Artificial Intelligence in Predictive Modeling of Drug–Drug Interactions: Advances, Applications, and Future Directions

Abstract

Drug–drug interactions (DDIs) pose a significant clinical challenge, contributing to adverse drug events (ADEs), hospitalizations, and increased healthcare costs. Conventional DDI detection methods such as in vitro assays, in vivo studies, and post-marketing surveillance—provide valuable mechanistic insights but are limited by high cost, time requirements, and incomplete coverage of possible drug combinations. The rapid expansion of available pharmaceuticals has intensified the need for scalable, accurate, and timely predictive approaches. Artificial intelligence (AI) and machine learning (ML) have emerged as transformative tools capable of integrating large, heterogeneous datasets spanning chemical structures, pharmacokinetics, pharmacodynamics, electronic health records (EHRs), and biomedical literature to uncover complex interaction patterns that may not be evident through traditional methods. This review provides a comprehensive overview of AI-based predictive modeling techniques for DDI identification, encompassing traditional ML algorithms, deep learning architectures, graph-based models, and natural language processing (NLP) approaches. It examines key data sources, including structured databases (e.g., DrugBank, ChEMBL, KEGG), real-world clinical data, and literature mining, and discusses strategies for data preprocessing, feature engineering, and model evaluation. Special attention is given to challenges such as data sparsity, class imbalance, lack of interpretability, and limited generalizability to novel drug combinations. Emerging solutions—including explainable AI (XAI), transfer learning, multimodal data integration, and federated learning—are highlighted as promising directions to enhance transparency, robustness, and clinical applicability. Applications of AI-driven DDI prediction span clinical decision support systems, drug development pipelines, and pharmacovigilance frameworks, with case studies demonstrating reductions in medication errors and improvements in patient safety. The review also addresses regulatory perspectives, ethical considerations, and integration challenges in clinical workflows. Ultimately, AI-based predictive modeling offers a powerful, adaptive, and scalable approach to mitigating DDI risks, supporting precision medicine, and advancing drug safety across healthcare systems.

Keywords

Introduction

Figure 1. Schematic representation of pharmacokinetic and pharmacodynamic processes involved in drug–drug interactions

Figure 2. Metabolic processing of xenobiotics in hepatocytes via CYP-mediated oxidation, conjugation, and elimination, with pathways leading to detoxification or reactive metabolite formation.

Figure 3. Workflow of Random Forest–based predictive modeling for drug–drug interaction (DDI) prediction.

Figure 4: Workflow of the Know DDI framework combining DDI and biomedical knowledge graphs to generate embeddings, extract subgraphs, and classify predicted DDI types.

This schematic illustrates a graph-based AI approach for predicting drug–drug interactions by merging DDI graphs with knowledge graphs containing biomedical entities. Through embedding generation, subgraph extraction, and feature integration, the model captures both structural and semantic relationships between drugs, enabling accurate prediction of interaction types.

6.4 Natural Language Processing (NLP)

Natural Language Processing (NLP) plays a crucial role in the prediction and discovery of drug-drug interactions (DDIs) by enabling the automated extraction of relevant information from vast amounts of unstructured textual data. Scientific literature, clinical notes, electronic health records (EHRs), and adverse event reports contain valuable insights about DDIs that are often buried within complex narratives, making NLP indispensable for efficient knowledge extraction. A primary NLP application in DDI prediction is text mining, which involves identifying, retrieving, and structuring information related to drug interactions from heterogeneous text sources. Text mining techniques can process millions of biomedical articles and clinical documents to uncover previously unreported or poorly characterized interactions, thereby enriching existing DDI databases and supporting clinical decision-making (67) Key tasks within NLP for DDI extraction include Named Entity Recognition (NER) and Relation Extraction (RE). NER focuses on identifying and classifying drug names, dosages, adverse events, and other biomedical entities within text. Accurate recognition of these entities is fundamental for subsequent analysis and linking to structured databases. Following entity recognition, Relation Extraction (RE) aims to detect and characterize the relationships between identified entities—specifically, interactions between drugs. RE models utilize rule-based, supervised, or deep learning approaches to determine whether two or more drugs mentioned in a text have a potential interaction, along with the nature and severity of the interaction. Recent advances have leveraged transformer-based architectures, such as BERT and its biomedical adaptations, to significantly improve the precision and recall of DDI extraction from complex sentences. NLP techniques face challenges including ambiguous terminology, diverse writing styles, and incomplete reporting in clinical texts. Nonetheless, combining NLP with AI-driven predictive modeling enhances the comprehensiveness and accuracy of DDI prediction frameworks by integrating knowledge from both structured databases and unstructured textual data (68).

7. Model Evaluation and Validation Strategies

The development of robust and reliable AI-based models for drug-drug interaction (DDI) prediction demands rigorous evaluation and validation to ensure performance, generalizability, and clinical relevance. Evaluation not only helps identify model strengths and limitations but also guides model selection and optimization for real-world deployment. The assessment of DDI models typically involves quantitative performance metrics, validation techniques, external benchmarking, and growing concerns regarding interpretability and explainability.

7.1 Common Evaluation Metrics

o effectively evaluate the performance of drug-drug interaction (DDI) predictive models, a variety of statistical metrics are utilized. These metrics provide comprehensive insights into different aspects of model accuracy, reliability, and discriminative power, which are critical for both model comparison and optimization during training and validation phases.

7.1.1 Accuracy

Accuracy represents the proportion of total correct predictions made by the model, encompassing both true positive and true negative outcomes. While accuracy provides a straightforward measure of overall correctness, it can be misleading in the context of DDI prediction due to the common issue of dataset imbalance. Specifically, since non-interacting drug pairs vastly outnumber interacting pairs, a model that predominantly predicts no interaction can achieve high accuracy without effectively identifying true interactions (69).

7.1.2 Precision

Precision quantifies the ratio of correctly predicted positive interactions to all predicted positive instances. This metric reflects the model’s capacity to minimize false positives, thereby indicating how reliably the model identifies actual interactions among the predicted ones. High precision is especially valuable in clinical settings to avoid unnecessary alerts or interventions based on incorrect interaction predictions.

7.1.3 Recall (Sensitivity)

Recall, also known as sensitivity, measures the proportion of actual interacting drug pairs that the model successfully identifies. It assesses the model’s ability to detect true DDIs, which is essential for ensuring potentially harmful interactions are not overlooked. A high recall reduces the risk of false negatives, which is crucial for patient safety and effective pharmacovigilance.

7.1.4 F1-Score

The F1-score is the harmonic mean of precision and recall, providing a balanced evaluation metric that accounts for both false positives and false negatives. This measure is particularly useful in datasets with class imbalance, such as those encountered in DDI prediction, where it offers a more nuanced assessment of model performance compared to accuracy alone.

7.1.5 Area Under the Receiver Operating Characteristic Curve (ROC-AUC)

The ROC-AUC metric evaluates the model’s ability to discriminate between interacting and non-interacting drug pairs across all classification thresholds. By plotting the true positive rate against the false positive rate, ROC-AUC summarizes the trade-offs between sensitivity and specificity. This metric is widely adopted in biomedical binary classification tasks due to its robustness and interpretability, with higher values indicating superior discriminative performance. Collectively, these metrics form an essential toolkit for assessing, comparing, and refining AI models designed for DDI prediction, thereby enhancing their clinical relevance and reliability (70).

7.2 Cross-Validation Techniques

Cross-validation represents a fundamental methodology for assessing the performance of predictive models, particularly in scenarios where labeled datasets are limited, as is often the case in drug-drug interaction (DDI) research. This technique helps to prevent overfitting by providing an unbiased evaluation of the model's generalizability to unseen data.

7.2.1 K-Fold Cross-Validation

The most widely used form of cross-validation is k-fold cross-validation. In this approach, the entire dataset is randomly partitioned into k approximately equal-sized subsets, known as folds. During each iteration, the model is trained on k–1 folds and subsequently tested on the remaining single fold. This process is repeated k times, ensuring that each fold serves as the test set exactly once. By averaging performance metrics over the k iterations, this method yields a more reliable and stable estimate of model effectiveness compared to a single train-test split (71).

7.2.2 Stratified Cross-Validation

When dealing with DDI datasets, which typically exhibit significant class imbalance between interacting and non-interacting drug pairs, stratified cross-validation becomes particularly valuable. This variant preserves the original class distribution within each fold, ensuring that the proportion of positive and negative samples remains consistent across all training and testing subsets. Maintaining class balance prevents biased performance estimates that could arise if some folds contain disproportionately few positive interaction examples, thereby enhancing the robustness and clinical relevance of the model evaluation. Together, these cross-validation strategies are essential for optimizing machine learning models in DDI prediction, facilitating reliable performance assessment even in the context of limited or imbalanced datasets (72).

7.3 External Validation and Benchmarking Datasets

To rigorously assess the generalizability of drug-drug interaction (DDI) predictive models, external validation using independent datasets is indispensable. Unlike cross-validation, which primarily evaluates model robustness within a single dataset, external validation tests the model’s performance on entirely separate and unseen data sources. This step is critical to confirm that predictive models maintain accuracy and reliability when applied beyond their original training environment, thereby supporting their practical utility in diverse clinical settings (73). Several widely accepted benchmark datasets serve as standard resources for external validation in DDI modeling. These include:

Drug Bank, which provides comprehensive, curated information on approved and experimental drugs, including detailed interaction profiles.

TWOSIDES, a database compiling statistically inferred adverse drug interaction side effects derived from large-scale clinical datasets.

SIDER, a resource cataloging marketed drugs alongside their documented adverse drug reactions, facilitating identification of clinically relevant DDI outcomes. Datasets curated from Bio Creative DDI extraction challenges, which contain annotated drug interaction data extracted from biomedical literature, supporting evaluation of text-mining based prediction methods. Utilizing these datasets for external validation offers several advantages. First, they establish a standardized framework for benchmarking, allowing researchers to objectively compare novel computational algorithms against established approaches. Second, testing models across multiple independent datasets reflects the heterogeneity and complexity of real-world data, including variations in drug combinations, patient populations, and clinical contexts. This broader evaluation enhances the confidence in model robustness and increases the likelihood of successful translation into clinical practice (74).

7.4 Interpretability and Explainability in DDI Models

As artificial intelligence (AI) models, particularly those based on deep learning and graph-based architectures, grow in complexity, ensuring interpretability and explainability has become a crucial requirement for fostering trust and promoting clinical adoption. Interpretability is defined as the degree to which human users can comprehend the underlying logic and decision-making process of a model, whereas explainability pertains to the ability to provide meaningful and understandable justifications for specific model predictions. To achieve these goals, various computational tools have been developed. Notably, SHAP (SHapley Additive explanations) and LIME (Local Interpretable Model-agnostic Explanations) are widely employed to elucidate the influence of individual features on a model’s output. These methods assign importance scores to input features, thereby allowing researchers and clinicians to assess whether the model’s decisions align with known biological mechanisms or pharmacological principles (75). In the context of drug-drug interaction (DDI) prediction, explainability techniques can reveal the molecular or mechanistic basis underlying a predicted interaction. This insight not only supports hypothesis generation for further experimental validation but also enhances drug safety evaluations by clarifying potential risk factors.Moreover, embedding explainable AI (XAI) practices within DDI predictive frameworks aligns with emerging ethical and regulatory standards in healthcare. Transparency in AI-driven decision-making processes is increasingly mandated to ensure patient safety, accountability, and fairness, ultimately facilitating broader acceptance and integration of AI-assisted drug interaction prediction tools in clinical practice (76).

8. APPLICATIONS AND USE CASES

The adoption of AI-based predictive modeling for drug-drug interactions (DDIs) has transitioned from purely academic research into practical applications across clinical, pharmaceutical, and regulatory domains. These models are proving instrumental in enhancing patient safety, streamlining drug development, and improving post-market surveillance.

8.1 Integration of Predictive Models into Clinical Decision Support Systems (CDSS)

One of the most transformative applications of artificial intelligence (AI)-driven drug-drug interaction (DDI) prediction lies in its incorporation within Clinical Decision Support Systems (CDSS). These systems are designed to assist healthcare professionals by providing evidence-based guidance during clinical workflows, particularly at the point of prescribing or medication reconciliation. By embedding AI-based DDI predictive models, CDSS can proactively identify and flag potentially harmful drug combinations, thereby significantly mitigating the risk of adverse drug events (ADEs) and improving patient safety outcomes. Conventional CDSS tools typically depend on static rule sets or curated interaction databases. While useful, these approaches are often limited by their inability to detect novel, rare, or complex DDIs that fall outside established knowledge bases. In contrast, AI-enhanced CDSS leverage dynamic algorithms capable of analyzing molecular structures, pharmacological mechanisms, and real-world clinical data. This enables a more comprehensive and context-sensitive evaluation of potential interactions, adapting to new evidence and diverse clinical scenarios. Moreover, advanced AI-powered CDSS have begun integrating patient-specific clinical parameters, such as age, renal function, hepatic status, and the presence of comorbidities, to tailor interaction alerts to individual risk profiles. Such personalized alerting systems help reduce the problem of “alert fatigue” — where clinicians become desensitized to frequent, non-specific warnings — thereby improving the clinical relevance and usability of interaction notifications. Hospitals employing these machine learning-driven DDI alerts have reported enhanced decision-making efficiency and a reduction in preventable medication errors, illustrating the practical benefits of AI integration into routine healthcare delivery (77).

8.2 Use in Drug Development and Post-Marketing Surveillance

Artificial intelligence (AI)-based drug-drug interaction (DDI) prediction is increasingly transforming various stages of drug discovery and development pipelines. During the preclinical and clinical phases, pharmaceutical companies utilize AI-driven predictive models to proactively identify potential interaction risks associated with new drug candidates. By anticipating harmful DDIs early in the development process, these models help mitigate the likelihood of costly late-stage failures, streamline the drug development timeline, and improve overall drug safety profiles before regulatory submission. In the post-marketing phase, AI models play a critical role in pharmacovigilance activities by continuously analyzing large-scale real-world data sources. These include electronic health records (EHRs), spontaneous adverse event reporting systems, and patient registries, which provide comprehensive and diverse patient-level information beyond the controlled environment of clinical trials. AI algorithms can detect emerging signals of previously unrecognized or rare DDIs, including those not captured during clinical studies due to limited sample sizes, population diversity, or study durations (78). Regulatory agencies such as the U.S. Food and Drug Administration (FDA) and other international counterparts have recognized the potential of AI technologies to enhance traditional pharmacovigilance methods. These bodies increasingly support the integration of AI tools to supplement existing safety monitoring frameworks, particularly for the early identification and evaluation of rare but clinically significant drug interactions, thereby contributing to improved public health outcomes (79).

8.3 Real-World Case Studies

Numerous real-world case studies have demonstrated the effectiveness of artificial intelligence (AI)-driven models in predicting drug-drug interactions (DDIs) and improving healthcare outcomes. These applications span drug discovery, clinical decision support, and pharmacovigilance, highlighting the broad utility of AI technologies in both research and clinical domains. One prominent example is IBM Watson for Drug Discovery, which integrates machine learning algorithms and natural language processing (NLP) techniques to mine extensive biomedical literature. This platform has been instrumental in uncovering previously unknown DDIs, particularly in complex therapeutic areas such as oncology and polypharmacy. By processing vast textual datasets, Watson has been able to provide actionable insights that inform drug safety research and guide clinical decisions in high-risk populations. In another case, researchers constructed a graph-based neural network model leveraging curated datasets from DrugBank and SIDER. This model was designed to predict novel DDIs by capturing the complex relationships between drugs, targets, and side effects within a structured network. The predictions were then validated using data from the FDA Adverse Event Reporting System (FAERS), and the model demonstrated superior performance in both precision and recall compared to traditional rule-based and machine learning approaches, reinforcing the potential of deep learning in DDI identification. A further example of clinical impact comes from a collaboration between Mayo Clinic and computational biology researchers, who implemented an AI-powered Clinical Decision Support System (CDSS) specifically tailored to identify harmful drug combinations frequently prescribed to elderly patients with multiple chronic conditions. By integrating patient-specific factors and pharmacological data, the system significantly reduced DDI-related hospitalizations, underscoring its role in enhancing patient safety, particularly in vulnerable populations. These case studies illustrate the transformative role of AI-based DDI prediction models not only in improving individual patient care but also in influencing broader public health strategies, regulatory frameworks, and comprehensive drug lifecycle management. As these technologies continue to evolve, they hold immense promise for advancing precision medicine and ensuring safer therapeutic practices (80).

9. CHALLENGES AND LIMITATIONS

While AI-based predictive modeling has significantly advanced the field of drug-drug interaction (DDI) prediction, several key challenges and limitations continue to hinder its full integration into clinical and pharmaceutical settings. These challenges range from data-related issues and model interpretability to regulatory barriers and ethical considerations.

9.1 Data Sparsity and Imbalance

A fundamental challenge in drug-drug interaction (DDI) predictive modeling is data sparsity, which arises because only a limited subset of all possible drug pairs has been experimentally tested or clinically documented. Considering the vast combinatorial space of potential drug combinations, the known and annotated DDIs represent only a small fraction, resulting in highly sparse interaction matrices that hinder comprehensive learning by AI models (81). In addition to sparsity, class imbalance poses a significant obstacle: the number of negative examples (i.e., drug pairs with no known interaction) greatly exceeds the number of positive, interacting pairs. This imbalance can bias predictive algorithms toward favoring the majority class, leading to an increased rate of false negatives and reducing the model’s sensitivity to true DDIs. These challenges of data sparsity and class imbalance not only limit the predictive accuracy but also impair the generalizability of models, especially when dealing with newly approved drugs or rare interaction events that lack sufficient training examples. To mitigate these issues, several strategies have been employed, including synthetic data generation through techniques like data augmentation, oversampling of minority classes, and cost-sensitive learning approaches that penalize misclassification of positive interactions more heavily. Despite these efforts, developing datasets that are both truly representative and balanced remains an ongoing challenge, underscoring the need for innovative solutions to enhance model robustness and reliability (82).

9.2 Black-Box Nature of Deep Models

Deep learning has become a cornerstone of modern drug-drug interaction (DDI) prediction, with numerous models based on architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and graph neural networks (GNNs) showing promising predictive capabilities. However, a critical limitation of these models is their black-box nature—their internal decision-making processes are highly complex and not easily interpretable by human users. This opacity in reasoning poses substantial challenges, particularly in clinical contexts where decisions informed by AI must be transparent, explainable, and ethically justifiable. Physicians, pharmacists, and other healthcare professionals need to understand the basis of AI-generated predictions to confidently incorporate them into patient care, especially when those decisions involve potential risks to patient safety. To address these concerns, recent advancements in explainable artificial intelligence (XAI) have introduced methods aimed at improving interpretability of complex models. Tools such as Shapley Additive explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) have been developed to attribute predictions to specific input features, providing insights into how individual variables contribute to the final output of an AI model. These tools offer post hoc explanations that are crucial for verifying the clinical plausibility of predicted DDIs and for fostering trust among medical practitioners. However, while XAI techniques have demonstrated potential, their applicability and reliability in high-stakes clinical environments remain an active area of research. Limitations include difficulty in interpreting explanations in very high-dimensional or interdependent data, and questions regarding the consistency and fidelity of the generated justifications. As such, further validation of these tools is necessary before they can be confidently relied upon in real-time clinical decision support systems (83).

9.3 Generalization to Unseen Drug Combinations

A significant and persistent limitation in artificial intelligence (AI)-based drug-drug interaction (DDI) prediction is the limited generalizability of models to previously unseen drug combinations. While many machine learning and deep learning models demonstrate high accuracy when evaluated on familiar compounds within established datasets, their performance often declines when applied to novel drugs, experimental agents, or drug repurposing scenarios. This limitation becomes particularly critical in contexts where rapid and accurate interaction assessments are needed—such as during global health emergencies (e.g., pandemics), or in personalized medicine, where patient-specific therapies may involve unique or non-standard drug combinations. The challenge arises partly due to the incomplete coverage of existing DDI datasets, which tend to be biased toward well-studied drugs and interactions. As a result, AI models trained on these datasets may overfit to known data distributions and struggle to extrapolate beyond them. This lack of extrapolative power restricts the clinical utility of AI systems in dynamic drug development pipelines and limits their effectiveness in emerging therapeutic areas.To address this issue, researchers have explored various strategies aimed at improving model generalizability. Transfer learning has been proposed as a promising approach, wherein knowledge gained from one domain or dataset is transferred to another related task involving novel drug entities. Data augmentation techniques, including the generation of synthetic drug pairs based on known interaction patterns, have also been employed to enrich training datasets and expose models to a broader chemical space. Additionally, incorporating chemical and biological similarity metrics—such as molecular fingerprints, target homology, or pathway overlap—into model architectures has shown potential for enhancing predictive accuracy on previously unseen drug combinations.  Despite these advancements, ensuring robust generalization across diverse drug spaces remains an open challenge in the field. Continued efforts are needed to design models that can reliably predict interactions involving newly approved drugs, off-label uses, or therapies introduced in response to emerging diseases (84).

9.4 Ethical, Legal, and Regulatory Concerns

While artificial intelligence (AI)-based models hold immense promise for enhancing drug-drug interaction (DDI) prediction, they also raise a number of ethical, legal, and regulatory concerns that must be carefully addressed to ensure their responsible and equitable use in healthcare settings. One of the foremost issues is data privacy, especially when using sensitive patient information from electronic health records (EHRs) for training predictive models. The utilization of such data must adhere to stringent privacy laws, including the Health Insurance Portability and Accountability Act (HIPAA) in the United States and the General Data Protection Regulation (GDPR) in the European Union, which mandate safeguards for patient confidentiality, informed consent, and data security (85). Another major concern is algorithmic bias, which can arise from imbalanced or non-representative training datasets. For instance, underrepresentation of certain demographic groups—such as racial minorities, older adults, or patients with rare diseases—can result in models that generate inequitable predictions, disproportionately affecting these populations and potentially reinforcing existing healthcare disparities. Addressing such biases requires deliberate efforts to ensure diversity in training data and transparency in model development processes. From a regulatory perspective, oversight of AI-driven DDI tools remains an evolving field. Regulatory agencies, including the U.S. Food and Drug Administration (FDA), have recognized the transformative potential of AI in drug safety monitoring and personalized medicine. However, these agencies emphasize the need for robust validation, clinical evidence, and post-market surveillance to ensure the safety, efficacy, and reliability of AI-based systems before and after deployment. Regulatory guidance documents increasingly call for transparency in model design, interpretability of outputs, and clearly defined accountability mechanisms, particularly when AI tools are integrated into clinical decision-making processes. Collectively, these ethical and regulatory dimensions underscore the importance of a responsible AI framework that promotes fairness, transparency, and accountability while safeguarding patient rights. Navigating these challenges is essential for building public trust and enabling the safe integration of AI technologies into clinical pharmacology and drug safety practices (86).

9.5 Integration into Clinical Practice

Despite substantial technical progress in the development of artificial intelligence (AI) models for drug-drug interaction (DDI) prediction, integration into clinical practice remains a considerable challenge. One of the primary obstacles is clinician skepticism regarding the reliability and usability of AI tools. Healthcare professionals may be reluctant to adopt these systems due to concerns about predictive accuracy, lack of interpretability, and the potential for workflow disruption during patient care. In high-stakes medical environments, tools that are not transparent or intuitive may be perceived as intrusive or even risky, especially if they generate false alerts or fail to provide justifiable recommendations. In addition, the technological infrastructure within many healthcare institutions is often inadequately prepared to support AI integration. Electronic health record (EHR) systems and clinical decision support systems (CDSS) may not be designed with the interoperability necessary to accommodate complex AI algorithms. The integration of AI models requires seamless data exchange, real-time processing, and compatibility with existing clinical workflows—requirements that are not universally met across health systems (87). Addressing these barriers will require more than just technical refinement of AI models. It demands a multifaceted implementation strategy involving interdisciplinary collaboration between data scientists, clinicians, informaticians, and administrators. Education and training programs should be developed to familiarize healthcare providers with the capabilities and limitations of AI-based tools, thereby enhancing user confidence and promoting responsible usage. Furthermore, clear operational guidelines must be established for the deployment, maintenance, and monitoring of these models in real-world healthcare environments, ensuring that AI integration enhances rather than disrupts clinical care. By focusing on these organizational and human factors—alongside technical innovation—stakeholders can facilitate successful adoption of AI-based DDI tools, ultimately contributing to safer prescribing practices and improved patient outcomes.

10. FUTURE DIRECTIONS

The evolution of AI-based predictive modeling for drug-drug interactions (DDIs) is poised to transform drug safety, clinical decision-making, and personalized therapy. Despite significant progress, future advancements must address current limitations and expand the scope of DDI prediction to make it more precise, interpretable, patient-centered, and ethically responsible. Several promising directions are emerging at the intersection of data science, medicine, and regulatory innovation.

10.1 Multimodal and Patient-Specific Data Integration

The next generation of drug-drug interaction (DDI) prediction models is anticipated to move beyond single-source data and embrace multimodal data integration as a core principle. This evolution involves the convergence of diverse and complementary data types—including genomics, proteomics, metabolomics, electronic health records (EHRs), drug labels, and biomedical literature—to develop more comprehensive and accurate representations of drug behavior and interaction mechanisms (88). By synthesizing these heterogeneous datasets, future models will be better equipped to capture the multifactorial and system-wide nature of drug interactions, which often vary based on biological context and population-specific characteristics. A major advancement will be the incorporation of patient-specific variables, such as genetic polymorphisms (e.g., variations in cytochrome P450 enzymes), comorbidities, hepatic or renal function, and concomitant therapies, all of which significantly influence pharmacokinetics and pharmacodynamics. This paradigm shift from population-level to individual-level modeling will enable context-aware DDI predictions, allowing for tailored assessments that consider the unique physiological and molecular profile of each patient. Such personalized modeling holds particular promise for improving drug safety and efficacy in vulnerable populations, including individuals experiencing polypharmacy, the elderly, or patients with rare diseases, where traditional DDI knowledge is often sparse or generalized. Additionally, advances in artificial intelligence (AI), such as deep learning and knowledge graph-based models, will further facilitate the dynamic integration of these multimodal inputs, enhancing predictive performance and clinical relevance. In summary, the future of AI-driven DDI prediction is characterized by a move toward precision pharmacology, where individualized risk assessments replace static, one-size-fits-all alerts, thus supporting safer, more effective, and equitable medication practices (89).

10.2 Personalized DDI Prediction and Precision Medicine

The incorporation of artificial intelligence (AI) into precision medicine is playing a transformative role in advancing personalized drug-drug interaction (DDI) prediction systems. Unlike traditional models that generalize interaction risks across broad populations, emerging AI frameworks are being designed to deliver individualized risk assessments by leveraging patient-specific data. These models integrate pharmacogenomic information, such as genetic polymorphisms influencing drug metabolism, along with longitudinal health records, including medication history, comorbidities, and clinical laboratory results, to more accurately predict how particular drug combinations may affect a specific patient (90). Such personalization enables dynamic risk stratification tailored to the patient’s current physiological and clinical context, allowing for proactive intervention. In addition to risk prediction, personalized DDI models contribute to dose optimization, suggesting adjusted dosing regimens based on predicted metabolic capacity or organ function. They also facilitate therapeutic substitution, recommending safer alternatives when high-risk interactions are identified—thus reducing the likelihood of adverse drug events (ADEs) and improving overall treatment outcomes. Looking forward, these intelligent systems could be further integrated into wearable health technologies, mobile health applications, or digital therapeutics platforms, offering real-time DDI alerts not only to clinicians but also to patients and caregivers. Such innovations would empower users with on-demand decision support, enhancing medication safety in outpatient and home-care settings where direct clinical supervision may be limited. This synergy between AI and precision medicine underscores the shift toward patient-centric pharmacotherapy, where treatment strategies are guided by individual data profiles, leading to more effective and safer healthcare delivery (91).

10.3 Explainable AI (XAI) in Pharmacovigilance

As artificial intelligence (AI) tools used in drug-drug interaction (DDI) prediction continue to evolve in complexity—particularly with the adoption of deep learning and graph-based neural networks—the need for explainable AI (XAI) has become increasingly urgent. Unlike traditional statistical models, which offer inherent transparency, many AI models operate as "black boxes", producing highly accurate outputs without providing insight into their internal reasoning. This lack of interpretability poses a significant barrier to clinical implementation, especially in high-stakes domains like pharmacovigilance, where decisions informed by AI can influence drug labeling, risk communication, or even regulatory approval and market withdrawal. To address this, future DDI models must prioritize transparent and interpretable outputs that are understandable and actionable by clinicians, researchers, and regulators alike. Explainability in this context goes beyond technical curiosity; it fosters trust, accountability, and regulatory compliance, ensuring that the AI system’s rationale aligns with known pharmacological mechanisms and clinical evidence. Several XAI techniques have emerged to enhance model interpretability. Notably, SHAP (SHapley Additive Explanations) provides a unified framework for attributing model predictions to specific input features, offering insights into how much each variable (e.g., enzyme inhibition, drug structure) contributed to a predicted DDI. Similarly, LIME (Local Interpretable Model-agnostic Explanations) approximates complex model behavior with locally interpretable surrogate models, revealing which factors most influenced a prediction in a particular case. Attention mechanisms, commonly used in transformer-based architectures, also help by highlighting relevant sections of molecular sequences or clinical data that the model “attends to” during prediction. Incorporating these XAI tools enables interpretation of DDI predictions in terms of chemical structure similarities, shared molecular pathways, or documented adverse clinical outcomes, making the models not only accurate but also scientifically and clinically meaningful. This approach is essential for integrating AI safely and effectively into drug development pipelines, regulatory workflows, and real-world clinical practice (92).

10.4 Federated Learning and Privacy-Preserving AI

Federated learning (FL) has emerged as a transformative paradigm in artificial intelligence (AI), particularly valuable in domains such as healthcare where data privacy and security are of paramount concern. Unlike conventional centralized learning frameworks, where data is aggregated into a central server for model training, FL allows AI models to be trained locally on decentralized data sources—such as individual hospitals, clinics, or healthcare institutions—without the need to transfer sensitive patient-level data. Instead of data, only model parameters or gradients are shared and aggregated centrally, thereby preserving privacy while still facilitating collaborative learning. This approach significantly enhances compliance with data protection regulations such as the Health Insurance Portability and Accountability Act (HIPAA) in the United States and the General Data Protection Regulation (GDPR) in the European Union. By ensuring that personally identifiable information (PII) remains within the originating institution, federated learning minimizes legal and ethical risks while maintaining the utility of real-world datasets for machine learning applications. In the context of drug-drug interaction (DDI) prediction, FL offers a promising solution to overcome the fragmentation of healthcare data across geographic and institutional boundaries. By enabling secure, multi-institutional collaboration, federated DDI models can benefit from heterogeneous patient populations, diverse prescribing patterns, and region-specific drug usage trends, thereby improving model robustness and generalizability. For example, data from hospitals treating geriatric patients in Europe can be combined with pediatric data from North America and oncology datasets from Asia, all without exposing patient identities. Moreover, federated frameworks can accelerate pharmacovigilance efforts, especially in detecting rare or delayed-onset DDIs that may not be evident in isolated datasets. The decentralized nature of FL also supports real-time model updates, enabling dynamic learning as new drug interactions or adverse events emerge in different parts of the world. By balancing data privacy, regulatory compliance, and AI-driven discovery, federated learning represents a crucial step forward in making large-scale, trustworthy, and globally relevant DDI prediction models a clinical reality (93).

10.5 Collaboration Among Academia, Healthcare, and Regulatory Bodies

The advancement and successful implementation of artificial intelligence (AI) models for drug-drug interaction (DDI) prediction rely fundamentally on interdisciplinary collaboration among various stakeholders. These stakeholders include researchers, clinicians, data scientists, pharmaceutical companies, and regulatory agencies, who must work together to establish and adhere to comprehensive standards for model validation, deployment, and ongoing monitoring to ensure safety, efficacy, and clinical utility. Such collaboration facilitates the development of robust, clinically relevant AI tools that are both scientifically sound and aligned with healthcare needs. For instance, researchers provide algorithmic innovation and technical expertise, clinicians contribute domain knowledge and practical insights, data scientists ensure proper handling and interpretation of complex datasets, pharmaceutical companies integrate models into drug development pipelines, and regulatory agencies oversee compliance with legal and ethical standards. Several collaborative initiatives exemplify this multidisciplinary approach. The U.S. Food and Drug Administration (FDA) has launched the Artificial Intelligence and Machine Learning (AI/ML) Action Plan, which outlines a framework for advancing regulatory science and fostering innovation in AI applications for healthcare. Similarly, the European Medicines Agency (EMA) has established the Big Data Task Force, aimed at creating guidance and policies for the use of big data and AI in medicine, including pharmacovigilance and drug safety assessment. Additionally, multi-institutional consortia such as Observational Health Data Sciences and Informatics (OHDSI) promote the harmonization of real-world data and open science principles to enhance collaborative research and benchmarking of AI models across diverse healthcare datasets .Together, these efforts demonstrate the critical importance of cooperation and shared governance in overcoming the technical, ethical, and regulatory challenges inherent in deploying AI-driven DDI prediction systems. They foster an ecosystem where innovation is balanced with patient safety and public trust, ultimately enabling the translation of AI advances into improved clinical outcomes. Such efforts are essential for aligning AI innovation with clinical needs and public health goals, ensuring that predictive DDI models are safe, equitable, and beneficial for all stakeholders (94).

CONCLUSION

Artificial intelligence has emerged as a transformative approach for predicting drug–drug interactions (DDIs), offering the capability to integrate heterogeneous biomedical data, model complex pharmacokinetic and pharmacodynamic relationships, and identify novel interaction patterns beyond the reach of conventional methods. By leveraging machine learning, deep learning, graph-based models, and natural language processing, AI-driven systems can analyze chemical structures, biological pathways, clinical records, and real-world pharmacovigilance data with remarkable scalability and precision. These models hold promise for early DDI detection, personalized risk assessment, and continuous post-marketing surveillance, thereby improving patient safety and optimizing therapeutic strategies. Despite these advances, several challenges remain. Data sparsity, imbalance, and variability across sources hinder model generalizability, while the “black-box” nature of many deep learning architectures raises concerns over interpretability and clinical trust. Ethical, regulatory, and infrastructural barriers particularly regarding data privacy, interoperability, and equitable performance across diverse populations must also be addressed before widespread clinical adoption is feasible. Future directions lie in multimodal and patient-specific data integration, explainable AI frameworks, privacy-preserving approaches such as federated learning, and stronger collaboration between academia, industry, and regulatory bodies. By aligning technological innovation with clinical needs and regulatory safeguards, AI-based predictive modeling can evolve into a reliable, transparent, and globally applicable tool for drug safety assessment. Ultimately, the convergence of AI, precision medicine, and pharmacovigilance has the potential to reshape how DDIs are predicted, prevented, and managed, fostering safer prescribing practices and more effective healthcare delivery.

REFERENCE

1. Percha B, Altman RB. Informatics confronts drug–drug interactions. Trends in pharmacological sciences. 2013 Mar 1;34(3):178-84.
2. Tatonetti NP, Ye PP, Daneshjou R, Altman RB. Data-driven prediction of drug effects and interactions. Science translational medicine. 2012 Mar 14;4(125):125ra31-.
3. Li AP, Jurima-Romet M. Overview: pharmacokinetic drug-drug interactions. Advances in pharmacology. 1997 Jan 1; 43:1-6.
4. Rombolà L, Scuteri D, Marilisa S, Watanabe C, Morrone LA, Bagetta G, Corasaniti MT. Pharmacokinetic interactions between herbal medicines and drugs: their mechanisms and clinical relevance. Life. 2020 Jul 4;10(7):106.
5. Das S, Manjunath K, Biswas A, Sekar K. Therapeutic target identification and evaluation using machine learning. In Therapeutic Protein Targets for Drug Discovery and Clinical Evaluation: Bio-Crystallography and Drug Design 2023 (pp. 73-101).
6. Sudsakorn S, Bahadduri P, Fretland J, Lu C. 2020 FDA drug-drug interaction guidance: a comparison analysis and action plan by pharmaceutical industrial scientists. Current drug metabolism. 2020 May 1;21(6):403-26.
7. Hazell L, Shakir SA. Under-reporting of adverse drug reactions. Drug safety. 2006 May;29(5):385-96.
8. Vilar S, Friedman C, Hripcsak G. Detection of drug–drug interactions through data mining studies using clinical sources, scientific literature and social media. Briefings in bioinformatics. 2018 Sep;19(5):863-77.
9. Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, Chou K, Cui C, Corrado G, Thrun S, Dean J. A guide to deep learning in healthcare. Nature medicine. 2019 Jan;25(1):24-9.
10. Ryu JY, Kim HU, Lee SY. Deep learning improves prediction of drug–drug and drug–food interactions. Proceedings of the national academy of sciences. 2018 May 1;115(18):E4304-11.
11. Zhang W, Chen Y, Liu F, Luo F, Tian G, Li X. Predicting potential drug-drug interactions by integrating chemical, biological, phenotypic and network data. BMC bioinformatics. 2017 Jan 5;18(1):18.
12. Neuvonen PJ, Niemi M, Backman JT. Drug interactions with lipid?lowering drugs: mechanisms and clinical relevance. Clinical Pharmacology & Therapeutics. 2006 Dec;80(6):565-81.
13. Tanaka H, Mizojiri K. Drug-protein binding and blood-brain barrier permeability. The Journal of pharmacology and experimental therapeutics. 1999 Mar 1;288(3):912-8.
14. Guengerich FP. Cytochrome p450 and chemical toxicology. Chemical research in toxicology. 2008 Jan 21;21(1):70-83.
15. Levy G. Pharmacokinetics of salicylate elimination in man. Journal of pharmaceutical sciences. 1965 Jul 1;54(7):959-67.
16. Zanger UM, Schwab M. Cytochrome P450 enzymes in drug metabolism: regulation of gene expression, enzyme activities, and impact of genetic variation. Pharmacology & therapeutics. 2013 Apr 1;138(1):103-41.
17. Chen J, Raymond K. Roles of rifampicin in drug-drug interactions: underlying molecular mechanisms involving the nuclear pregnane X receptor. Annals of clinical microbiology and antimicrobials. 2006 Jan;5(1):3.
18. Zhou S, Chan E, Lim LY, Boelsterli UA, Li SC, Wang J, Zhang Q, Huang M, Xu A. Therapeutic drugs that behave as mechanism-based inhibitors of cytochrome P450 3A4. Current drug metabolism. 2004 Oct 1;5(5):415-42.
19. Rawlins MD, Thompson JW, Davies DM. Textbook of adverse drug reactions. by DM Davies. 1977:95.
20. Katzung BG, Trevor AJ, Pacheco MD. Farmacologia. El Manual Moderno; 1991.
21. Droogan C, Patel C, Yan GX, Kowey PR. Role of antiarrhythmic drugs: frequent implantable cardioverter-defibrillator shocks, risk of proarrhythmia, and new drug therapy. Heart Failure Clinics. 2011 Apr 1;7(2):195-205.
22. Watanabe JH, McInnis T, Hirsch JD. Cost of prescription drug–related morbidity and mortality. Annals of Pharmacotherapy. 2018 Sep;52(9):829-37.
23. Al-Ghazaly ZM, Jassim NA. Polypharmacy and potential drug–drug interactions in patients with rheumatoid arthritis. Medical Journal of Babylon. 2022 Jul 1;19(3):396-403.
24. Trissel LA. Handbook on injectable drugs. InHandbook on Injectable Drugs 2007 (pp. 1720-1720).
25. Newton DW. Drug incompatibility chemistry. American Journal of Health-System Pharmacy. 2009 Feb 15;66(4):348-57.
26. Bourget P, Fernandez H. Incompatibilities and interactions of injectable drugs. Ann Pharmacother. 1999;33(3):347–354. doi:10.1345/aph.17217.
27. Pirmohamed M, James S, Meakin S, et al. Adverse drug reactions as cause of admission to hospital: prospective analysis of 18,820 patients. BMJ. 2004;329(7456):15–19. doi:10.1136/bmj.329.7456.15.
28. Budnitz DS, Lovegrove MC, Shehab N, Richards CL. Emergency hospitalizations for adverse drug events in older Americans. N Engl J Med. 2011;365(21):2002–2012. doi:10.1056/NEJMsa1103053.
29. Nobili A, Garattini S, Mannucci PM. Multiple diseases and polypharmacy in the elderly: challenges for the internist of the third millennium. J Comorb. 2011; 1:28–44. doi:10.15256/joc.2011.1.13.
30. Juurlink DN, Mamdani M, Kopp A, et al. Drug–drug interactions among elderly patients hospitalized for drug toxicity. JAMA. 2003;289(13):1652–1658. doi:10.1001/jama.289.13.1652.
31. Wysowski DK, Nourjah P, Swartz L. Bleeding complications with warfarin use: a prevalent adverse effect resulting in regulatory action. Arch Intern Med. 2007;167(13):1414–1419. doi:10.1001/archinte.167.13.1414.
32. Patel JP, Desai NR, Huang ES. Risk of hospitalization for bleeding with warfarin and NSAID co-administration. J Clin Pharm Ther. 2014;39(6):576–581. doi:10.1111/jcpt.12147.
33. Bellosta S, Paoletti R, Corsini A. Safety of statins: focus on clinical pharmacokinetics and drug interactions. Circulation. 2004;109(23 Suppl 1): III50-57.
34. Maher RL, Hanlon J, Hajjar ER. Clinical consequences of polypharmacy in elderly. Expert Opin Drug Saf. 2014;13(1):57–65. doi:10.1517/14740338.2013.827660.
35. Mangoni AA, Jackson SHD. Age-related changes in pharmacokinetics and pharmacodynamics: basic principles and practical applications. Br J Clin Pharmacol. 2004;57(1):6–14. doi:10.1046/j.1365-2125.2003.02007. x.
36. Spinewine A, Schmader KE, Barber N, et al. Appropriate prescribing in elderly people: how well can it be measured and optimized? Lancet. 2007;370(9582):173–184. doi:10.1016/S0140-6736(07)61091-0.
37. European Medicines Agency. Guideline on the Investigation of Drug Interactions. EMA/CHMP/ICH/572198/2012; 2012.
38. U.S. Food and Drug Administration. Guidance for Industry: Drug Interaction Studies — Study Design, Data Analysis, Implications for Dosing, and Labeling Recommendations. FDA; 2020.
39. Flockhart DA. Drug Interactions: Cytochrome P450 Drug Interaction Table. Indiana University School of Medicine; 2007. Available from: https://drug-interactions.medicine.iu.edu/.
40. Vilar S, Harpaz R, Uriarte E, Santana L, Rabadan R, Friedman C. Drug-drug interaction through molecular structure similarity analysis. J Am Med Inform Assoc. 2012;19(6):1066–1074. doi:10.1136/amiajnl-2012-000942.
41. Ryu JY, Kim HU, Lee SY. Deep learning improves prediction of drug-drug and drug-food interactions. Proc Natl Acad Sci U S A. 2018;115(18):E4304–E4311. doi:10.1073/pnas.1803294115.
42. FDA. Use of Computational Modeling and Simulation in Regulatory Review. Guidance for Industry. FDA; 2021.
43. Lin JH, Lu AYH. Role of pharmacokinetics and metabolism in drug discovery and development. Pharmacol Rev. 1997;49(4):403–449.
44. Houston JB. Utility of in vitro drug metabolism data in predicting in vivo metabolic clearance. Biochem Pharmacol. 1994;47(11):1823–1831.
45. Rendic S, Di Carlo FJ. Human cytochrome P450 enzymes: a status report summarizing their reactions, substrates, inducers, and inhibitors. Drug Metab Rev. 1997;29(1-2):413–580.
46. LeCluyse EL, Witek RP, Andersen ME, Powers MJ. Organotypic liver culture models: meeting current challenges in toxicity testing. Crit Rev Toxicol. 2012;42(6):501–548. doi:10.3109/10408444.2012.682115.
47. Fahmi OA, Kharasch ED, Thummel KE. The cytochrome P450 3A inhibitor ketoconazole selectively inhibits intestinal CYP3A in humans. Clin Pharmacol Ther. 2009;85(6):573-579.
48. Lin JH, Lu AYH. Role of pharmacokinetics and metabolism in drug discovery and development. Pharmacol Rev. 1997;49(4):403–449.
49. Kuepfer L, Niederalt C, Wendl T, et al. Applied concepts in PBPK modeling: how to build a PBPK/PD model. CPT Pharmacometrics Syst Pharmacol. 2016;5(10):516–531. doi:10.1002/psp4.12110.
50. Bailey DG, Dresser GK. Interactions between grapefruit juice and cardiovascular drugs. Am J Cardiovasc Drugs. 2004;4(5):281–297. doi:10.2165/00129784-200404050-00002.
51. Temple R. Clinical drug interaction studies — study design, data analysis, and clinical implications. Clin Pharmacol Ther. 2010;87(6):718–720. doi:10.1038/clpt.2010.64.
52. Vilar S, Harpaz R, Uriarte E, et al. Drug-drug interaction through molecular structure similarity analysis. J Am Med Inform Assoc. 2012;19(6):1066–1074. doi:10.1136/amiajnl-2012-000942.
53. Tatonetti NP, Ye PP, Daneshjou R, Altman RB. Data-driven prediction of drug effects and interactions. Sci Transl Med. 2012;4(125):125ra31. doi:10.1126/scitranslmed.3003377.
54. Preissner SC, Hoffmann MF, Preissner R. Computational approaches to predict drug-drug interactions: a review. Brief Bioinform. 2020;21(6):1802–1817. doi:10.1093/bib/bbz147.
55. Zhang W, Chen Y, Liu F, et al. Predicting potential drug-drug interactions by integrating chemical, biological, phenotypic and network data. BMC Bioinformatics. 2017;18(1):18. doi:10.1186/s12859-016-1422-9.
56. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56. doi:10.1038/s41591-018-0300-7.
57. Chen H, Engkvist O, Wang Y, Olivecrona M, Blaschke T. The rise of deep learning in drug discovery. Drug Discov Today. 2018;23(6):1241–1250. doi: 10.1016/j.drudis.2018.01.039.
58. Zhang W, Chen Y, Liu F, et al. Predicting potential drug-drug interactions by integrating chemical, biological, phenotypic and network data. BMC Bioinformatics. 2017;18(1):18. doi:10.1186/s12859-016-1422-9.
59. Beam AL, Kohane IS. Big data and machine learning in health care. JAMA. 2018;319(13):1317–1318. doi:10.1001/jama.2017.18391.
60. Deo RC. Machine learning in medicine. Circulation. 2015;132(20):1920–1930. doi:10.1161/CIRCULATIONAHA.115.001593.
61. Zhang W, Chen Y, Liu F, et al. Predicting potential drug-drug interactions by integrating chemical, biological, phenotypic and network data. BMC Bioinformatics. 2017;18(1):18. doi:10.1186/s12859-016-1422-9.
62. Liu Y, Wei Q, Yu G, Gai L. An overview of computational methods for drug–drug interaction prediction. Curr Drug Metab. 2021;22(6):462–474. doi:10.2174/1389200222666210323122946.
63. Rajkomar A, Dean J, Kohane I. Machine learning in medicine. N Engl J Med. 2019;380(14):1347–1358. doi:10.1056/NEJMra1814259.
64. Chen H, Engkvist O, Wang Y, Olivecrona M, Blaschke T. The rise of deep learning in drug discovery. Drug Discov Today. 2018;23(6):1241–1250. doi: 10.1016/j.drudis.2018.01.039.
65. Vamathevan J, Clark D, Czodrowski P, et al. Applications of machine learning in drug discovery and development. Nat Rev Drug Discov. 2019;18(6):463–477. doi:10.1038/s41573-019-0024-5.
66. Liu Y, Wei Q, Yu G, Gai L. An overview of computational methods for drug–drug interaction prediction. Curr Drug Metab. 2021;22(6):462–474. doi:10.2174/1389200222666210323122946.
67. Breiman L. Random forests. Mach Learn. 2001;45(1):5-32.
68. Vilar S, Uriarte E, Santana L, et al. Detection of drug–drug interactions through data mining studies using clinical sources, scientific literature and social media. Brief Bioinform. 2018;19(4):693-710.
69. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20(3):273-297.
70. Yamanishi Y, Kotera M, Kanehisa M, Goto S. Drug-target interaction prediction from chemical, genomic and pharmacological data in an integrated framework. Bioinformatics. 2010;26(12):i246-i254.
71. Rogers D, Hahn M. Extended-connectivity fingerprints. J Chem Inf Model. 2010;50(5):742-754.
72. Keiser MJ, Roth BL, Armbruster BN, et al. Relating protein pharmacology by ligand chemistry. Nat Biotechnol. 2007;25(2):197-206.
73. Kuhn M, Letunic I, Jensen LJ, Bork P. The SIDER database of drugs and side effects. Nucleic Acids Res. 2016;44(D1):D1075-D1079.
74. Ching T, Himmelstein DS, Beaulieu-Jones BK, et al. Opportunities and obstacles for deep learning in biology and medicine. J R Soc Interface. 2018;15(141):20170387.
75. Goodfellow I, Bengio Y, Courville A. Deep Learning. MIT Press; 2016.
76. Rifaioglu AS, Atas H, Martin MJ, et al. Recent applications of deep learning and machine intelligence on in silico drug discovery: methods, tools and databases. Brief Bioinform. 2020;21(6):1878-1912.
77. Duvenaud DK, Maclaurin D, Aguilera-Iparraguirre J, et al. Convolutional networks on graphs for learning molecular fingerprints. Adv Neural Inf Process Syst. 2015; 28:2224-2232.
78. Lee I, Keum J, Nam H. DeepConv-DTI: prediction of drug-target interactions via deep learning with convolution on protein sequences. PLoS Comput Biol. 2019;15(6): e1007129.
79. Gilmer J, Schoenholz SS, Riley PF, et al. Neural message passing for quantum chemistry. Proc Mach Learn Res. 2017; 70:1263-1272.
80. Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput. 1997;9(8):1735-1780.
81. Zhang W, Liu F, Zhang C, et al. Predicting drug–drug interactions using a novel graph neural network with feature attention. BMC Bioinformatics. 2021;22(1):1-12.
82. Wang F, Nie F, Huang H, et al. Drug–drug interaction extraction via convolutional neural networks. J Biomed Inform. 2017; 76:125-135.
83. Lee H, Kang J, Hwang S, et al. Predicting drug–drug interactions using deep autoencoders. Front Pharmacol. 2020; 11:273.
84. Wu Z, Pan S, Chen F, et al. A comprehensive survey on graph neural networks. IEEE Trans Neural Netw Learn Syst. 2021;32(1):4-24.
85. Zitnik M, Agrawal M, Leskovec J. Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics. 2018;34(13): i457-i466.
86. Huang K, Xiao C, Glass LM, Sun J. MolRNN: molecular graph neural networks with recurrent neural networks for DDI prediction. J Chem Inf Model. 2021;61(2):647-659.
87. Saito T, Rehmsmeier M. The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS One. 2015;10(3):e0118432. doi: 10.1371/journal.pone.0118432.
88. Chicco D, Jurman G. The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation. BMC Genomics. 2020;21(Suppl 1):6. doi:10.1186/s12864-019-6413-7.
89. Arlot S, Celisse A. A survey of cross-validation procedures for model selection. Stat Surv. 2010; 4:40–79. doi:10.1214/09-SS054.
90. Japkowicz N, Stephen S. The class imbalance problem: A systematic study. Intell Data Anal. 2002;6(5):429–49. doi:10.3233/IDA-2002-6504.
91. Steyerberg EW, Harrell FE Jr. Prediction models need appropriate internal, internal-external, and external validation. J Clin Epidemiol. 2016; 69:245–7. doi: 10.1016/j.jclinepi.2015.04.005.
92. Wishart DS, Feunang YD, Guo AC, Lo EJ, Marcu A, Grant JR, et al. DrugBank 5.0: A major update to the DrugBank database for 2018. Nucleic Acids Res. 2018;46(D1): D1074–82. doi:10.1093/nar/gkx1037.
93. Tatonetti NP, Ye PP, Daneshjou R, Altman RB. Data-driven prediction of drug effects and interactions. Sci Transl Med. 2012;4(125):125ra31. doi:10.1126/scitranslmed.3003377.
94. Kuhn M, Letunic I, Jensen LJ, Bork P. The SIDER database of drugs and side effects. Nucleic Acids Res. 2016;44(D1): D1075–9. doi:10.1093/nar/gkv1075.