Why Drug-Drug Interactions Escapes Preauthorization Testing and Traditional Pharmacovigilance (And How AI Can Help)

Drug-drug interactions (DDIs) represent a significant challenge in modern healthcare, often eluding detection during preauthorization testing and traditional pharmacovigilance methods. As a follow-up to my previous article on AI’s role in pharmacovigilance, this article aims to elucidate the complex factors contributing to this phenomenon. This comprehensive analysis will explore the multifaceted nature of DDIs, the limitations of current detection methods, and the implications for patient safety and clinical practice.

Combinatorial Complexity

The increasing number of approved drugs and the prevalence of polypharmacy have exponentially increased the potential for DDIs. This combinatorial complexity poses a significant challenge to recognizing and understanding these interactions.¹

As Roughead² points out, the addition of multiple drugs to a patient’s regimen over time can lead to synergistic interactions that may be mistakenly attributed to aging or intercurrent illness. This misattribution can obscure the true nature of the DDI, making it more difficult to detect and manage.³

Furthermore, the complexity is not limited to drug-drug interactions alone. Drug-drug-disease and drug-drug-gene interactions amplify the combinatorial possibilities.^4,5,6 For instance, consider a scenario where a patient with chronic renal failure is administered a drug primarily eliminated via one major and one alternative backup cytochrome P450 metabolic pathway. The backup pathway will typically compensate if a second drug that inhibits the primary pathway is added. However, in the presence of chronic renal failure, the expression of the secondary cytochrome may be inhibited, reducing its compensatory effect.⁷ This interplay between drugs, diseases, and genes illustrates the challenges in predicting and identifying DDIs.

Mechanistic Complexity

The mechanisms underlying DDIs can be highly complex, often involving multiple intermediate steps that obscure the direct linkage between drugs. This complexity can make interactions less recognizable and more challenging to anticipate.

A prime example of this mechanistic complexity is the reported interaction between etanercept and cyclosporine.⁸ In this case, etanercept reduces inflammation, leading to decreased cytokines and cytokine-induced cytochrome P450 3A4 inhibition. This cascade of events ultimately results in increased metabolism of cyclosporine and reduced trough levels. Notably, the therapeutic effect of etanercept serves as the starting point for this interaction, adding another layer of complexity to its detection and understanding.

As a researcher, I find this example particularly fascinating as it highlights the intricate web of physiological processes that can be affected by drug interactions. It underscores the need for a systems-level approach to understanding and predicting DDIs, moving beyond simple one-to-one drug interactions.

Limited Data and Testing

One of the most significant challenges in identifying DDIs is the limited scope of preauthorization testing and the underrepresentation of certain populations and substances in these studies. Gabay and Spencer⁹ highlight several categories that are often overlooked in DDI testing, including:

• Children
• Individuals with variant drug metabolizing or transporting genotypes
• Herbal supplements
• Foods
• Alcohol
• Tobacco
• Cannabis

This lack of comprehensive testing creates significant gaps in our knowledge base, making it difficult to anticipate potential DDIs in these populations or with these substances. Furthermore, the inherent limitations of small studies, such as power limitations, further restrict our ability to detect less common or subtle interactions.

I believe this underscores the need for more inclusive and diverse clinical trials, as well as the importance of post-marketing surveillance to capture DDIs that may have been missed during preauthorization testing.

Relative Paucity of Guidance

While there are numerous guidelines and paradigms for detecting pharmacokinetic DDIs during drug development, similar frameworks are lacking for pharmacodynamic DDIs. This disparity in guidance can lead to an underestimation of the potential for pharmacodynamic interactions, which can be equally significant in clinical practice.

In my opinion, this gap in guidance represents a critical area for future research and policy development. Establishing comprehensive frameworks for assessing both pharmacokinetic and pharmacodynamic DDIs could significantly enhance our ability to predict and manage these interactions.

Under-attention to Potential Beneficial DDIs

An often overlooked aspect of DDIs is the potential for beneficial interactions. Traditional pharmacovigilance tends to focus primarily on adverse effects, potentially missing signals of beneficial interactions.

Kuss and Rathmann¹⁰ provide an intriguing example of this phenomenon, discussing a missed protective drug-drug interaction between DPP-4 inhibitors and statins on myopathy risk. This case highlights the importance of considering adverse and beneficial interactions in our DDI research approach and pharmacovigilance approach.

From what I have gleaned, I believe this represents an exciting area for future investigation. By expanding our focus to include potential beneficial interactions, we may uncover new therapeutic strategies and optimize existing treatment regimens.

Unknown Mechanisms

Despite our increasing understanding of the molecular, cellular, and systemic aspects of DDIs, the mechanisms of some interactions remain unknown. This lack of mechanistic understanding decreases the likelihood of early detection and complicates management strategies.

According to a recent analysis by Xiong et al.¹¹, approximately 17.7% of DDIs in a comprehensive database are classified as having unknown mechanisms. This significant proportion of unexplained interactions underscores the ongoing challenges in this field and highlights the need for continued research into the fundamental mechanisms underlying DDIs.

AI Solutions to DDI Challenges

Artificial intelligence (AI) and machine learning (ML) offer promising solutions to many challenges discussed above. Here are some potential AI-driven approaches to address each of the identified problems:

1. Combinatorial Complexity:

AI can efficiently handle the vast combinatorial possibilities of DDIs. Machine learning models, particularly deep learning networks, can process large datasets of drug combinations, patient characteristics, and interaction outcomes to identify patterns that may elude human researchers.

For example, Deng et al.¹² developed a graph neural network model that can predict DDIs with high accuracy, even for drug combinations not seen during training. This approach could help identify potential interactions among newly approved drugs or in complex polypharmacy scenarios.

2. Mechanistic Complexity:

AI can assist in unraveling the complex mechanisms underlying DDIs. Natural language processing (NLP) techniques can extract relevant information from scientific literature and electronic health records to build comprehensive knowledge graphs of drug interactions and their mechanisms.

Burkhardt et al.¹³ demonstrated the use of NLP and knowledge graph embedding to predict DDI mechanisms. Their model could identify both direct and indirect interactions, potentially capturing complex, multi-step DDIs like the etanercept-cyclosporine example mentioned earlier.

3. Limited Data/Testing:

AI can help address the limitations of current DDI testing by:

• Augmenting existing data: Generative models can create synthetic data to supplement limited real-world data, especially for underrepresented populations or substances.
• Transfer learning: Models trained on adult data can be fine-tuned for pediatric populations, potentially addressing the lack of DDI data in children.
• In silico testing: AI models can simulate DDIs in various scenarios, including different genetic profiles or disease states, reducing the need for extensive human trials.

Vilar et al. ¹⁴ developed a large-scale DDI prediction model using machine learning that could predict interactions for drugs not included in their training data, demonstrating the potential for AI to address data limitations.

4. Relative Paucity of Guidance:

AI can assist in developing more comprehensive guidelines for DDI detection, including both pharmacokinetic and pharmacodynamic interactions. Machine learning models can analyze vast amounts of clinical data to identify patterns and factors contributing to various types of DDIs.

For instance, Zheng et al.¹⁵ proposed a multi-modal deep learning approach that integrates molecular structure, drug-target interactions, and known DDIs to predict new interactions. Such comprehensive models could inform the development of more holistic DDI guidelines.

5. Under-attention to Potential Beneficial DDIs:

AI models can be designed to identify both adverse and beneficial DDIs. By training on diverse outcomes, including positive therapeutic effects, these models can help researchers and clinicians consider the full spectrum of potential interactions.

Gottlieb et al.¹⁶ developed a machine-learning approach that predicts both adverse and synergistic drug combinations, demonstrating the potential for AI to uncover beneficial interactions that might be overlooked by traditional methods.

6. Unknown Mechanisms:

AI can help elucidate unknown DDI mechanisms through various approaches:

a) Hypothesis generation: AI models can analyze patterns in known DDIs to suggest potential mechanisms for unexplained interactions.

b) Molecular docking simulations: AI-enhanced molecular modeling can simulate drug-drug interactions at the molecular level, providing insights into potential mechanisms.

c) Multi-omics integration: AI can integrate data from genomics, proteomics, and metabolomics to provide a systems-level view of DDI mechanisms.

Huang et al. ¹⁷ developed a deep learning model that predicts DDI mechanisms based on molecular structure and known interaction data. This approach could help researchers generate hypotheses about unknown mechanisms for further investigation.

Conclusion

The challenges in detecting and understanding drug-drug interactions are complex. Numerous factors contribute to the elusiveness of DDIs in preauthorization testing and traditional pharmacovigilance.

As researchers and clinicians, we must continue to advance our understanding of these interactions through comprehensive studies, improved testing methodologies, and enhanced surveillance systems. Furthermore, we should strive to develop more inclusive guidelines that address both pharmacokinetic and pharmacodynamic interactions while also considering potential beneficial DDIs.

While AI offers promising solutions to many challenges in DDI research and detection, it’s important to note that these approaches should complement, not replace, traditional pharmacological research and clinical judgment. The integration of AI-driven insights with expert knowledge and rigorous clinical testing can significantly enhance our ability to predict, understand, and manage drug-drug interactions, ultimately improving patient safety and therapeutic outcomes.

References:

1. Bode C. The nasty surprise of a complex drug-drug interaction. Drug Discov Today. 2010;15:391-395. ↩︎
2. Roughead E. Multidrug interactions: the current clinical and pharmacovigilance challenge. J Pharm Pract Res. 2015;45:138-139. ↩︎
3. Roberts AG, Gibbs ME. Mechanisms and the clinical relevance of complex drug-drug interactions. Clin Pharmacol. 2018;10:123-134. ↩︎
4. Storelli F, Samer C, Reny JL, et al. Complex drug–drug–gene–disease interactions involving cytochromes P450: systematic review of published case reports and clinical perspectives. Clin Pharmacokinet. 2018;57:1267-1293. ↩︎
5. Malki MA, Pearson ER. Drug-drug-gene interactions and adverse drug reactions. Pharmacogenomics J. 2020;20:355-366. ↩︎
6. Bruckmueller H, Cascorbi I. Drug-drug-gene interactions: a call for clinical consideration. Clin Pharmacol Ther. 2021;110:549-551. ↩︎
7. Verbeurgt P, Mamiya T, Oesterheld J. How common are drug and gene interactions? Prevalence in a sample of 1143 patients with CYP2C9, CYP2C19 and CYP2D6 genotyping. Pharmacogenomics. 2014;15:655-665. ↩︎
8. Wen H, Chen D, Lu J, et al. Probable drug interaction between etanercept and cyclosporine resulting in clinically unexpected low trough concentrations: first case report. Front Pharmacol. 2020;11:939. ↩︎
9. Gabay M, Spencer SH. Drug Interactions: Scientific and Clinical Principles. Pharmacotherapy Self-Assessment Program (PSAP) Book 3. Chronic Conditions and Public Health: American College of Clinical Pharmacy. 2021:7-28. ↩︎
10. Kuss O, Rathmann W. A missed protective drug-drug interaction of DPP-4 inhibitors and statins on myopathy risk. Acta Diabetol. 2020;57:113-114. ↩︎
11. Xiong G, Yang Z, Yi J, et al. DDInter: an online drug-drug interaction database towards improving clinical decision-making and patient safety. Nucleic Acid Res. 2022:50. ↩︎
12. Deng Y, Xu X, Qiu Y, et al. A multimodal deep learning framework for predicting drug–drug interaction events. Bioinformatics. 2020;36(15):4316-4322. ↩︎
13. Burkhardt HA, Subramanian D, Mower J, Cohen T. Predicting adverse drug-drug interactions with neural embedding of semantic predications. J Biomed Inform. 2019;96:103252. ↩︎
14. Vilar S, Friedman C, Hripcsak G. Detection of drug-drug interactions through data mining studies using clinical sources, scientific literature and social media. Brief Bioinform. 2018;19(5):863-877. ↩︎
15. Zheng Y, Peng H, Zhang X, et al. DDI-PULearn: a positive-unlabeled learning method for large-scale prediction of drug-drug interactions. BMC Bioinformatics. 2019;20(Suppl 19):661. ↩︎
16. Gottlieb A, Stein GY, Oron Y, Ruppin E, Sharan R. INDI: a computational framework for inferring drug interactions and their associated recommendations. Mol Syst Biol. 2012;8:592. ↩︎
17. Huang K, Xiao C, Glass LM, Sun J. MolTrans: Molecular Interaction Transformer for drug-target interaction prediction. Bioinformatics. 2021;37(6):830-836. ↩︎