Designing a DOE for CHO Cell Culture Media Optimization

Part of our complete guide to DOE for biologists.

You bump glucose feed rate up, titer improves a little. Next run you leave glucose alone and bump the amino acid feed instead — a little more improvement. You never test them together, so you never find out that the real gain only shows up when both move at once. That's the OFAT trap in cell culture media optimization, and it's exactly the interaction DOE is built to catch.

What actually varies in a media optimization run

A typical CHO titer-improvement project has a mix of variable types, and getting the mix right matters more than the exact list:

• Continuous variables: Glucose concentration, key amino acid feed rates, pH, dissolved oxygen setpoint.
• Categorical variables: Feed strategy (e.g. Feed A vs. Feed B vs. a blend), media lot or vendor.
• Hard-to-Change variables: Temperature shift timing is the classic one — resetting a bioreactor's temperature between every single run isn't practical, so it needs to be flagged and handled with a split-plot design rather than treated like the others.

Two responses that pull in different directions

Titer optimization rarely has one target. You're almost always maximizing titer while keeping viability above some floor — push feed too aggressively and titer climbs but the culture crashes early. This is a multi-goal optimization problem: define both responses, set a goal for each (maximize titer, keep viability ≥ some threshold), and let the design find settings that satisfy both instead of manually trading them off by eye.

Picking the design stage

If you're staring at five or six candidate variables and don't know which actually move titer, start with a screening design — it's built to cheaply rule out the ones that don't matter, not to find the exact optimum. Once you're down to the two or three that do, move to an optimization (response surface) design on just those, which can find a genuine peak rather than just a direction to keep pushing in.

Don't skip straight to a response surface design with six variables — the run count needed to characterize curvature across that many dimensions gets expensive fast. Screen first.

Sizing it before you commit bioreactor time

Titer assays have real run-to-run noise — plate reader variability, sampling error, day-to-day culture variability. Before committing bioreactor runs, it's worth knowing: given how noisy your titer measurement normally is, and the smallest improvement that would actually be worth changing your process for, does your planned run count have a realistic shot at detecting it? That's a five-minute calculation, not a guess.

Reading the result

When the runs are in, the significance table tells you which variables had a real effect on titer and viability independently, and whether any of them interact (a common one: feed rate and temperature shift often trade off against each other). The optimizer then reports predicted optimal settings for both responses at once — not just "titer is maximized here," but "titer is maximized here while viability stays above your floor."

Ready to try it on your own media optimization project? Start free — the workflow above is exactly what BioDOE's three panels walk you through.