Leveraging Artificial Intelligence in Drug and Biological Product Development: An FDA and Clinical Trial Transformation Initiative Workshop Report

Abstract

Artificial intelligence (AI) holds immense potential to transform drug development by improving the efficiency and accuracy of key processes across the drug product life cycle. However, the scalable adoption of this technology may be influenced by new and unique challenges. The U.S. Food and Drug Administration collaborated with the Clinical Trial Transformation Initiative to organize a public workshop on Artificial Intelligence in Drug and Biological Product Development in August 2024 with medical product sponsors, technology innovators, academicians, and regulators to discuss guiding principles for the use of AI in drug and biological product development in order to realize its transformative potential. This article synthesizes key insights from the workshop and discusses the emerging current need for policy development to enhance the integration of AI in drug and biological product development.

Introduction

Artificial intelligence (AI) has the potential to transform drug and biological product (henceforth “drug”) development and health care delivery by shortening timelines, reducing costs, and increasing success rates at every stage of drug development.¹ As AI technologies rapidly advance and the acceptance of AI methods grows, more organizations are leveraging AI tools throughout the drug product life cycle, from target identification and lead optimization to enhancing pharmacovigilance efforts.^2,3

Since 2016, the U.S. Food and Drug Administration’s (FDA) Center for Drug Evaluation and Research has received an increasing number of regulatory submissions with AI components. Most of these AI approaches were applied in the clinical research phase, including clinical trial design, end points, including clinical outcome assessments and biomarkers, real-world evidence generation, and postmarket safety surveillance.⁴ The FDA has undertaken a series of initiatives to share knowledge and provide regulatory clarity and predictability, including public workshops, listening sessions, educational materials, discussion papers,³ and the issuance of draft guidance documents.

The Clinical Trial Transformation Initiative (CTTI) is a public–private partnership cofounded by Duke University and the FDA. CTTI collaborates with more than 500 organizations, spanning academia, clinical investigators, industry leaders, patient advocates, and regulators to facilitate the modernization of clinical trials. The FDA, leveraging CTTI’s multisector collaboration expertise, convened a public workshop on Artificial Intelligence in Drug and Biological Product Development in August 2024.⁵ The purpose of the public workshop was to bring drug developers and AI experts together to discuss guiding principles for the responsible use of AI in the development of safe and effective biological products.

Speakers identified four main hurdles to large-scale AI adoption: (1) limited data access, (2) low confidence in AI model output, (3) lack of collaboration across siloed disciplines, and (4) regulatory complexities (Fig. 1).

Figure 1. Challenges Associated with the Integration of AI Technology in Drug Development.

AI denotes artificial intelligence.

Key Challenges

The following sections outline key challenges and highlight shared lessons to inform global policy development for emerging applications of AI technology in drug development. Although challenges are grouped by topic for clarity, several issues span multiple categories and are interconnected. Successful AI implementation, in spite of the current challenges, includes analyzing genomic databases for clinical decisions, optimizing trial site selection, automating image analysis for disease scoring, enabling remote patient monitoring, accelerating literature reviews, and enhancing pharmacovigilance through automated adverse event detection and classification systems (Table S1 in the Supplementary Appendix).

ENABLING ACCESS TO FIT-FOR-USE DATA AND IMPROVING PROSPECTIVE DATA COLLECTION

The availability of high-quality, fit-for-use data is crucial for AI models to produce valuable output. In this context, “fit for use” means that the data used to develop and test AI models should be both relevant and reliable to the specific context of use.⁶ Importantly, the application of fit-for-use data as a concept is not absolute. Instead, it emphasizes that data quality should be assessed relative to the specific purpose it is meant to serve (i.e., its context of use).

Fit-for-use data exist across the drug development ecosystem but often remain siloed within academic institutions, industry sponsors, health care systems, payers, and regulatory agencies. Access to these data is further constrained by regulatory and privacy requirements (e.g., the Health Insurance Portability and Accountability Act in the United States and the General Data Protection Regulation in the European Union), as well as by complex contractual frameworks, inconsistent data formats, and the labor-intensive processes required to clean and harmonize data from multiple sources. For rare diseases, sufficient historical data for model development are often unavailable or nonexistent.

Fragmented data assets constrain the full potential of AI, highlighting the need for collaborative data platforms that support secure data sharing. Such frameworks should necessarily protect patient privacy while safeguarding intellectual property and proprietary information. For example, federated computing platforms, which enable analysis across distributed data sources without requiring data centralization, may enhance data accessibility while respecting localization requirements across institutions and jurisdictions.

To address challenges related to data quality, speakers emphasized several key strategies, such as ensuring data reliability, adopting standardized data models and vocabularies (e.g., Observational Medical Outcomes Partnership Common Data Models, PCORnet Common Data Model), and developing frameworks to assess the quality of real-world data for AI model development. Other suggested approaches included using synthetic data to fill gaps — while acknowledging its limitations — and validating data through independent external partners not involved in initial development. Strong data governance was also highlighted, especially the importance of maintaining separation between training and testing datasets to prevent data leakage and reduce the potential risk of model overfitting.

Selection, confirmation, and measurement biases introduce systematic errors into the data used to develop AI models, which can compromise their performance, validity, and generalizability — especially when applied to diverse patient populations and real-world settings. To manage bias in data, workshop participants recommended collecting extensive, representative data across all relevant areas; testing models across various demographics to find weaknesses; implementing human oversight and review, particularly for high-risk applications; and documenting limitations in the data used for developing these models. Developers should incorporate bias mitigation strategies during initial model development and conduct ongoing monitoring to detect any unexpected biases.

BUILDING TRUST THROUGH INTERPRETABILITY, EXPLAINABILITY, AND TRANSPARENCY

Another key challenge in AI adoption for drug development is the “black box” nature of many models, where understanding the fundamentals of how they operate and why they produce specific outputs is difficult. Building collective trust in AI models is essential for widespread adoption, and this trust can be strengthened through interpretability, explainability, and transparency.

In the context of AI, interpretability and explainability are related but separate concepts. Interpretability is generally understood as how well a human can understand the internal workings of an AI system (i.e., how the AI functions, based on the clarity of the relationship between input data and output data). Explainability, on the other hand, is typically seen as how well a human can understand the reasons behind an AI system’s decisions or predictions (i.e., why the AI made a specific prediction or decision).⁷ Low levels of interpretability and explainability can potentially hinder the ability to identify and rectify errors and biases effectively, creating challenges for regulatory agencies when evaluating AI technologies in drug development.

Workshop participants suggested various strategies to overcome the limitations of interpretability and explainability. These include creating models with explainable AI techniques such as Shapley Additive Explanations values and Local Interpretable Model-Agnostic Explanations, using parameters that represent real-world concepts, adopting human-in-the-loop methods, verifying results with independent measurements or existing data, performing thorough testing and sensitivity analyses, and combining AI with traditional modeling techniques. Participants also highlighted the need to ensure that end users can understand and effectively utilize AI outputs. Full interpretability may not always be feasible, particularly for complex models; however, it is essential to validate outputs, maintain human oversight, and foster trust in the credibility of AI model outputs through transparency.

Transparency is a broad concept that includes interpretability, explainability, metadata, documentation, and disclosures beyond just technical understanding. When AI technologies lack transparency, it diminishes trust and confidence among stakeholders such as sponsors, researchers, regulators, and patients. Even if the inner workings of complex AI models cannot be fully explained, clear communication about the methods used is essential to build trust and enable effective use of AI outputs.

DEVELOPING A RISK-BASED REGULATORY APPROACH FOR AI USE IN DRUG DEVELOPMENT

Workshop participants recommended that regulators adopt a flexible, risk-based approach to AI oversight — one that is proportionate to the level of risk posed by the AI application and tailored to its specific context of use. It is important to clearly communicate regulatory expectations by publishing high-level principles and guidance, supported by supplementary materials such as frequently asked questions, white papers, and illustrative examples to promote consistent implementation. Addressing risk aversion during the deployment of emerging technologies can be achieved by sharing case studies of AI integration, starting with small, low-risk use cases to demonstrate risk-based approaches.

Furthermore, international convergence and harmonization of AI regulations and terminologies would support the global development and deployment of AI technologies. Consistent data privacy and sharing rules could overcome challenges related to accessing suitable data needed to train AI models, especially in rare disease areas or with under-represented populations.

FOSTERING MULTIDISCIPLINARY COLLABORATIONS

The integration of AI technologies in drug development faces challenges because of the siloed nature of the disciplines involved in developing, implementing, and evaluating AI tools. Different scientific fields often have varying interpretations of key concepts, creating a terminology barrier that hinders communication and collaboration. To unlock AI’s full potential in health promotion, several key actions are needed: encouraging multidisciplinary collaboration by involving experts from different fields in developing AI tools, establishing standardized terminology through joint efforts, increasing AI literacy across the drug development community, and ensuring clear communication tailored to specific audiences. Although collaboration across disciplines is important, true “multifluency” (i.e., “a person’s ability to operate across multiple fields,” as one participant described it) is even more essential. This multifluency can be developed through training in related disciplines, promoting interdisciplinary learning, and offering incentives for cross-domain expertise. By adopting these strategies, organizations can integrate knowledge across disciplines, lower barriers to AI adoption, and build the expertise necessary to effectively leverage AI in drug development.

FUTURE OPPORTUNITIES TO ADVANCE THE FIELD

By convening a wide range of stakeholders from academia, industry, regulatory agencies, and other sectors, the workshop emphasized both the transformative potential of AI in drug development and the critical gaps that need to be addressed for responsible and scalable implementation. Key policy areas identified for further development included (1) establishing clear guidance for assessing and managing risks associated with AI across various stages of drug development; (2) defining standards for data quality and evaluating the fitness for use of training and testing data for AI model development, including guidance on the use of synthetic data; (3) out-lining expectations for transparency, explainability, and interpretability — especially for high-risk applications; (4) developing common terminology and glossaries for AI that span scientific disciplines; (5) encouraging public–private partnerships to enable responsible data sharing and collaborative innovation; and (6) creating policies and incentives to develop multidisciplinary expertise and multifluency across clinical research, data science, and regulatory fields. All parties involved have a crucial role in this effort (Fig. 2, Table S2).

Figure 2. Recommended Next Steps for the Interested Parties to Successfully Integrate AI in Drug Development.

AI denotes artificial intelligence.

Conclusion

This workshop, along with other efforts, including the discussion paper published in 2023³ and revised in 2025, contributed to the development of the FDA’s draft AI guidance titled ‘Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products,’ which was published in January 2025. The draft guidance offers recommendations for industry and other stakeholders on using AI to generate data or information intended to support regulatory decisions regarding drug safety, effectiveness, or quality. At its core, the guidance features a seven-step, risk-based credibility assessment framework as a resource to support evaluating an AI model’s reliability for its specific context of use.

The FDA continues to prioritize a strong and predictable regulatory infrastructure that advances the field responsibly. Across the agency, teams are focusing on external engagements to support awareness and mutual learning; anticipating policy needs where users may need more regular and robust input from the agency on utilizing AI; exploring pilot projects to further inform policy and review; upskilling staff to ensure awareness; and engaging globally to explore opportunities for harmonizing terminology and basic principles. The FDA is committed to harnessing innovation in the development of safe and effective medical products for all.