Definitive Screening Designs (DSD), Explained

Part of our complete guide to DOE for biologists.

A standard screening design has a blind spot: it tests every variable at just two levels, low and high, so it can tell you a variable matters without ever being able to tell you the relationship curves. Finding that out has traditionally meant a second experiment — a characterization design with extra center points, run after screening is already done. A Definitive Screening Design (DSD) closes most of that gap in a single pass.

What it actually adds

A DSD tests each variable at three levels instead of two — low, center, and high — but does it in roughly the same number of runs a standard screen would need, not the much larger number a naive three-level design would cost. That third level is what lets it catch curvature directly: if a variable's effect isn't a straight line between low and high, a DSD has a chance of seeing that during screening instead of only finding out later.

The construction is a real piece of combinatorics (it's built from something called a conference matrix, mirrored and folded around a single center run), but you don't need any of that to use one — the run count and structure are what matter, and both come out looking almost identical to a screening design you'd already recognize, just with a center point built in for every factor.

Where it fits against the other options

Which DOE Design Should I Use? covers the three-way choice of screen, characterize, or optimize. A DSD doesn't replace that choice — it's a better way to build the screening option specifically, when your factor count makes it a genuinely good trade. It isn't a replacement for characterization or optimization either: it tells you which factors curve, not the exact shape of that curve or where a peak sits. If you find curvature and need the precise optimum, you still move on to a response surface design from there.

When BioDOE uses one

You never have to ask for a DSD by name. When you pick "Find what matters" (screening) in BioDOE, it checks whether a Definitive Screening Design is available for your exact factor count without costing you meaningfully more runs than a standard screen would. If it is, that's what gets generated — no extra step, no new terminology in the interface. If your factor count doesn't line up well with one, BioDOE falls back to the same Resolution IV fractional-factorial screening it would otherwise use, and you lose nothing by that fallback existing: it's the same design you'd have gotten before DSD support existed.

The honest limitation

DSDs support a specialized analysis technique in the statistical literature that can tease apart which specific two-factor interactions and curvature terms are active, even when several are confounded with each other. BioDOE doesn't implement that specific technique yet — it analyzes DSD results the same way it analyzes every other design, fitting the richest model your data can actually support. That's a legitimate, honest fit; it's just not the most statistically sophisticated option available for this particular design type. In practice this mainly matters if you have several curving factors at once and need to know precisely which interactions are driving it, rather than just that curvature is present.

Start free — screening designs, DSD included, are on BioDOE's free Starter plan.