The FDA has released a discussion paper on the use of artificial intelligence (AI) in the drug manufacturing process, raising a host of questions about issues such as the use of cloud services for manufacturing data management. Perhaps the key consideration posed by the paper is whether AI will enable the broad adoption of advanced drug manufacturing, which the FDA has been trying to encourage for nearly 20 years.
There is a significant contextual consideration for this paper, such as the fact that the FDA’s Center for Drug Evaluation and Research (CDER) established the Emerging Technology Program (ETP) in 2014 for advanced drug manufacturing. Also, the Government Accountability Office (GAO) released a paper in early March which states that few pharmaceuticals are manufactured in the U.S. with advanced manufacturing technologies, in part because of industry concerns regarding other regulatory authorities’ perspectives on advanced manufacturing. These questions apply to biologics as well, as the FDA’s Center for Biologics Evaluation and Research (CBER) is a co-author of the discussion paper.
Manufacturers Wary of Attitudes of FDA Premarket Staff
The GAO report states that the FDA has no benchmarks for determining whether its efforts to encourage the use of advanced manufacturing techniques are having the intended effect on adoption, but also that pharmaceutical companies are uncertain as to whether FDA reviewers will approve these applications. Even when the application is approved, there may be delays while FDA review staff are updated on the use of these technologies for that specific drug product, the GAO report states.
The FDA relies on the International Council for Harmonization (ICH) for a significant portion of its guidance and standards, such as analytical methods validation (as we explained here), and analytical procedures for drug development (which we reviewed here). It is difficult to know the degree to which the FDA seeks to internationally harmonize its regulation of AI for drug manufacturing, but the agency participated in at least one forum on the topic involving the European Medicines Agency (EMA). It seems reasonable to assume that the question of harmonization would influence the timing of any FDA guidance at least as much as the content, a predicament that may be affecting the agency’s approach to AI change control for software as a medical device (SaMD) products as well.
The discussion paper does not offer any specific proposals as to how current good manufacturing practices (cGMPs) would be applied to AI. The broad themes of interest for the FDA are:
- How AI manufacturing controls would interact with cloud-based manufacturing data storage and/or processing;
- The impact of AI on the volume of data generated on a drug manufacturing process, and the difficulty of managing that increased volume of data;
- Which applications of AI in manufacturing and supply chain management would be subject to regulatory oversight;
- The need for standards for developing and validating AI models for manufacturing process control and finished product release testing; and
- The challenges presented by machine learning algorithms, including the issue of algorithm change control.
The question of change control for machine learning software is all too familiar to our clients that develop SaMD products, but it is not clear whether CDER and CBER would follow the approach eventually adopted by the Center for Devices and Radiological Health (CDRH). Whether internal FDA harmonization of change control is a priority at CDER and CBER is impossible to determine, let alone whether this is a higher priority than harmonization with other pharmaceutical regulatory frameworks. This question becomes exceptionally complicated where combination products are concerned, and there are no signals from the FDA on any of these questions.
CDRH’s Collaboration on GMLPs Potentially Useful
The paper poses 8 specific questions to respondents, such as whether guidance would be helpful. In addition, the FDA seeks feedback on:
- The elements that are essential to deploying AI-based models in a pharmaceutical manufacturing environment;
- The mechanisms needed for managing the data used to create AI models in pharmaceutical manufacturing; and
- The best practices for validation and maintenance of self-learning AI software in this environment.
There may already be an answer for the question of best practices for self-learning software in the form of a collaboration on good machine learning practices (GMLPs) between CDRH, Health Canada, and the U.K. Medicines and Healthcare products Regulatory Agency (MHRA). This may be another instance of a resource that CDER and CBER find is not particularly well-suited to their purposes, although some of the principles described in this collaboration would seem to apply readily to AI in drug manufacturing.
While Part 11 for electronic signatures is clearly implicated in this discussion paper, it is likely that the scope of Part 11 considerations will be limited, albeit potentially important. The FDA released a guidance in 2003 on electronic signatures that applies to CBER, CDER, and CDRH, but there may be a need to revise this guidance to account for the effects of AI on the authorship of changes to an electronic record. This may be especially true with the combination of AI and any drug manufacturing data services that are provided via the cloud.
This paper is only the opening in a more extensive examination of the question, but it is not clear whether the agency intends to hold a public meeting for a more in-depth conversation with industry. The FDA is accepting feedback through May 1, 2023, but it may be prudent to view this paper as the start of a process that will run through the rest of 2023 — and well into 2024 — before any guidance is available.
Discussion Paper References Draft Guidances as Resources
In the interim, manufacturers may resort to compliance with existing guidances and standards, including the June 2022 draft guidance for an update to ICH Q9 for quality risk management, and the FDA’s 2016 draft guidance on data integrity and compliance with cGMPs. While the agency has finalized neither of these policies, the discussion paper highlights them as useful resources, even if such a practice would seem to resurrect the ongoing debate over when it is permissible for the FDA to regulate on the basis of a draft guidance.
There are several takeaways from this discussion paper, such as that existing guidance and standards are of only limited help to manufacturers seeking to employ AI in their use of advanced manufacturing technologies. It may also be the case that the FDA will have to sort out several pressing questions about the use of AI in this context before advanced drug manufacturing technologies become both economical and broadly acceptable to regulators. We urge our clients in the AI software business to closely track developments in this space.
