Transforming Deviation Management in Biopharmaceuticals: Harnessing Generative AI and Emerging Technologies at MSD

Transforming Deviation Management in Biopharmaceutical Manufacturing with Generative AI

Co-written by Hossein Salami and Jwalant Vyas from MSD

In the biopharmaceutical industry, maintaining drug product quality and ensuring patient safety hinges on diligent deviation management. Each deviation in the manufacturing process is meticulously documented and analyzed to uphold compliance and high standards. For leading pharmaceutical companies, effective deviation management is not just a regulatory requirement; it’s essential for minimizing disruptions and maintaining operational excellence.

Recently, the Digital Manufacturing Data Science team at Merck & Co., Inc. (MSD) identified a pivotal opportunity to enhance their deviation management processes through emerging technologies, including vector databases and generative AI, leveraging AWS services such as Amazon Bedrock and Amazon OpenSearch. This innovative strategy focuses on transforming historical deviation data into a robust, reliable knowledge source. By harnessing insights from past deviations, MSD aims to reduce research time and improve efficiency when addressing new incidents, while adhering to the rigorous standards set by Good Manufacturing Practices (GMP).

Industry Trends: AI in Pharmaceutical Manufacturing

The integration of advanced technologies, particularly AI—especially generative AI—into pharmaceutical manufacturing operations is rapidly gaining traction. From drug discovery to quality control, companies are increasingly exploring how AI can streamline processes that traditionally required significant human expertise and time. The shift towards AI-assisted workflows is not only aimed at improving efficiency but also at enhancing the quality and consistency of critical outcomes, such as deviation management.

Innovative Solution: Generative AI for Deviation Management

To tackle the complexities of deviation management, MSD’s Digital Manufacturing Data Science team crafted a forward-thinking solution utilizing generative AI. Central to this approach is the creation of a comprehensive knowledge base derived from past deviation reports, allowing for intelligent querying that yields valuable insights for new cases.

This knowledge base compiles both structured metadata and unstructured data—such as observations and analysis processes, typically recorded in natural language. Different user personas across manufacturing sites can quickly access relevant information on similar past incidents, facilitating hypotheses about potential root causes and defining resolutions for current cases. This synergy is amplified by a hybrid search mechanism implemented using Amazon OpenSearch Service, which effectively processes and presents information tailored to users’ needs.

Solution Overview: Goals, Risks, and Opportunities

The traditional process of deviation investigations is labor-intensive and prone to human error. Investigation teams often dedicate extensive hours to collecting and analyzing information. MSD aims to achieve several core objectives through this innovative solution:

• Efficiency: Significantly reduce the time and effort required for deviation investigations.
• Accessibility: Provide users easy access to relevant knowledge and historical information, ensuring high accuracy and flexibility based on user roles.
• Traceability: Ensure that the information leading to conclusions is verifiable and traceable.

However, the team remains vigilant about potential risks, such as over-reliance on AI suggestions or outdated information affecting current investigations. To mitigate these risks, the solution limits generative AI content creation to low-risk areas while incorporating necessary human oversight. An automated data pipeline ensures that the knowledge base remains current, and robust security measures protect sensitive information.

Exciting opportunities exist for further enhancements, including agents capable of handling specific user requests, such as high-level statistics and visualizations tailored for site managers.

Technical Architecture: RAG Approach with AWS Services

At the heart of the solution is a Retrieval-Augmented Generation (RAG) architecture designed for efficiency and traceability in deviation investigations. This system seamlessly integrates multiple AWS managed services, creating a scalable, secure, and domain-aware AI-driven framework.

The architecture features a hybrid retrieval module leveraging Amazon OpenSearch Service, which combines semantic (vector-based) and keyword (lexical) search for optimal information retrieval. OpenSearch indexes embeddings from past deviation reports enriched with domain-specific metadata, enabling both deep semantic searches and efficient filtering.

For managing structured data, the Amazon Relational Database Service (Amazon RDS) is employed, supporting complex queries and ensuring compliance and reporting capabilities.

In this architecture, when a user submits a query, relevant documents from OpenSearch, along with structured data from RDS, are retrieved and provided as context to a large language model (LLM) hosted in Amazon Bedrock. This results in contextualized, grounded outputs such as:

• Summarized investigation histories
• Root cause patterns
• Comparable past incidents
• Suggested next steps or knowledge gaps

Conclusion and Next Steps

In conclusion, MSD is leveraging the transformative potential of generative AI and database technologies to optimize deviation management processes. By establishing a multi-faceted knowledge base of past deviations, MSD aims to enhance efficiency while maintaining stringent quality standards.

Looking forward, the company plans to explore additional use cases in the pharmaceutical quality domain, striving to develop a generative AI-driven enterprise-scale product that integrates structured and unstructured data. Enhanced capabilities, such as improved data architecture and advanced retrieval methods, are on the horizon. Utilizing Amazon Bedrock’s capabilities will pave the way for a new standard in deviation management and ultimately enrich manufacturing quality processes across the industry.

About the Authors

• Hossein Salami: Senior Data Scientist at MSD, leveraging over 9 years of R&D experience in chemical engineering to develop AI/ML solutions addressing business challenges.

• Jwalant (JD) Vyas: Digital Product Line Lead at MSD, with over 25 years of biopharmaceutical experience across various domains and a focus on enhancing efficiency in Quality Operations.

• Duverney Tavares: Senior Solutions Architect at AWS, specializing in digital transformation for Life Sciences, with extensive experience in data management and analytics.