The use of AI may also find fertile ground in the production of Atmps, a complex and expensive process that requires rigorous quality control. AI-enabled automated systems can detect anomalies in manufacturing processes in real time and offer greater reproducibility by ensuring faster and more robust adjustments to critical process parameters (CPPs), thereby reducing costs and improving efficiency.
Compared to traditional pharmaceuticals, ATMPs exhibit less predictable behaviour and manufacturing processes are less standardised. This makes the application of AI more challenging, but also potentially very innovative. ATMPs operate in complex and dynamic biological systems. The interactions between therapeutic cells and the patient’s microenvironment (such as the immune system and surrounding tissues) are highly variable and difficult to predict. This extreme variability leads to significant heterogeneity, both in terms of genetic and functional characteristics from individual to individual, and in terms of cellular responses to gene modifications, which can vary and lead to different outcomes.
Autologous CAR-T therapies (chimeric antigen receptor T-cell therapies), for example, require manufacturing processes tailored to each patient, making it difficult to apply standardised models. From a manufacturing perspective, processes for traditional drug therapies are generally well established and can be easily standardised at scale, allowing for easier integration of AI for process optimisation and quality control. ATMP, on the other hand, can vary significantly from batch to batch, and this variability requires more sophisticated and adaptable AI models.
The data problem
The limited amount of consistent historical and clinical data compared to the information landscape available for small molecules is another aspect to be assessed. The variability in data quality and the lack of standardised protocols for searching and generating atmps could be an obstacle to the training of AI models, whose efficiency depends precisely on the availability of a very large and standardised dataset. It is precisely the complexity of use in this context that makes the framework somewhat uncertain and unclear.
However, while this inherent variability of ATMPs makes the application of AI difficult, it is also seen as an opportunity. AI is characterised by autonomous learning and is an attractive technology precisely because it can interpret and adapt to highly variable input data. As variability is an intrinsic property of biological components (e.g. cells) and thus of Atmps, AI is particularly attractive for application at all stages of the life cycle of an Atmp, from preclinical development to post-therapeutic monitoring.
Model validation
AI technologies need to be validated in order to be used in a Gmp environment. Currently, regulatory agencies have not provided specific guidance on the validation of AI applications. However, in 2019, the Fda published a ‘Proposed regulatory framework for modifications to artificial intelligence/machine learning (AI/ML)-based software as a medical device (SaMD)’, which was also cited by the QbD Group.
With this initial document, the FDA recognises that AI and ML (machine learning)-based software used as medical devices are constantly evolving through updates and modifications that improve their performance and functionality. However, the document does not specifically mention the integration of AI into the manufacturing or drug development environment.
Rather, the aim is to outline a regulatory framework that allows these changes to be made safely and effectively throughout a well-defined product lifecycle that includes
- Pre-planning of changes
- Continuous performance monitoring
- Proactive risk management
The proposed framework provides considerations for developing AI applications to meet high quality standards:
- Data must be relevant to the intended problem and current practice be collected in a consistent and generalisable way
- There must be a separation between training, tuning and testing data sets
- There must be a sufficient level of clarity about the output and the algorithm