FairIntellect Technologies Book a call
Book a 30-minute call
contact@fairintech.com +1 (440) 760-5692

Design of Experiments

Stop testing one factor at a time.

Design of experiments (DoE) shows which conditions matter, how they interact and where the optimum is, with fewer runs than testing every combination. We plan the design with you, get it ready for your bench, and turn your results into answers for the next experiment.

Book a 30-minute call Try the design planner

Why DoE

The optimum hides where factors interact.

Change one factor at a time and you only ever look along one line at a time. When the best glucose level depends on the temperature, you stop on a ridge below the peak. A designed experiment maps the whole ridge instead, so you see where it peaks.

  • Interactions, measured. Factors change together in a planned pattern, so effects that only appear in combination show up.
  • More certainty per run. Every run informs every effect, instead of one comparison at a time.
  • A model of the whole range. A prediction for every setting in between, not just the points you tested.
  • Ranges you can defend. The multivariate evidence that Quality by Design (ICH Q8, Q11 and Q14) builds on, if the work later moves toward a filing.
Fig. 1. Temperature and glucose for an illustrative production strain. One factor at a time (10 runs) stops at 33 °C and 18 g/L, at 76% of the best titer. A central composite design (13 runs, after the 20-run screen in Fig. 3) maps the ridge; its fitted model puts the optimum at 35.1 °C and 17.7 g/L, at 99%. Simulated data.

Two ways to work with us

One experiment, or a pipeline for every experiment.

Start with a single campaign, or have the whole loop built once and run it yourselves. Either way the analysis is yours to rerun, with or without a statistics license.

For one campaign

DoE sprint

For teams without a statistician on hand, before a media, assay or formulation campaign. We plan the experiment with you, you run it, and we turn your results into a decision.

  • A design session: factors, ranges, constraints and controls
  • Plate maps and a randomized layout, ready to pipette
  • The analysis: effects, interactions and a model checked for fit
  • A report with the optimum, its uncertainty and the next experiment
  • The analysis script, so the next round needs no license

Fixed scope · one round of experiments

For teams that run DoE often

DoE pipeline

For labs with a liquid handler and a plate reader that run designed experiments regularly. We build the loop once; you own it and run it.

  • Designs generated from your factors and labware rules
  • Liquid-handler worklists, with volumes calculated from your stock concentrations
  • Plate-reader exports parsed and matched to each well
  • Analysis and report generated after every plate
  • Results written to your ELN or LIMS, with a trail back to the raw files

Built in stages · you own the code

Already using JMP? Keep it: we can work from its exported designs and data tables. Not using it? Every analysis we deliver runs on open-source R or Python.

DoE pipeline

From design to decision, without the spreadsheets.

Most of the work in a DoE isn't statistics. It's turning concentrations into pipetting volumes, laying out plates and matching exports back to runs. That's the part we automate.

  • Randomized plate layouts, with edge wells filled with water or medium
  • Worklists checked against your stock concentrations and minimum volumes
  • Every well traced from design point to raw reading to model

Typical connections

  • Liquid handlers
  • Acoustic dispensers
  • Plate readers
  • Parallel bioreactors
  • ELN and LIMS
Fig. 2. One design point on its way to the bench: placed in a random well, then turned into liquid-handler transfers whose volumes follow from the stock concentrations, 200 µL per well. It is point 7 of the planner's default plan below. Illustrative.

Why a microbiologist

Biology is noisy. The
design is not.

A design that works on paper can still fail at the bench. Living systems add variation a statistics package can't see, so we plan for it from the start.

i.

Plates have geography

Edge wells evaporate faster and change temperature sooner. We randomize positions, blanks included, and fill the edge wells with water or medium, so a well's position can't pose as an effect.

ii.

Days are not identical

Inoculum, passage number, reagent lots and the person at the bench all vary. We block by day or batch, so that variation is measured instead of mixed into the result.

iii.

Replicates mean biology

Three wells from one culture are not three experiments. We plan biological replicates where they count, and controls on every plate.

iv.

Some factors set the whole plate

Temperature and shaking speed can't change from well to well. We use split-plot designs for them, so the analysis stays honest about what was randomized.

Try it

Plan a screening experiment in silico.

Choose how many factors to test. The planner picks a standard design, randomizes it across a 96-well plate and tells you what the design can and cannot separate. Nothing is uploaded: it all runs in your browser.

Factor names and levels

Randomization seed 1

20 wells: 16 design points and 4 center points, resolution V.

Wells
20
Design points
16
Resolution
V
Full factorial
32 runs

Every main effect and every two-factor interaction is estimated separately, if interactions of three or more factors are negligible. The 4 center points test for curvature and estimate pure error.

One row per well, in plate order. Low levels plain, high levels shaded, center points in green.
WellPointGlucose (g/L)Yeast extract (g/L)pHNaCl (g/L)Inoculum (% v/v)
B6C312.56753
B10C412.56753
B111620107.4105
C55527.401
C662027.405
C119526.6101
D735106.601
D8C112.56753
D11102026.6105
E8115106.6105
E1122026.601
F2142027.4101
F4C212.56753
F8420106.605
F10820107.401
G41220106.6101
G575107.405
G91526.605
G1013527.4105
G11155107.4101

Two-level designs from the NIST/SEMATECH e-Handbook of Statistical Methods: full factorials, the fractional factorials of its Table 3.17 with up to 8 factors and 32 runs, and the 12-run Plackett-Burman design. Only factors you can pipette are listed, because temperature and shaking need a split-plot design. A demo, not the service: a real design also handles blocking by day or batch, constraints between factors and a second round to find the optimum.

What you get back

Answers, with the uncertainty attached.

  • What matters

    Effects ranked by size, interactions included, and factors dropped on evidence rather than habit.

  • Where the optimum is

    The best settings and the result predicted there, with a 95% prediction interval, and whether the peak lies inside the range you tested.

  • How far to trust it

    Fit, lack of fit and residual checks, explained in plain language.

  • What to run next

    Confirmation runs at the optimum, or the next design, sized to your plates.

Fig. 3. An illustrative two-round report from the study behind Fig. 1: five factors screened in 20 parallel bioreactor runs, then temperature and glucose optimized in 13 more. Simulated data.

Next step

About to run an experiment one factor at a time?

Tell us in a 30-minute call. You'll leave with a clear next step, whether or not we end up working together.

Book a 30-minute call Email us