How to Design a Media Optimization Study Using DOE

Why DOE for Media Optimization?

Design of experiments (DOE) is the most efficient statistical method for optimizing cell culture media because it reveals both individual component effects and interactions between components in far fewer runs than one-factor-at-a-time (OFAT) approaches. A chemically defined CHO medium typically contains 50–80 components, and OFAT testing even 10 of these at 3 levels would require 30 independent experiments that miss every interaction effect. DOE detects interactions by design.

Media composition directly controls cell growth, product titer, and critical quality attributes like glycosylation. Amino acids alone can account for 25–40% of total media cost in large-scale mammalian cell culture, making optimization both a productivity and an economics problem. DOE-guided media optimization routinely improves titers by 30–70%, with published studies reporting mAb titers increasing from 3–4 g/L to 7–10 g/L after systematic amino acid rebalancing.

The core advantage of DOE over OFAT is the ability to detect interaction effects. If glutamine and asparagine both individually increase growth by 10%, their combination might yield a 30% increase (synergy) or only 5% (antagonism). OFAT experiments cannot detect this—only DOE with appropriately designed factor combinations can.

The Sequential DOE Workflow

Media optimization follows a staged approach that progressively narrows the design space from dozens of candidate components down to the optimal concentration of 3–5 critical factors. Each stage uses a different DOE design type matched to its objective.

Step 1: Screening — Identifying Critical Components

The screening stage identifies which media components significantly affect your responses from a large initial list. For a CHO fed-batch process, you might start with 10–15 candidate components drawn from spent media analysis, literature, and supplier recommendations. The goal is to reduce this to 3–5 critical factors for detailed optimization.

Choosing Candidate Components

Start by analyzing spent media from your baseline process. Components that are depleted below 20% of their initial concentration by harvest are candidates for supplementation. Amino acids are almost always the highest-impact group: glutamine, asparagine, serine, and cysteine are commonly limiting in CHO cultures. Trace metals (iron, zinc, copper, manganese) and vitamins (B12, biotin, folic acid) are also frequent screening candidates.

Group related components when the total factor count exceeds 15. For example, test "branched-chain amino acids" (leucine, isoleucine, valine) as a single group factor at low and high levels rather than as three separate factors. This reduces the design size while preserving the ability to detect group-level effects.

Setting Factor Levels

For screening, use two levels per factor: a "low" level (typically 0.5–0.75× the current concentration) and a "high" level (1.5–2.0× the current concentration). The range must be wide enough to produce a measurable response difference but not so extreme that it causes toxicity or precipitation.

• Amino acids: 0.5–2.0× baseline (never exceed solubility limits; tyrosine and cysteine are poorly soluble above 2–4 mM)
• Trace metals: 0.5–3.0× baseline (copper is toxic above 0.1–0.5 µM for most CHO lines)
• Vitamins: 0.5–5.0× baseline (water-soluble vitamins are rarely toxic, so wider ranges are safe)
• Glucose/feed rate: 0.7–1.5× baseline (lower bound avoids glucose depletion; upper bound avoids lactate accumulation)

The Plackett–Burman Design

A Plackett–Burman (PB) design is the most run-efficient screening design, requiring only N+1 experiments (rounded to the nearest multiple of 4) for N factors. It estimates main effects independently but aliases all two-factor interactions with main effects, so any significant "main effect" might actually be a disguised interaction. This is acceptable for screening because the purpose is to identify the important factors, not to build a precise model.

Always include 3–5 center-point replicates (all factors at their midpoint). Center points serve three purposes: (1) they estimate pure error for the F-test, (2) they detect curvature (if the center-point mean differs significantly from the factorial mean, at least one factor has a quadratic effect), and (3) they verify that the baseline condition is reproducible.

Step 2: Optimization — Mapping the Response Surface

Once screening identifies 3–5 critical components, optimization DOE maps the curved response surface to find the true optimum concentration for each. The two most common designs are the central composite design (CCD) and the Box–Behnken design (BBD).

Central Composite Design (CCD)

A CCD consists of three parts: the 2^k factorial points, 2k axial (star) points at a distance α from the center, and center-point replicates. For k = 3 factors, this means 8 factorial + 6 axial + 5–6 center = 19–20 total runs. The CCD fits a full second-order polynomial model:

Y = β₀ + Σ β_iX_i + Σ β_iiX_i² + Σ β_ijX_iX_j + ε

Set α = (2^k)^1/4 for rotatability (equal prediction variance at equal distances from the center in all directions). For 3 factors, α = 1.682. Use face-centered CCD (α = 1) if you cannot run factor levels beyond the original screening range.

Box–Behnken Design (BBD)

A BBD avoids extreme corners of the design space by only testing factor combinations at the midpoints of edges and the center. For 3 factors, it requires only 13 runs (vs. 20 for a CCD). This is advantageous when corner combinations might cause cell death (e.g., simultaneously high copper and low cysteine). The trade-off is that BBD cannot detect effects at the extremes of the design space.

Choosing Between CCD and BBD

• Use CCD when you can safely test extreme combinations and want rotatable prediction variance.
• Use BBD when corner conditions risk cell toxicity, precipitation, or other practical constraints.
• For k ≥ 5 factors, consider a face-centered CCD with fractional factorial core to keep runs manageable (≤ 32).

Definitive Screening Designs: Screen and Optimize in One Step

A definitive screening design (DSD) combines the efficiency of screening with the ability to detect curvature and some two-factor interactions in a single experimental campaign. Introduced by Jones and Nachtsheim in 2011, DSDs require only 2k+1 runs for k factors (e.g., 17 runs for 8 factors) while providing three levels per factor.

DSDs achieve this efficiency through a conference-matrix construction that ensures all main effects are orthogonal to each other, orthogonal to all two-factor interactions, and orthogonal to all quadratic effects. In practice, this means you can fit main effects + any significant quadratic or interaction terms without confounding, provided the number of active effects is fewer than about half the number of runs.

When to Use DSD vs. Sequential PB + RSM

• Choose DSD when you have 5–16 factors, suspect curvature in the response, and want to reduce the total number of experimental campaigns from two (screening + optimization) to one.
• Choose PB + RSM when you have more than 16 factors, need Resolution IV aliasing structure (PB is Res III), or when the screening step is run at smaller scale (e.g., 96-well plates) while optimization moves to shake flasks or mini-bioreactors.

A 2024 study by Bai et al. compared fractional factorial designs and DSD for CHO cell culture process development in Ambr 15 miniature bioreactors. The DSD was more effective at detecting both main and quadratic effects while using fewer runs, with models that transferred well from the 15 mL to 3 L bioreactor scale.

Worked Example: CHO mAb Media Optimization

This worked example walks through a realistic CHO mAb media optimization study from screening through verification, showing how to size each stage and interpret the results.

Stage 1: PB Screening (12 runs + 3 center points)

Stage 2: Path of Steepest Ascent (6 runs)

The steepest ascent method moves along the gradient defined by the screening model coefficients. Starting from the center point, each step increases glutamine, asparagine, and iron in proportion to their estimated effects. After 6 steps, titer peaked at step 4 (4.6 g/L) then declined at step 5, indicating the region near the optimum.

Stage 3: CCD Optimization (20 runs)

Stage 4: Verification (5 runs)

Five replicate shake flasks at the predicted optimum yielded titers of 5.3, 5.5, 5.7, 5.4, and 5.6 g/L (mean = 5.5 ± 0.16 g/L, CV = 2.9%). The predicted 5.6 g/L fell within the 95% confidence interval of the observed mean. The optimized medium was then scaled to 3 L bioreactors, achieving 5.8 g/L with improved mixing and DO control.

Titer improvement: 3.2 → 5.5 g/L = 72% increase
Total experiments: 15 + 6 + 20 + 5 = 46 runs over 10 weeks

Common Pitfalls and How to Avoid Them

Media optimization DOE studies fail most often due to avoidable experimental design and execution errors, not statistical complexity. Here are the seven most common pitfalls and their remedies.

1. Factor Ranges Too Narrow

If the low and high levels are too close, the signal is buried in noise. A rule of thumb: the expected response difference between low and high levels should be at least 2–3× the standard deviation of your measurement assay. For mAb titer measured by Protein A HPLC (typical CV = 3–5%), factor levels should produce at least a 10–15% titer difference.

2. Ignoring Center Points

Skipping center-point replicates removes your ability to detect curvature and estimate pure error. Without pure error, you cannot compute a lack-of-fit test. Always run at least 3 center points in any screening or optimization design.

3. Confounding Run Order with Factor Effects

If you run all "high glutamine" conditions on Monday and all "low glutamine" on Friday, passage number, media age, or operator fatigue become confounded with the glutamine effect. Always randomize run order. If complete randomization is impractical (e.g., hard-to-change factors like temperature), use a split-plot design.

4. Measuring Too Few Responses

Optimizing titer alone can inadvertently degrade glycosylation, charge variants, or aggregation. Always measure at least 3 responses: growth (VCD or IVC), productivity (titer), and quality (glycan profile, SEC monomer %, or charge heterogeneity by icIEF).

5. Oversized Optimization Designs

Running a CCD with 5 factors when only 3 were significant in screening wastes resources. Carry forward only factors with p < 0.05 (or p < 0.10 if borderline and biologically plausible). Use the screening model's lack-of-fit and curvature tests to decide whether augmentation is needed before jumping to RSM.

6. Skipping Verification

The predicted optimum from the response surface model is a statistical estimate with confidence intervals. Running 3–5 verification replicates at the predicted optimum is essential to confirm reproducibility and check that the model's prediction error is acceptable (<10% relative error for titer).

7. Not Accounting for Scale Effects

Media optimization in 96-well plates or shake flasks may not transfer perfectly to stirred-tank bioreactors. Oxygen transfer, mixing time, and pH control differ across scales. Run the verification stage at the target scale (or at least at miniature bioreactor scale with controlled DO and pH) to confirm robustness.

Frequently Asked Questions

Related Tools

• Media Cost Estimator — Calculate media costs per liter across scales and compare vendor formulations.
• Fed-Batch Calculator — Model nutrient feeding profiles and predict metabolite trajectories.
• ELISA 4PL Analyzer — Fit 4-parameter logistic curves to titer ELISA data from DOE experiments.

References

1. Torkashvand F. et al. "Designed Amino Acid Feed in Improvement of Production and Quality Targets of a Therapeutic Monoclonal Antibody." PLOS ONE, 10(10): e0140597, 2015. doi: 10.1371/journal.pone.0140597
2. González-Leal I.J. et al. "Use of a Plackett–Burman statistical design to determine the effect of selected amino acids on monoclonal antibody production in CHO cells." Biotechnology Progress, 27(6): 1709–1717, 2011. doi: 10.1002/btpr.674
3. Zhou T. et al. "A review of algorithmic approaches for cell culture media optimization." Frontiers in Bioengineering and Biotechnology, 11: 1195294, 2023. doi: 10.3389/fbioe.2023.1195294
4. Bai Y. et al. "Enhancing early-stage cell culture process development efficiency using an integrated approach of high-throughput miniaturized bioreactors and definitive screening design." Biochemical Engineering Journal, 203: 109217, 2024. doi: 10.1016/j.bej.2024.109217
5. Narayanan H. et al. "Accelerating cell culture media development using Bayesian optimization-based iterative experimental design." Nature Communications, 16: 4539, 2025. doi: 10.1038/s41467-025-61113-5