Design of Experiments for Biologists & Bioprocess Engineers: The Complete Guide
If you've ever changed one variable, run the experiment, changed the next variable, run it again — you've done One-Factor-at-a-Time (OFAT) experimentation. It's the default approach in most biology and bioprocess labs, and it has a real problem: it can't see interactions. If your yield only improves when temperature and pH move together, an OFAT sweep of temperature alone will never find that, no matter how many runs you throw at it.
Design of Experiments (DOE) fixes this by changing several variables together, in a small, carefully chosen set of runs, so you can see both the individual effect of each variable and how they interact. It's been standard practice in engineering and quality control for decades. It's underused in biology mostly because the software built for it was written for statisticians, not bench scientists — this guide skips that barrier and explains DOE the way you'd actually use it.
→ Full walkthrough: DOE vs. OFAT: Why Changing One Variable at a Time Wastes Runs
The three things DOE can tell you
Every DOE project starts with picking which of these three questions you're actually asking:
"I have a lot of variables and don't know which ones matter." This is a screening design. You're not trying to find the exact optimum yet — you're trying to cheaply rule out the variables that don't do anything, so you can focus follow-up work on the two or three that do.
"I know my important variables and want to understand how they interact." This is a characterization design. It adds enough runs to estimate interaction effects cleanly and checks whether the relationship is a straight line or has curvature — i.e., whether there's a peak somewhere inside your tested range rather than the response just climbing or falling monotonically.
"I know my key variables and want the exact peak settings." This is an optimization (response surface) design. It fits a full curved model across your variables and lets you solve for the settings that maximize (or minimize) your outcome — including balancing two competing outcomes at once, like maximizing yield while keeping impurity below a threshold.
Most projects move through these in order: screen a long list down to the vital few, then optimize those few. You don't have to plan this upfront — see the blocking section below for why post-hoc flexibility matters here too.
→ Full walkthrough: Screening Design vs. Response Surface: Which Do You Need?
→ Not sure which of the three you're in, or want the edge cases? Which DOE Design Should I Use? A Decision Guide
→ Screening designs can also catch curvature without a second experiment — see Definitive Screening Designs (DSD), Explained
→ Want the mechanics behind any specific design type? Full Factorial Designs, Fractional Factorial Designs and Aliasing, Plackett-Burman Designs, and Response Surface Methodology each get their own deep dive.
Handling constraints real labs actually have
Textbook DOE assumes every run is equally easy to perform in any order. Real lab work rarely is.
Some variables are expensive to change. Incubator temperature, which bioreactor vessel you're using, which analyst is running the assay — resetting these between every single run can turn a one-day experiment into a two-week one. Flagging a variable as Hard-to-Change tells the design to group runs so that variable changes as few times as possible, and the analysis (a split-plot / REML model) accounts for the fact that runs sharing a hard-to-change setting aren't fully independent of each other. Get this wrong — treating them as independent when they're not — and you'll see "significant" effects that are really just noise from whichever settings happened to run together.
→ Full walkthrough: Hard-to-Change Factors and Split-Plot Designs, Explained
You don't always know you need blocks until you've already generated the design. Splitting work across two lab days, or across two analysts, is usually a logistics decision that only becomes obvious once you're staring at a 16-run list. You don't need to have planned for this from the start — an existing design can be split into balanced blocks after the fact, without adding runs or redoing the design.
→ Full walkthrough: How to Block an Experiment After You've Already Designed It
Seeing this applied to a real workflow
Two worked examples, if you want to see these ideas applied to an actual project rather than in the abstract:
→ Designing a DOE for CHO Cell Culture Media Optimization
→ Using DOE to Optimize ELISA and Immunoassay Development
Where regulatory expectations fit in
If you're working under ICH Q8 or Q2, the DOE work above isn't just good practice — it's the evidence base those frameworks ask for directly. Screening candidate Critical Process Parameters and mapping their effect on your Critical Quality Attributes is how a Design Space gets built and documented; testing method robustness by varying parameters together instead of one at a time is how ICH Q2 robustness testing is supposed to work in the first place.
→ Full walkthrough: DOE and ICH Q8: Building the Design Space
→ Full walkthrough: DOE for Analytical Method Validation (ICH Q2)
Sizing the experiment before you run it
The question every design ultimately has to answer is: how many runs do you actually need? Too few, and a real effect hides in the noise. Too many, and you've spent weeks and reagents on precision you didn't need.
The inputs are just two numbers you already know from experience: how much your assay's result normally varies between identical repeat runs, and the smallest improvement that would actually be worth acting on. From those two numbers, you can calculate — before running a single sample — whether your planned run count has a realistic chance of detecting that improvement, or whether you're setting yourself up to run an experiment that can't answer the question you're asking.
→ Full walkthrough: How Many Runs Do You Actually Need for a Design of Experiments?
Reading what comes out the other end
Once results are in, two outputs do most of the work: a significance table (which variables had a real, non-random effect) and a Pareto chart of the same information ranked by size. A third check — Actual vs. Predicted — tells you whether the model fits your data well enough to trust the first two at all. Skipping that check is the single most common way people over-interpret a DOE result.
→ Full walkthrough: Reading Your ANOVA Table and Pareto Chart
Where to start
If you want to see this end-to-end rather than in the abstract — defining variables, generating a design, and reading a real result — the walkthrough below covers a full run of the process. Or open BioDOE and try it on your own experiment; the free tier covers everything in this guide.