How AI Assisted Bioprocessing Can Transform Biotech

Biotechnology is evolving, but some methods haven’t caught up. Slow experiments, costly materials, and outdated assumptions are holding back innovation. As demand grows for faster, smarter, and more scalable processes, the industry is turning to Artificial Intelligence (AI), modelling, and data-driven strategies to break through the bottlenecks. From regulatory shifts to machine learning breakthroughs, it’s time to rethink how AI assisted bioprocessing can help build the future of biotech. 

Old Methods: Slow, Costly, Limited 

The industry relies on statistical methods to improve biological systems, but traditional approaches like Design of Experiments (DoE) and one-factor-at-a-time experiments are often slow and resource-heavy. Limited data and high material costs make these methods less effective, especially when scaling new modalities. 

To meet growing demands for faster development and scalable processes, both regulators and industry leaders are turning to data-driven decision-making. The FDA’s 2025 draft guidance on AI highlights the need for model transparency and risk awareness in regulated environments. 

Despite its promise, AI and Machine Learning (ML) are often misunderstood. One myth is that they require massive datasets, when in fact, data quality is just as important. Another is that one model fits all, but the best approach depends on the data, process complexity, and specific goals. 

AI in MSAT: Predict, Optimise, Scale 

Manufacturing Science and Technology (MSAT) teams are increasingly using AI and modelling tools to improve bioprocess efficiency. Techniques like ML, Bayesian optimisation, together with empirical models help predict outcomes, reduce lab work, and support scale-up. Their uses include: 

• ML models: Predict DNA-based therapeutic yield using only the sequence, replacing wet lab experiments and modality-based heuristics. 

• Bayesian optimisation (using Gaussian Processes): Identifies optimal restriction digest conditions with far fewer experiments than traditional DoE, saving time and costly reagents. 

• Empirical models: Translate lab observations into predictive tools for scale-up using power-law models for viscosity or concentration and polarisation models for TFF flux prediction. 

Touchlight’s Approach: Flexible, Predictive, Proven 

Touchlight manufactures DNA, from discovery to GMP, to support the development of genetic medicines. One of our challenges is predicting how long each step in the process will take. Since our operations run on a fixed working day, accurate timing is essential for scheduling. 

A critical step of our process involves tangential flow filtration (TFF) using hollow fiber membranes. These membranes are ideal for processing DNA, but the time it takes to run a batch can vary depending on factors like membrane size, dimensions, DNA concentration, and shear rate. 

To improve predictability, we have developed a hybrid model that combines physics with machine learning: 

• Mechanistic Model: The core of the model is based on mass transfer theory, incorporating concentration polarisation effects to describe solute transport through the membrane. This layer captures the fundamental physics governing flux behaviour, influenced by shear rate, membrane dimensions, and solute concentration gradients. 

• Discrepancy Modelling: While mechanistic models offer valuable insight, they may fail to capture all real-world effects.  To overcome this, we implemented a Gaussian Process (GP) model to learn the discrepancy between theoretical predictions and observed data. This discrepancy model captures residual behaviours not accounted for by the mass transfer framework, such as non-ideal flow patterns, membrane fouling, or subtle interactions between operating parameters. 

Advantages Over Traditional Approaches 

• Enhanced Predictive Accuracy: By correcting for model bias, hybrid models outperform purely mechanistic or empirical approaches in real-world scenarios. 
• Generalisation across scales: The mechanistic layer supports extrapolation to new equipment and scales, while the GP model adapts to specific operational contexts. 

• Uncertainty quantification: GP modelling enables probabilistic predictions, supporting risk-aware decision-making and robust scheduling. 

By combining these approaches, we can accurately predict processing times across different scales and equipment setups. This enables more efficient process design, improved planning, and accelerated delivery, ultimately helping genetic medicines reach patients faster.
 

Want to fast-track your project with cell-free DNA?