Methodology
BioDOE's interface is written in plain language on purpose — "Confidence Score" instead of p-value, "Experimental Dial" instead of factor — because most of the scientists using it aren't statisticians, and shouldn't need to be. This page is for the people who need the version underneath: what's actually being computed, which methods are used where, and why. Nothing here is simplified.
1. How the app's intents map to real designs
Panel 1 asks "what are you trying to do?" instead of "pick a design class." Underneath, that choice selects a specific, standard DOE construction:
| In-app choice | What it actually generates |
|---|---|
| Find what matters (Screening) | A Resolution IV (or higher) two-level fractional factorial — main effects are unconfounded with each other and with two-factor interactions, at the smallest run count that supports it. |
| Map the shape (Characterization) | A full two-level factorial plus center points, so the design can support per-factor pure-quadratic (curvature) terms in addition to main effects and two-way interactions — not just detect that curvature exists, but locate it. |
| Find the peak (Optimization) | A response-surface design (center points plus, where applicable, augmented runs) built to support a full second-order model for direct optimization. |
2. Model fitting
Each response is fit independently by ordinary least squares against the richest model the design's degrees of freedom actually support, in this priority order: main effects + two-way interactions + per-factor quadratic terms, falling back to main effects + two-way interactions, falling back to main effects only. Significance is reported per term via standard t-tests on the fitted coefficients (what the UI shows as "Confidence Score" is the model's p-value for that term; anything below the conventional 0.05 threshold is flagged significant).
An unreplicated, fully-saturated screening design (e.g. 8 runs testing 7 two-level factors) has zero residual degrees of freedom, so no p-values can be computed in the usual sense — this is expected, standard practice for these designs, not a limitation of BioDOE. In that case, effects are ranked by magnitude using Lenth's pseudo standard error (PSE) method, shown as the Half-Normal Plot.
Curvature and quadratic terms
A factor only gets a pure-quadratic term (I(F² ) in the fitted formula) if
the design actually visited a third distinct level for it (e.g. a center point) — otherwise
there's nothing to estimate curvature from, and the term is correctly omitted rather than
silently fabricated.
3. Hard-to-change factors and Plate/Plot Capacity
When one or more factors are flagged Hard-to-Change, BioDOE doesn't just group the run
order — it fits a genuine split-plot mixed-effects model via statsmodels.MixedLM,
with REML estimation, treating the whole-plot grouping (the hard-to-change combination) as
a random effect. This correctly separates whole-plot error (variation between hard-to-change
settings) from subplot error (variation between easy-to-change runs within a setting),
rather than understating the standard error of the hard-to-change factors the way a naive
OLS fit on the same data would.
If a whole plot's run count exceeds what physically fits in one container (a 96-well
plate, a single bioreactor run), Plate/Plot Capacity adds a second, nested random-effect
level — Plate within Whole Plot (vc_formula={'Plate': '0 + C(Plate_Global)'})
— so plate-to-plate noise is absorbed into its own variance component instead of
contaminating the hard-to-change factor's effect estimate.
If the REML fit doesn't converge to a stable solution (most often too few whole-plot groups for the number of terms being estimated), BioDOE falls back to a standard independent-runs model and says so explicitly in the results — it never silently substitutes a less appropriate model without telling you.
4. Multi-response optimization ("Sweet Spot")
When more than one response has a stated goal (Maximize / Minimize), BioDOE combines them into a single desirability function per response — scaled 0 to 1 across each response's observed range in the direction of its goal — and searches a grid across the design space for the point that maximizes the combined (geometric-mean-style) desirability across all goals simultaneously. This is the same overall approach as classical desirability-function optimization (Derringer & Suich), not a black-box "AI recommendation." The contour plot shows that desirability surface directly; the star marks the grid optimum.
5. What BioDOE doesn't do
- It doesn't claim GxP, 21 CFR Part 11, or any specific regulatory certification — it's a statistical modeling tool, and validation of its use within a regulated process is your organization's responsibility, same as any statistics software.
- It doesn't fabricate significance or curvature where the design doesn't support estimating it — see Section 2.
- Your experiment data is stored in your browser while signed out, or in our database tied to your account once signed in (so it follows you across devices); see the Privacy Policy for exactly what is stored where.
6. Questions
If you're evaluating BioDOE for a regulated or validated workflow and want more detail than this page covers — specific formulas, a worked example, or anything else — email info@biodoe.com.