute-supervised tri-factorization model for predicting DDIs (Zhu et al., 2021). Additionally, the AutoDDI model proposed by Gao et al., in 2024, which employs automated graph neural networks, improves both the accuracy and efficiency of DDIs predictions (Gao et al., 2024).

Artificial intelligence (AI) has demonstrated significant promise in the management of polypharmacy, particularly by enhancing the accuracy of predicting drug interactions among elderly individuals and patients with multiple chronic conditions. A review conducted in Saudi Arabia highlighted how AI models can effectively identify high-risk drug combinations associated with polypharmacy (Alsanosi and Padmanabhan, 2024). In this regard, deep learning techniques have made notable strides in predicting DDIs by analyzing complex patterns within extensive datasets. For instance, models like DrugBERT utilize large-scale textual data to achieve a deeper semantic understanding, thereby improving the accuracy of DDIs predictions by extracting relevant information from clinical literature (Liu et al., 2025). Additionally, other models, such as Graph Neural Networks (GNNs), leverage molecular graphs to capture the intricate relationships between different drugs, demonstrating promising results in terms of both accuracy and interpretability (Yao et al., 2024).

The integration of Transformer-based models such as DrugBERT and BioBERT has become increasingly popular, facilitating more advanced interactions and improving predictions through a deeper semantic understanding derived from extensive medical literature (Qi et al., 2025; Hu et al., 2024). These innovations, when paired with efficient natural language processing (NLP) pipelines like DDIExtractor and specialized knowledge bases, enable the early identification of emerging DDIs, often before these interactions are documented in structured clinical or experimental databases.

Systems pharmacology aims to model the effects of drugs within interconnected biological networks by integrating pharmacokinetics (PK) and pharmacodynamics (PD). Knowledge graphs (KGs) serve as a semantic representation of the relationships among drugs, targets, pathways, diseases, and phenotypes. To effectively encode these semantic relationships and facilitate cross-platform integration, standardized information models are crucial (Hochheiser et al., 2021). Although KGs enhance reasoning capabilities and interpretability, their construction and maintenance demand significant expert curation and computational resources. Notable platforms that support these efforts include:

• DrugBank and PharmGKB–Curated molecular and clinical pharmacology data.

These tools facilitate explainable and mechanistically grounded inference of DDIs. The SAO semantic structure-based forecasting method offers a robust framework for predicting adverse drug reactions, especially in cases of polypharmacy, where the risks associated with drug interactions can be intricate and challenging to detect (Wang JY. et al., 2024).

In recent years, AI-based knowledge graph mining methods have demonstrated significant promise in predicting DDIs. By utilizing deep learning techniques and graph neural networks (GNNs), researchers can automatically analyze extensive biomedical data to identify relationships between various drugs, thereby aiding in the discovery of potential DDIs. Some of the prevalent AI-based methods include:

• Graph Neural Networks (GNNs): GNNs excel at identifying intricate, nonlinear relationships among entities, including drugs, targets, and diseases within knowledge graphs. By utilizing node and edge embeddings, GNNs can effectively forecast various types of DDIs. For example, methods such as GraphSAGE and Graph Convolutional Networks (GCN) are commonly employed for the classification and prediction of DDIs, demonstrating their utility in this domain (Wang Yaqing et al., 2024).

These AI methods improve the accuracy of drug-drug interaction (DDI) predictions and offer a scientific foundation for personalized medication therapy, especially in situations involving multiple medications.

Electronic health records provide valuable real-world evidence that can help identify potential DDIs. By retrospectively analyzing medication histories, lab results, and clinical outcomes, researchers can uncover associations that may not have been recognized before. Time-aware models, particularly those based on transformer architectures, are capable of capturing the temporal aspects of drug exposure and the associated risks of interactions. Nonetheless, there are challenges to consider, such as the variability in data, missing information, and confounding factors. To address these issues, methods like federated learning and differential privacy are being investigated to facilitate large-scale DDI analysis while preserving patient privacy across different institutions. Additionally, electronic prescription systems are crucial for monitoring DDIs, especially through real-time alerts that notify healthcare providers of potential risks. A study conducted in Greece demonstrated that CDSS can significantly enhance the accuracy of clinical decision-making, particularly in managing DDIs (Grammatikopoulou et al., 2024b).

Pharmacogenomic (PGx) variability is essential in understanding how individuals respond differently to DDIs. Variations in genes like CYP2C9, CYP2D6, and SLCO1B1 can significantly influence how drugs are metabolized or transported in the body. For example, individuals who are poor metabolizers of CYP2D6 may face a higher risk of experiencing toxicity from medications such as codeine or tricyclic antidepressants. By incorporating PGx data into clinical DDIs risk models, healthcare providers can enhance precision medicine, tailoring treatments to individual genetic profiles. Tools like PharmCAT and CPIC guidelines support the integration of genetic information into treatment decisions, which is particularly important in high-risk areas such as oncology, psychiatry, and cardiology. Personalized treatment strategies are vital for managing complex drug interactions, especially among psychiatric patients. Research by Cuomo et al. (2024) highlights that clinical responses to vortioxetine can vary widely among different patient profiles. In psychiatric patients with additional health conditions, the risk of DDIs may increase due to the simultaneous use of antidepressants and medications for physical illnesses (Berk et al., 2023). Furthermore, the growing use of psychedelics, whether prescribed or recreational, underscores the need for a better understanding of potential DDIs and effective communication between clinicians and patients (Boehnke et al., 2023). Additionally, gender-diverse individuals undergoing hormone therapy may present unique DDIs profiles, particularly in the context of psychiatric pharmacotherapy (Kim et al., 2023).

Despite substantial progress, several challenges remain:

• Data quality and standardization–Heterogeneous ontologies and inconsistent annotations hinder model interoperability. Efforts toward minimal information models for DDIs are a step toward solving these limitations (Hochheiser et al., 2021).

Looking ahead, the integration of artificial intelligence with systems pharmacology, digital twin modeling, and real-time electronic health record (EHR) analytics presents promising opportunities for the advancement of next-generation DDIs surveillance. Key elements such as explainable AI (XAI), causal inference frameworks, and collaboration with regulatory bodies will be essential in maintaining clinical relevance, ensuring transparency, and prioritizing patient safety. Recent advancements in artificial intelligence, particularly in areas like graph neural networks and multimodal deep learning, have notably enhanced the accuracy of predicting potential drug-drug interactions (Zhang et al., 2023; Wang N-N. et al., 2024; Huang et al., 2023; Zhao et al., 2024; Li et al., 2023). Additionally, insights from patients and caregivers play a vital role in recognizing and addressing DDI-related risks, especially in outpatient settings (Sharma et al., 2024).

As biomedical data continues to expand, the challenge of analyzing health data across different institutions while protecting data privacy has become increasingly critical. Traditional methods of data analysis often involve centralized data storage, which poses significant privacy risks. In contrast, federated learning techniques enable data to be processed locally, thereby preserving patient privacy and facilitating data sharing between institutions. This approach is especially vital for predicting and managing drug-drug interactions. For instance, the privacy-preserving federated learning framework introduced by Sinaci et al. (2024) effectively tackles privacy concerns associated with cross-institutional data collaboration, all while improving the efficiency of data analysis.

Polymorphic variants in drug-metabolizing enzymes, such as CYP2C9/2C19, CYP2D6, and CYP3A5*3, play a crucial role in influencing the risk of drug-drug interactions (DDIs). For instance, individuals identified as poor metabolizers of CYP2D6 may be at an increased risk of developing serotonin syndrome when paroxetine is used in conjunction with tramadol. By incorporating pharmacogenomic (PGx) profiling into standard healthcare practices, utilizing platforms like PharmCAT or YouScript, healthcare providers can conduct personalized risk assessments and make proactive dose adjustments tailored to individual patient needs.

• Dynamic DDIs Modeling: Integration of real-time PK/PD data and digital phenotyping for adaptive therapy.

The U.S. Food and Drug Administration (FDA) has established a thorough and risk-informed framework for evaluating drug-drug interactions (DDIs), which includes both laboratory (in vitro) and clinical (in vivo) studies. The guidance updated in 2020 highlights the importance of using physiologically based pharmacokinetic (PBPK) modeling as a key method for predicting interactions that occur due to metabolism and transport processes. The FDA supports model-informed drug development (MIDD), which combines PBPK with quantitative systems pharmacology (QSP) to enhance the design of studies and improve regulatory decision-making. Additionally, the FDA has promoted the use of artificial intelligence (AI)-driven models that utilize extensive clinical data to facilitate real-time detection of DDIs.

The FDA is progressively incorporating AI tools into the assessment of DDIs, especially within CDSS. To ensure these AI models are effective in real-world settings, they must undergo thorough validation that includes not just data from clinical trials but also real-world evidence (RWE) sourced from electronic health records (EHRs). This approach helps confirm that the AI tools are applicable in everyday clinical practice. Additionally, the FDA emphasizes the necessity for clear labeling that accurately conveys the underlying mechanisms of the identified DDIs and their therapeutic consequences. It is crucial that AI-generated predictions are presented in a way that is clinically actionable, thereby providing valuable support to both healthcare providers and patients.

The European Medicines Agency (EMA) offers comprehensive guidance on evaluating pharmacokinetic interactions, placing significant importance on mechanistic classification, especially concerning cytochrome P450 enzymes and drug transporters. Within the EMA framework, there is a clear requirement for strong justification when waiving clinical interaction studies, frequently necessitating sensitivity analyses or supplementary modeling to substantiate these claims. In contrast to the FDA, the EMA applies more stringent regulatory oversight regarding the reasoning behind the decision not to perform certain clinical trials.

The European Medicines Agency (EMA) has begun to integrate artificial intelligence (AI) into the evaluation of DDIs, acknowledging its ability to predict intricate interactions and categorize DDIs based on underlying mechanisms. Nevertheless, the application of AI tools must be firmly rooted in thorough scientific validation and must adhere to the agency’s commitment to transparency and data sharing. To support this initiative, the EMA provides an open-access DDIs database, which aids in the incorporation of AI-driven tools designed to uncover previously underreported or newly emerging DDIs.

The Pharmaceuticals and Medical Devices Agency (PMDA) in Japan generally aligns with international best practices while also taking into account specific regional considerations. A key focus for the PMDA is ethnic sensitivity; they often require bridging studies to ensure that DDIs data can be accurately extrapolated to Japanese populations. Additionally, the PMDA has been increasingly supportive of modeling and simulation techniques, including the use of AI-based systems, to predict potential DDIs effectively.

The PMDA has approached the endorsement of AI-driven models with caution, yet it is gradually advancing towards supporting these technologies. These models have the potential to incorporate population-specific data, which is particularly important for enhancing DDIs predictions. This is especially relevant in the context of Japanese populations, where unique genetic factors can significantly influence drug metabolism.

While national regulatory agencies have their own specific nuances, there is a growing trend towards convergence in DDIs evaluation strategies, largely influenced by ongoing efforts to harmonize guidelines through the International Council for Harmonisation (ICH). The upcoming ICH M12 guideline marks a significant step forward, as it seeks to standardize the scientific principles that govern DDIs study design, data interpretation, and labeling requirements. Among the leading regulatory bodies, the U.S. Food and Drug Administration (FDA) has taken a proactive approach by advocating for model-informed and simulation-based regulatory science. In contrast, the European Medicines Agency (EMA) and the Pharmaceuticals and Medical Devices Agency (PMDA) are gradually incorporating these methodologies, reflecting their unique institutional and regional priorities. Despite these differences, it is crucial for these agencies to align their efforts to reduce redundant testing, speed up global drug development, and maintain consistency in communicating clinical risks. As these global agencies strive for harmonization, recent regulatory guidance—such as the FDA’s protocols for CYP450 and transporter studies (U.S. Food and Drug Administration FDA, 2020)—illustrates a developing consensus. Additionally, emerging artificial intelligence tools, like the affinity models summarized on arXiv, provide further translational benefits in this evolving landscape (Vefghi et al., 2025).

There is a growing movement aimed at achieving global consistency in the use of artificial intelligence (AI) tools for predicting DDIs. AI-driven models must comply with international regulations concerning data privacy, transparency of the models, and their validation processes. New AI technologies, such as graph neural networks (GNNs) and transformer-based models, provide significant advantages, allowing for more precise and context-sensitive predictions of DDIs. As international organizations strive for harmonization, it is essential that these AI models adhere to the same stringent standards that are applied to conventional DDIs assessment tools, thereby ensuring their relevance in clinical settings and safeguarding patient safety.

Many current DDIs prediction models are based on limited and skewed datasets, which primarily focus on well-characterized drug classes. This focus leads to a significant underrepresentation of rare interactions, population-specific pharmacokinetics, and the complexities of real-world treatment scenarios. Additionally, clinical trials frequently exclude vulnerable populations, including the elderly, pediatric patients, pregnant individuals, and those with multiple health conditions. Such exclusions undermine the external validity of DDIs evidence, particularly for high-risk groups.

A considerable number of machine learning-based DDIs prediction models face challenges related to interpretability, which diminishes trust among clinicians and hampers acceptance by regulatory bodies. These models frequently function as “black boxes,” delivering precise predictions without offering insights into their underlying mechanisms or pharmacological reasoning. Consequently, this lack of transparency restricts their incorporation into clinical workflows and the labeling of drugs.

Current DDIs frameworks inadequately address the role of herbal medicines, dietary supplements, and traditional therapies, even though these are commonly used, especially in low- and middle-income countries (LMICs). The diversity in therapeutic practices across different geographic and cultural contexts adds another layer of complexity that is seldom reflected in conventional pharmacovigilance systems or regulatory guidelines.

Most DDIs risk stratification tools are static, which means they do not adapt to the dynamic physiological changes, evolving comorbidities, or treatment adjustments that patients experience over time. Currently, the integration of real-time patient data, including electronic health records (EHRs), therapeutic drug monitoring (TDM), and longitudinal health trends, remains underutilized in modern DDIs management systems. This lack of adaptability and real-time data incorporation limits the effectiveness of these tools in providing personalized and timely assessments of DDIs risks, ultimately impacting patient safety and treatment outcomes.

Future research should concentrate on creating explainable and clinically validated DDIs prediction models that combine mechanistic pharmacology with sophisticated data-driven techniques. It is essential to prioritize the generalizability of these models across various patient populations and clinical contexts, including intensive care units, oncology settings, and resource-limited environments.

The convergence of pharmacogenomics, mobile health technologies, and digital infrastructure presents remarkable opportunities for personalized assessment of DDIs risks. By integrating genotype-guided dosing algorithms with real-time decision support tools and mobile patient engagement platforms, we can facilitate a dynamic and individualized approach to pharmacovigilance. This integration allows healthcare providers to tailor medication plans based on a patient’s genetic profile, ensuring safer and more effective treatment options while actively engaging patients in their healthcare journey.

The establishment of global DDIs surveillance networks, along with the implementation of interoperable data standards, can significantly improve consistency across different regulatory jurisdictions. It is essential for regulators, clinicians, informaticians, and technology developers to work together across sectors to effectively translate computational advancements into practical tools that enhance the safety of pharmacotherapy.

To tackle the aforementioned challenges, we propose an integrative translational framework that connects diverse data sources, predictive modeling strategies, and various application domains. This framework, illustrated in Figure 6, is designed to enhance the prediction of DDIs through the use of AI-powered models. It effectively integrates heterogeneous data sources with AI-based prediction methodologies and real-world applications, emphasizing the challenges faced and the solutions available for translating these insights into clinical practice.

The foundational data inputs essential for DDIs prediction come from various sources:

• Molecular Mechanisms: The data encompasses how drugs interact at the molecular level, including details about chemical structures, biological pathways, and the interactions that occur within biological systems.

The second section of the framework involves AI-based prediction methods that process the heterogeneous data sources mentioned above. These methods include:

• Machine Learning: Algorithms that learn from large datasets to predict drug interactions based on patterns and correlations.

The predictions generated by AI-based models are directed toward practical applications:

• CDSS: AI-powered tools that assist healthcare providers in making informed decisions about drug prescriptions, minimizing harmful interactions.

While the framework holds promise, several challenges remain in implementing AI-powered DDI predictions:

• Interpretability: AI models, especially deep learning, can be difficult to interpret, making it challenging to understand the rationale behind predictions.

Artificial intelligence (AI) and machine learning (ML) are transforming the detection and prediction of DDIs by facilitating scalable, data-driven methodologies. Advanced deep learning models, such as graph neural networks (GNNs) and transformer-based architectures like BERT-DDI, are capable of revealing intricate molecular interactions and identifying novel DDIs from extensive compound libraries and biomedical literature. Additionally, natural language processing (NLP) techniques enhance the analysis of electronic health records (EHRs) and adverse event reports, allowing for the identification of hidden interaction signals. Nevertheless, for these technologies to gain wider regulatory acceptance, there is a need for improvements in model interpretability, clinical validation, and effective risk communication.

The increasing occurrence of polypharmacy necessitates the implementation of systems-level strategies. Tools such as network pharmacology models, agent-based simulations, and cumulative DDIs risk scores allow researchers to assess multi-drug treatment plans in a comprehensive manner. These methodologies are particularly significant in fields like geriatrics, oncology, and psychiatry, where the simultaneous administration of various medications is frequently observed.

DDIs mechanisms are evolving beyond traditional pathways, introducing new complexities such as epigenetic modulation, microbiome-driven metabolism, and immune-mediated enzyme regulation. For instance, histone deacetylase inhibitors have been shown to influence the expression of CYP3A4, an important enzyme in drug metabolism. Additionally, the gut microbiota plays a significant role in drug metabolism, as it can either activate or inactivate certain medications and alter the host’s metabolic capacity. These emerging mechanisms highlight the need for their incorporation into existing DDIs prediction frameworks to enhance our understanding and management of drug interactions.

Genetic polymorphisms play a significant role in influencing the drug behaviors of both perpetrators and victims. New models that focus on gene–drug–drug interactions (GDDIs) are being developed alongside electronic health record (EHR)-integrated CDSS to provide alerts tailored to specific genotypes. By incorporating polygenic risk scores and dynamic enzyme activity models, these systems may improve the assessment of risks associated with drug interactions. Pharmacogenomics is essential for personalizing treatment, particularly in the realm of psychiatric medications, where it can help prevent drug intolerance. For example, pharmacogenetic testing has been shown to reduce the risk of adverse drug reactions in patients prescribed antipsychotics, as highlighted in a specific case report (Hudnik et al., 2024).

To improve study efficiency and ensure relevance to real-world applications, innovative methodologies like adaptive trials, microdosing studies, and physiologically based pharmacokinetic (PBPK)-informed trial designs are becoming more popular. Regulatory agencies are increasingly supporting simulation-based strategies as an integral component of model-informed drug development (MIDD).

The adoption of ICH M12 and the establishment of open-access DDIs databases, such as ONC-DDI and FDA DDIR, signify a significant shift towards harmonized regulatory science and enhanced transparency in the field. These initiatives are supported by collaborative consortia that are actively working to set standards, promote data sharing, and train artificial intelligence models. This collaborative effort is crucial in accelerating global advancements in DDIs safety, ensuring that stakeholders can access and utilize vital information effectively.