Biopharma · Statistical Methods

Design of Experiment (DoE) for Bioprocess : Design Evaluation Guide

Choosing a DoE design type is only the first step. Before running a single experiment, evaluating whether your chosen design is statistically sound sufficient power, manageable aliasing, acceptable prediction variance, and adequate degrees of freedom — is what separates robust bioprocess studies from ones that yield inconclusive or misleading results.

Topic: DoE Evaluation Level: Advanced Application: Bioprocess, CMC, Pharma

A poorly evaluated DoE design can result in missing critical effects, confounding important factors, or generating a model with insufficient predictive capability — all costly outcomes in biopharmaceutical development. Design evaluation answers four fundamental questions before any experiment is run.

Why Design Evaluation Matters?

Before running experiments, you must evaluate whether your design is statistically sound. This evaluation ensures you can detect real effects, avoid confounding, and generate reliable models.

Statistical Power

Can the design detect an effect of size that matters to you? Power is the probability of correctly identifying a real effect when it exists.

Target: ≥ 80% power
Aliasing & Confounding

Are any effects entangled with each other? Aliases prevent you from distinguishing which factor truly caused an observed response change.

Target: No main effect aliasing
Prediction Variance

How precisely can the model predict responses at any point in design space? Prediction variance should be low and uniform across the region of interest.

Target: Stable FDS profile
Residual Degrees of Freedom

Does the design have enough runs beyond what the model parameters require to estimate error and validate the model? Too few residual df means unreliable statistics.

Target: ≥ 3 residual df
Design Efficiency (D-optimal)

How efficiently does the design estimate model coefficients? D-efficiency measures the volume of the confidence ellipsoid — higher is better.

Target: D-efficiency ≥ 80%
Variance Inflation Factor (VIF)

VIF measures how correlated model terms are with each other. High VIF indicates multicollinearity — inflated parameter uncertainty and unstable estimates.

Target: VIF < 10 (ideally ≈ 1)

Statistical Power : Detecting What Matters

Power analysis determines whether your design has enough runs to detect an effect of a given size with acceptable certainty. In bioprocess development, power is typically evaluated at a signal-to-noise ratio relevant to your critical quality attributes (CQAs).

Power Benchmarks at Signal-to-Noise Ratio = 2 (Typical Bioprocess Scenario)
≥ 90%
Excellent : high confidence to detect real effects
Recommended for CQA-linked CPPs
80–89%
Acceptable : industry standard minimum
Suitable for most screening studies
< 80%
Insufficient : high risk of missing true effects
Add centre points or replicates
How to improve power without adding many runs: Adding replicates of centre points is the most efficient approach, it improves power for curvature detection and provides a pure error estimate. For screening designs, reducing the number of factors (after OFAT pre-screening) also increases per-factor power significantly.

Aliasing and Confounding : Resolving Effect Ambiguity

In fractional factorial designs, not all effects can be independently estimated. Some effects are mathematically indistinguishable — they are aliased. Understanding the alias structure before running the experiment is critical to avoiding uninterpretable results.

Alias Structure : Effect Resolution Summary by Design Type
Design Type Resolution Main Effects Aliased? 2FI Aliased with 2FI? Curvature Estimable? Typical Use
Full Factorial (2ᵏ) Full (V+) No No With centre points Characterisation
Fractional Factorial (Res V) Resolution V No With 3FI only With centre points Screening → Char.
Fractional Factorial (Res IV) Resolution IV No Yes (2FI = 2FI) With centre points Screening only
Fractional Factorial (Res III) Resolution III Yes (ME = 2FI) Yes No Avoid if possible
Definitive Screening Design Inherent (IV+) No No Yes Screening + Char.
Response Surface (CCC / Box-Behnken) Full No No Yes (explicit) Optimisation

⚠ Resolution III designs in bioprocess: Avoid using Resolution III fractional factorials when interactions between bioprocess parameters (pH × DO, temperature × feed timing) are plausible. In cell culture systems, factor interactions are common — a Resolution IV or DSD is much safer even with the cost of additional runs.

Prediction Variance and FDS Plot

The Fraction of Design Space (FDS) plot is the gold standard tool for evaluating how uniformly a design can predict responses across the entire experimental region. It answers: "For what fraction of the design space can I predict with a given precision?"

Fraction of Design Space (FDS) — Scaled Prediction Variance Profile

0 1 2 3 4 0% 25% 50% 75% 100% Fraction of Design Space Scaled Pred. Variance Target ≤ 2
RSM / CCC — Ideal flat variance profile
Full Factorial + Centre Points — Acceptable
Low-resolution Fractional — Poor prediction at extremes

A good FDS profile stays flat and below SPV = 2 across at least 80% of the design space. Steep rises at the high end of the FDS plot indicate that predictions near the boundaries of the experimental region are unreliable — a major concern for bioprocess design space definition.

Residual Degrees of Freedom : Ensuring Model Validity

Residual degrees of freedom (df) are the runs remaining after fitting all model terms. They provide the error estimate needed for hypothesis testing, lack-of-fit assessment, and overall model validation.

≥ 6
Residual df
Excellent
Robust error estimation, reliable lack-of-fit test, strong model validation
4–5
Residual df
Good
Adequate for most bioprocess characterisation studies
3
Residual df
Minimum Acceptable
Add replicates if budget allows; lack-of-fit assessment limited
≤ 2
Residual df
Insufficient
Model cannot be validated; add centre point replicates immediately
Quick fix for low residual df: Adding 3–5 centre point replicates is the most cost-effective way to increase residual degrees of freedom. Centre points also enable detection of curvature without the full cost of a response surface design, and provide a pure error estimate independent of the model.

Comparing Designs Side by Side

When multiple design options exist for the same experimental objective, a structured comparison across evaluation criteria guides the optimal selection. Here is an example comparison for a 5-factor characterisation study in a fed-batch cell culture process.

Full Factorial 2⁵ = 32 runs
Power (SNR=2)98%
AliasingNone
Residual df16
D-Efficiency100%
CurvatureCPs only
Run count32 + CPs
Res V Frac. Factorial 2⁵⁻¹ = 16 runs
Power (SNR=2)91%
Aliasing2FI ↔ 3FI only
Residual df5
D-Efficiency100%
CurvatureCPs only
Run count16 + CPs
DSD (Definitive Screening) 11 runs
Power (SNR=2)82%
AliasingNone (MEs)
Residual df3–4
D-Efficiency~88%
CurvatureYes (explicit)
Run count11 + CPs

Note: DSD's key advantage is detecting curvature with far fewer runs. If budget is constrained and curvature is expected in your bioprocess (common in temperature and pH effects on cell growth), DSD is often the superior choice despite slightly lower power.

Pre-Experiment Design Evaluation Checklist

Before executing any DoE in a bioprocess setting, work through this checklist to confirm your design is statistically sound and fit for purpose.

Power ≥ 80% at minimum detectable effect size
Define the smallest effect on your response (e.g., titre, viability, glycoform) that is biologically or commercially meaningful. Verify design power at that effect size at your expected signal-to-noise ratio.
Alias structure reviewed : no main effects aliased with 2FIs
Print the alias structure for your fractional design and confirm that no main effects (CPPs) of interest are aliased with two-factor interactions that are scientifically plausible in your system.
!
FDS plot reviewed : SPV ≤ 2 for ≥ 80% of design space
Generate the FDS (Fraction of Design Space) plot in your software. Confirm that scaled prediction variance stays below 2 across most of the design space, especially in the region of greatest process interest.
Residual degrees of freedom ≥ 3 (target ≥ 6)
Count residual df = total runs − number of model parameters. Add centre point replicates if below 3. Verify that the lack-of-fit test will be estimable (requires pure error from replicates).
VIF < 10 for all model terms
Check Variance Inflation Factors for all terms in the planned model. VIF values above 10 indicate multicollinearity — terms are too correlated for stable parameter estimation. Reconfigure the design or reduce model terms.
!
Factor ranges set to span the design space of interest
Factor ranges that are too narrow miss real effects; too wide risks going outside safe operating windows. Validate ranges against prior knowledge, OFAT scouting data, and cell line or process constraints.
Randomisation plan confirmed and blocking considered
Confirm runs will be randomised to protect against time-trends and systematic bias. If experiments will be run across multiple days, cell culture passages, or bioreactor systems, introduce blocking to account for known sources of nuisance variation.
Do not proceed if any main effect is aliased with a plausible 2FI
If your alias structure shows a main effect of interest confounded with a two-factor interaction that could realistically occur in your bioprocess, upgrade to a higher-resolution design or switch to a DSD before running experiments.
GRAPHTAL SERVICES

Bioprocess Design of Experiment (DoE) — Full Suite

DoE Planning
Advanced
Statistical Analysis
Visualisation &
Report Writing
Process Optimisation,
Scale-up & Down
Troubleshooting
Support

Need expert DoE design evaluation for your bioprocess?

Graphtal evaluates power, aliasing, prediction variance, and model validity before you run a single experiment.

Go Back

Tell Us About Your Project

Our team of experienced professionals at Graphtal is ready to transform your project from idea to
reality, ensuring alignment with your organisation goals through advanced data analytics and
predictive modelling.


Let's simplify your work

Want to make Insightful Analytics for Smarter Decisions?

Let's connect.