Looking Beyond the Hype: Applied AI in Cell and Gene Therapy Manufacturing
AI is a broad umbrella, from rule-based systems through machine learning to generative models. In CGT, the traction is with applied AI that speeds process decisions, tightens control, and shortens iteration cycles. The panel set out to cut through hype and highlight concrete use cases in manufacturing sciences and technologies for cell and gene therapies.
18 Sep 2025
|
Ryan Leahy
One-line take: What works today with AI in CGT manufacturing, where it falls short, and how to move from pilots to practical value.
Focus on applied AI that reduces development time, improves decisions, and actually deploys on shop floors.
Use a simple frame for applied AI: descriptive, predictive, prescriptive.
Pair data-driven models with mechanistic models to extrapolate beyond training sets.
Start with low-risk use cases that keep a human in the loop, then scale.
Invest in process analytical technology and automation to generate the right time-series data.
Be realistic about small data in CGT and design experiments to cover the space efficiently.
Prioritize explainability for regulators and operators, not just accuracy.
Collaborate on standards and ethics so models are useful, safe, and trusted.
What counts as applied AI
Descriptive, predictive, prescriptive: A practical way to map the field. Descriptive AI helps you summarize and visualize data. Predictive AI forecasts outcomes such as growth or quality. Prescriptive AI recommends actions or automates them. In CGT, most deployed tools sit in the descriptive and predictive tiers, with selected prescriptive steps under human oversight.
Real-world use cases that exist today
Bioreactor “digital shadow.” Predictive analytics fuse online sensor streams such as glucose, lactate, perfusion rates, and dissolved oxygen to estimate cell expansion in real time and flag optimal harvest windows. A digital shadow informs operators without closing the loop.
Nuclease design optimization. Machine learning assists the selection and tuning of gene editing components such as TALEN architectures by learning from large design and performance sets.
Process modeling and digital twins. Hybrid models for vector processes such as AAV blend first-principles kinetics with data-driven components to predict outcomes and explore millions of in-silico scenarios during development.
Media and high-dimensional optimization. Teams use AI to handle noisy, multi-factor data in media development, often adapting foundation models from adjacent domains.
Why hybrid beats pure data-driven in CGT
Machine learning interpolates within known space. CGT often demands extrapolation. Mechanistic models grounded in biology, chemistry, and physics extend reach and improve explainability. Hybrid approaches combine the two: known mechanisms where we trust them, data-driven terms where uncertainty is high.
Data: get the right kind, not just more
Small data is the norm. Large homogeneous datasets are rare and expensive. Design of experiments matters.
Continuous signals unlock value. PAT and automation provide high-frequency time-series that power predictive tools.
Measure what matters. Surface markers and easy surrogates do not always map to function. Tie data collection to CQAs and process decisions.
Quality over quantity. Curate inputs, document context, control bias, and validate against reality.
Human in the loop and regulatory comfort
Keep operators in control for higher-risk decisions. Emphasise transparent model performance, versioning, and lifecycle management. Favour interpretable outputs and clear operator messaging. Treat AI like any other validated tool inside GMP systems.
Governance, ethics, and confidentiality
Train teams in safe AI use. Do not paste confidential data into public tools. Align legal, quality, and data security policies with model development and deployment. Build continuous assurance and audit trails.
Collaboration that actually helps
CGT data is costly. Work toward confidential ways to compare methods, share learnings, and standardise data formats without exposing IP. Industry working groups on data standardisation will accelerate usable benchmarks and trust.
)