Key Findings
Bayesian optimization (BO) enables bioprocess development teams to find optimal operating points in a fraction of the runs required by grid or response surface designs. In specific applications, such as recombinant protein production in Komagataella phaffii and cell culture media for peripheral blood mononuclear cells, BO reached comparable or better outcomes using 3–30× fewer experiments than standard DOE.
Technical Details
BO is a sequential strategy for optimizing expensive black-box functions where responses are noisy and mechanistic models are unavailable. The process consists of two main components: a surrogate model, typically a Gaussian process (GP), which predicts responses and uncertainty, and an acquisition function, such as Expected Improvement (EI), which scores the value of candidate conditions.
The operational loop follows four steps:
- Prior model: Fitting the GP to 5–10 initial space-filling experiments or historical data.
- Acquisition: Maximizing EI to select the next experiment.
- Experiment: Measuring responses such as titer or viability.
- Update: Adding results to the dataset and refitting the GP until convergence.
A typical campaign requires 15–30 experiments: 5–10 initial points followed by 10–20 sequential evaluations.
Background & Context
Unlike traditional DOE, which uses a fixed matrix to quantify main effects and interactions, BO is an adaptive approach. DOE is preferred for screening long factor lists, small factor spaces (2–3 factors), or regulatory filings under ICH Q8. BO is most effective when the list of influential factors has already been narrowed and the goal is maximizing titer rather than mapping the entire space. A recommended sequential workflow involves using a screening design to identify critical factors, BO to refine them, and a final confirmatory DOE around the optimum.
Significance & Outlook
BO addresses constraints in bioreactor time, media components, and analytical capacity. By concentrating experiments in promising regions and abandoning poor regions quickly, it accelerates the path to peak yield. For example, in a simulated three-factor media problem, BO reached 90% of its final value by run 7, whereas full factorial designs require the completion of the entire matrix before providing an answer.
Source: https://bioprocesstools.com/blog/bayesian-optimization-bioprocess/
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
