Digital & IT
AI bioprocess optimisation as a service
The statistics of learning a bioprocess from a few dozen expensive, correlated runs, why hybrid models exist, and the specific problems introduced when the model is hosted by someone other than the manufacturer.
A bioprocess model has to learn a high-dimensional, non-linear, time-dependent system from a training set that is unusually small and unusually correlated. This is not a shortcoming of any particular vendor’s mathematics; it follows from what a run costs. A development campaign generates tens of fermentations, not thousands, each taking one to three weeks, each consuming media and analyst time, and each executed under a protocol designed to keep the process on target. That last point is the awkward one: good process development deliberately holds variables steady, so the runs available for training carry little of the variation a model needs in order to separate causes.
Correlation is the binding constraint, not sample size
Feed rate, cell density, oxygen uptake, lactate and osmolality move together in a normal run. A model fitted to that data can predict titre accurately inside the operating envelope and still be wrong about which variable to change, because it has learned an association rather than a response. The remedy is designed experiments that deliberately break those correlations — which costs runs, in a setting where runs are the scarce resource. Every serious platform in this space is an attempt to buy information per run: Bayesian optimisation and design-of-experiment schedulers choose the next condition by expected information gain rather than by expected yield, and accept a deliberately poor run when it is informative.
The second family of answers is structural. A hybrid model keeps the parts that are actually known — mass balances, Monod-type kinetics, gas transfer — as mechanistic equations, and uses machine learning only for the residual it cannot write down, typically the specific rates. The mechanism supplies constraints for free, so far fewer runs are needed to fix the remaining parameters, and extrapolation degrades gracefully instead of catastrophically. A purely empirical model has no such floor.
Scale is where the model breaks
A model trained at bench scale encodes the transport regime of the vessel it was trained in. At production scale, mixing time is longer, gas hold-up and bubble size differ, and the culture experiences gradients in dissolved oxygen, pH and substrate that simply do not exist in three litres. Scale-up rules — constant volumetric power input, constant tip speed, constant oxygen transfer coefficient — cannot all be satisfied at once, so something changes. The measured variables still arrive; the relationship between them has moved. The sibling page on bioprocess analytics and PAT covers the same failure for spectroscopic calibration models and the statistics used to detect it, and that discussion applies here unchanged.
What is different about off-premises
Three things, none of them about the mathematics. First, process data is among a manufacturer’s most sensitive assets, and the value proposition of a multi-tenant platform — pooling runs across clients so models see more variation — is in direct tension with that. Pooling also runs into the transfer problem above: another client’s runs came from another plant. Second, the validated state is the manufacturer’s responsibility under EU GMP Annex 11 and 21 CFR Part 11 regardless of who hosts the software, so a vendor-side model update is a change to a validated system that the manufacturer did not schedule. Third, a model behind an interface is harder to interrogate; if the recommendation cannot be traced to a mechanism, the process engineer cannot judge whether it will hold at scale.