Artificial Intelligence (AI) has long been associated with science fiction movies about dystopian futures, leading to fear among the general public about its potential impact. This is especially the case today for those in academia who have graded countless papers written by ChatGPT. However, the truth is far from what we see in the movies. In fact, one industry where AI is making significant progress is the biosimilar industry. AI offers many possibilities, including optimizing process design and process control, smart monitoring and maintenance, and trend monitoring to drive continuous improvement. Recently, the FDA has participated in discussions around AI and biotechnology.

The FDA has already played an important role in the integration of AI in the biotechnology field. It has authorized more than 500 AI/ML-enabled medical devices, but last month, the FDA made two big contributions to the conversation. The first is its publication of a discussion paper on artificial intelligence in drug manufacturing to help proactively prepare for the implementation of AI in the field.[1] The second is an article the FDA published disclosing the implementation of AI-based modeling to analyze protein aggregation in therapeutic protein drugs.[2]

1. FDA Discussion Paper – Artificial Intelligence in Drug Manufacturing

In its discussion paper, the FDA requests public feedback to help inform its evaluation of the existing regulatory framework involving AI in drug manufacturing. The FDA suggests a number of areas for consideration.

One such area is standards for developing and validating AI models. The FDA admits that there are limited industry standards and FDA guidance available for the development and validation of models that impact product quality. The lack of guidance is a concern since AI has such great applicability during drug manufacturing. AI can be used in applications to control manufacturing processes by adapting process parameters based on real-time data, or in conjunction with interrogation of in-process material or the final product to: (1) support analytical procedures for in-process or final product testing, (2) support real-time release testing, or (3) predict in-process product quality attributes.

The FDA also notes the challenge applicants have in defining standards that validate an AI-based model and sustaining the ability to explain the model’s output and impact on product quality. As AI methods become more complex, it becomes more challenging to explain how changes in model inputs impact model outputs.

Another area for consideration is how continuously learning AI systems that adapt to real-time data may challenge regulatory assessment and oversight. AI models can evolve over time as new information becomes available. The FDA states that it may be challenging to determine when such an AI model can be considered an established condition of a process. It also may be challenging to determine the criteria for regulatory notification of changes to these models as a part of model maintenance over the product lifecycle. Applicants may need clarity on: (a) the expectations for verification of model lifecycle strategy, and (b) expectations for establishing product comparability after changes to manufacturing conditions introduced by the AI model.

Comments on these and other issues can be sent to the FDA at the link below.[3]

2. FDA’s AI/Machine Learning Modeling to Ensure Safety and Demonstrate Biosimilarity

Despite the limited guidance the FDA has for AI-based technologies, it recently published a study utilizing AI for characterizing protein aggregation, which will provide a more effective means of demonstrating biosimilarity and improve safety in therapeutic protein drugs.

One major challenge that biosimilar developers face with therapeutic protein drugs is characterizing these products in order to compare them with a reference product. Characterization is particularly an issue because of protein aggregates that can create subvisible particles with a wide variety of sizes, shapes, and compositions from a variety of stress conditions. Although a small fraction of the total protein, these aggregates may increase the risk of undesirable immune responses.

The FDA’s study characterized aggregate protein particles using flow imaging microscopy (FIM). This imaging technique can record multiple images of a single subvisible particle from a single sample. Although these image sets are rich in structural information, manual extraction of this information is cumbersome and often subject to human error, meaning that most of the information is underutilized.

To overcome the shortcomings of current optical image analysis, the FDA applied convolutional neural networks (CNNs), a class of artificial neural networks proven helpful in many areas of image recognition and classification. This AI/ML approach enables automatic extraction of data-driven features (i.e., measurable characteristics or properties) encoded in images. These complex features (e.g., fingerprints specific to stressed proteins) can potentially be used to monitor the morphological features of particles in biotherapeutics, and enable tracking the consistency of particles in a drug product.

CNNs can be trained with input data using supervised learning or a fingerprinting approach. For supervised learning, the AI model is trained using estimations of the most discriminatory parameters defined using images that are correctly labelled as either stressed or unstressed. Once trained, the CNN can predict which pre-defined labels best apply to a new image. The fingerprinting approach, on the other hand, is optimized to reduce the dimension of the spatially correlated image pixel intensities, resulting in a new lower dimensional (e.g., 2D) representation of each image. These lower dimensional representations can be used to analyze complex morphology encoded in a heterogeneous collection of FIM images since the full images can readily be mapped to a lower dimensional representation by the CNN.

The FDA found that flow microscopy combined with CNN image analysis could be applied to a range of products and will provide potential new strategies for monitoring product quality attributes. Such technology will enable processing of large collections of images with high efficiency and accuracy by distinguishing complex “textural features” which are not readily delineated with existing image processing software.

*            *            *

As AI becomes more advanced and more of those in the biosimilar industry utilize this technology, the more guidance the FDA will have to provide, and the sooner the better. These two contributions from the FDA indicate that it is well aware of this need and is even looking to promote AI’s use across the pharmaceutical and biopharmaceutical fields.

[1] https://www.fda.gov/media/165743/download

[2] https://www.fda.gov/drugs/regulatory-science-action/artificial-intelligencemachine-learning-assisted-image-analysis-characterizing-biotherapeutics?utm_medium=email&utm_source=govdelivery

[3] https://www.regulations.gov/  Docket No. FDA-2023-N-0487