Ready to unlock your cells’ potential?
No other teamis as dedicated to advancing cell culture media as Multus. Talk to our mediaspecialists today!
No other teamis as dedicated to advancing cell culture media as Multus. Talk to our mediaspecialists today!
.avif)
AI has arrived in biomanufacturing with a wave of claims and very little explanation. It’s common for companies to say they’re using AI, or that their process is AI-enabled. What is less common is a clear account of why applying AI to biomanufacturing is extremely difficult.
Most people working in bioprocess understand that media formulation is complex. Fewer appreciate the specific nature of the challenge, or why solving it requires more than a good algorithm. There are three distinct problems. Each is hard, and solving any one in isolation is not enough.
A typical cell culture media formulation contains 60 or more interacting ingredients, each at a concentration that affects every other. Testing all possible combinations at just three concentration levels per ingredient produces more potential formulations than there are grains of sand on the Earth.
This explains why conventional Design of Experiments (DoE) approaches hit a ceiling at around 10 to 12 variables. Beyond that threshold, the statistical coverage of the formulation space becomes so thin that the results are essentially meaningless. You’re not exploring the space, you’re sampling a fraction of a fraction of it.
A useful way to think about the problem is as a landscape. Peaks represent high-performing formulations. Valleys represent the poor ones. The job of media development is to find the peaks as efficiently as possible - without being able to walk the terrain in advance.
Bayesian Optimization is the right tool for this because it does not try to cover the landscape uniformly. Instead, it builds a probabilistic model of the terrain from each batch of experimental results, then uses that model to direct the next round of testing toward the regions most likely to contain a peak. It searches intelligently rather than exhaustively. Bayesian-guided development requires 3 to 30 times fewer experiments than conventional DoE methods to find comparable or better formulations (Nature Communications, 2025).
Transfer learning compounds this advantage further. Every project run through our MediOP™ platform adds structured experimental data to a proprietary dataset. When the next project begins, the system is starting from accumulated knowledge - meaning the landscape is already partially mapped. That means faster convergence, better outcomes, and it means the platform gets better with every engagement.
Media optimized purely for titer tends to be expensive. Media optimized purely for cost tends to underperform.
Modern biomanufacturing demands that multiple objectives are prioritized simultaneously. That means titer, cost of goods, batch-to-batch consistency, regulatory compliance, ingredient availability, and scalability must interact at once. Even when improving, one can mean sacrificing the other.
The concept that captures this is the Pareto front, where it is physically impossible to improve one objective without worsening another. Any solution sitting on the Pareto front is defensible, and any solution inside it is leaving performance on the table.
Multi-objective optimization samples the Pareto front deliberately. Rather than converging on a single answer, it produces a range of formulations that each represent a different trade-off between competing objectives.
This approach also explains why AI-discovered formulations are often unintuitive. The algorithm does not carry the preconceptions that guide human experimental design. It explores the full landscape, including the combinations an experienced scientist would not think to try. A 2025 European Pharmaceutical Review article puts this into perspective, describing how AI-driven biomanufacturing can optimize process parameters, improve robustness and scalability, and reduce batch variability by finding conditions that human-designed experiments may miss. This is reflected in our work - some of the highest-performing formulations MediOP™ has produced would not have been reasoned to from first principles.
Software platforms that provide Bayesian optimization tooling but hand experimental work back to the scientists are solving half the problem. The model is only as good as the data it receives. The data is only as good as the experimental system generating it. If the wet lab work is slow, inconsistent, or limited in throughput, the optimization loop stalls - regardless of the quality of the algorithm. In OpenAI’s wet‑lab evaluation, a closed, iterative experiment–model loop driven by GPT‑5 delivered a 79‑fold increase in cloning efficiency, underscoring that rapid, repeated experimental feedback can matter more than any single static model improvement (OpenAI, 2025).
The MediOP™ platform is not a software layer placed over conventional lab operations. It is a system where hardware runs experiments under software direction, with results feeding back into the model in real time. The model directs the next experimental batch, the platform executes it, and the outputs update the model. Human expertise is embedded at the right points - ingredient qualification, regulatory interpretation, cell biology judgment - rather than as the bottleneck in every iteration.
This kind of integration is a harder engineering problem than the algorithm itself. It requires hardware, software, and wet lab capability to be built and refined together over the years. A team that hires an AI specialist and licenses a Bayesian optimization library has the algorithm. They do not have the system. Building the system takes the iterative work that cannot be fast-tracked.
It also generates the dataset that makes transfer learning possible. Every MediOP™ experiment contributes structured data to a record of how ingredients behave across cell types and applications. That record is the accumulated output of years of integrated operation.
Mechanistic understanding of cell behavior is fundamentally incomplete. Cells are not machines. The way a Chinese Hamster Ovary (CHO) cell responds to a shift in lipid concentration at a given pH, under a given dissolved oxygen level, in a given bioreactor geometry, is not fully predictable from first principles.
The correct response to that reality is to build systems that extract maximum signal from experimental data and translate it into formulations that perform. That requires the formulation space to be navigated intelligently, multiple objectives to be optimized simultaneously, and the full experimental loop to be integrated and automated. Not any one of these things in isolation - all three, built together, working as a single system.
That is what MediOP™ is. A platform designed from the ground up for the specific nature of the problem.
If you are working on a media development challenge and want to understand where MediOP™ could apply, book a call with our team.
Sign up for email updates:
