FDA Releases Discussion Paper for AI in Drug Development

June 6, 2023

Earlier this year, the FDA released a discussion paper on the use of artificial intelligence in the manufacture of pharmaceuticals and biotech therapies, but has added another discussion paper for AI in drug development. The two papers will help the FDA develop guidance for these two uses of AI. While there is no plausible expectation that guidance is going to emerge anytime this year, indications are that such a software development project may require deep pockets.

As we discussed in March, the FDA’s paper on drug manufacturing lists several essential considerations, such as the regulatory complications associated with the use of cloud-based compliance systems. This new paper on AI and machine learning (ML) in drug development includes the use of any software that is regulated as a medical device for the drug development process.

Industry Ahead of FDA

There is a substantial body of work already in place in AI/ML in drug development, such as the FDA grant of an orphan drug designation for a product for idiopathic pulmonary fibrosis. According to one analysis, more than 15 drug products discovered with the help of AI/ML were already in clinical trials as of March 2022, a number that has surely grown significantly in the 15 months since.

There are three key regulatory considerations for the FDA as explained by the paper, which are:

  • Human-led governance, accountability, and transparency;
  • The quality, reliability, and representativeness of the data; and
  • Model development, performance, monitoring, and validation.

The discussion paper proposes that a risk management approach be applied to the first of these three areas to identify and mitigate risks. The FDA is also seeking feedback on what transparency means in this context, which may include disclosure of information about the performance of the algorithm to regulators and other stakeholders.

A Simpler Algorithm is Possibly the Better Algorithm

Not unexpectedly, the discussion of data quality includes a review of the hazards of unintended bias for which there are three categories. These are human, systemic, and statistical/computational bias, but the paper also asks stakeholders to consider methods for managing missing data in the discussion of this second point.

As for model development/performance/validation, the paper points to the fact that the risk associated with a given context of use and the consequence of the decision will be influenced by the model. A simpler model may be preferrable to a more complex model to help de-risk the use of an algorithm so long as the simpler model produces the same results as the more complex model. The FDA indicates that the developer will have to track the performance of the algorithm over time and document that performance with an eye toward potential corrective action should the performance of the algorithm drift out of specification.

Among the key potential uses of AI/ML in drug development are drug target identification, clinical research, and post-market safety surveillance. There is also a section in the FDA discussion paper on advanced pharmaceutical manufacturing as well as the use of digital health technologies (DHTs) in the drug development process, something we delved into recently.

Deep Learning May be Required to Manage Large Data Sets

Drug target identification is one of the more demanding potential uses of AI and ML in drug discovery, as it requires the mining of large data sets of genomic, proteomic, and transcriptomic information to aid in the process of selecting a target. The data sets required to fulfill this kind of approach may be quite complex and the developer of an AI/ML algorithm may have to draw on a variety of data sources. The process of developing software to perform such a task will obviously not be easy, and a conventional ML algorithm might not be sufficiently capable.

As the authors of a January 2021 journal article explained, these data sets may include millions of compounds, and a conventional ML computational model based on quantitative structure-activity relationships (QSARs) might not provide the computational lift needed to efficiently eliminate implausible drug targets. Drug manufacturers must be able to predict drug product characteristics such as absorption, metabolism, and toxicity in an efficient manner. QSAR-driven software can handle these tasks if aided by a deep learning capability, which will be crucial if the process of vetting potential drug candidates is to be reasonably efficient.

Generative Adversarial Networks Gaining in Popularity

Another obvious use of AI/ML is in drug repurposing, which the FDA stated can leverage real-world evidence (RWE) such as data drawn from registries, electronic health records, and DHTs. Even the de novo design of a drug is a potential use of AI/ML, but the FDA discussion paper points to several major hurdles, such as the need to ensure that the algorithm will yield a properly optimized 3-D structure. However, this may require incorporation of mutations that are not necessarily present in all widely available genetic data sets.

The FDA announced this latest discussion paper May 10 with the observation that cybersecurity is another point of interest in this AI-driven drug development world. Drug developers are increasingly leaning toward the use of advanced software products such as generative adversarial networks (GANs) in drug discovery and design, but as a recent paper in the Journal of Chemical Information and Modeling suggests, the computer hardware required to run this type of software will not be inexpensive to access, given that quantum computational power may be necessary. Developers who are considering this field will have to evaluate the resources required to fully develop and maintain the algorithm, as this type of software promises to be one of the most expensive software development projects most developers are likely to face.

Wondering which QMS is right for you?
This is default text for notification bar