Search-Space Limits: Why Labs Run Few Experiments

With three factors at five levels each, a complete DOE requires 125 runs, while a typical lab may perform only 8 to 12. That leaves 113 of 125 combinations unobserved. The mismatch grows when a problem includes more factors or continuous parameter ranges: the reachable region and the sampled region become substantially different. This is not a minor accounting discrepancy. It is the structural limit on how much scientific discovery can come from a finite experimental budget.

This search-space constraint occurs across drug discovery, materials optimization, synthetic biology pathway engineering, and process chemistry. The details differ, but the arithmetic remains: a laboratory can execute only a small fraction of the experiments needed to characterize the relevant space.

The Numbers

Consider a drug discovery program optimizing a lead compound across 10 structural parameters, including scaffold choices, substituent positions, and functional group variants. A thorough search may require 10^8 to 10^10 compounds, depending on parameter granularity. A medium-throughput synthesis effort might make 200 to 500 compounds per year. At that pace, comprehensive coverage is not a practical target. The portion of relevant space sampled is measured in fractions of a percent, not 10% or 1%.

Materials discovery has the same scaling problem. A five-element alloy with continuous concentration variables has mathematically infinite compositional space, though synthesis constraints impose practical bounds. A project may test 50 to 200 compositions, while the stable, processable region still contains thousands of untested compositions for each one synthesized and characterized.

These are ordinary search-space sizes for discovery work in these fields, not unusually selected examples.

Why Labs Run Few Experiments

The throughput ceiling reflects physical constraints, not a lack of effort. Wet-lab work consumes reagents, consumables, and instrument time; it also includes synthesis, incubation, characterization queues, skilled execution, and data interpretation before the next round can be specified. Automation has improved each constraint, but not enough to make comprehensive coverage of large parameter spaces feasible.

High-throughput screening has raised capacity for specific experiment classes, especially biochemical assays in microtiter plates. A fully equipped HTS facility can test hundreds of thousands of compounds against a target in a few weeks. That addresses the assay step only, and it assumes the compounds are already present in a screening library. For most discovery workflows, synthesis capacity and the cognitive work of interpreting screening data remain limiting steps.

The candid assessment is that standard assay throughput has risen by roughly one order of magnitude over the past two decades. The spaces we seek to explore have expanded faster, as biological understanding reveals more relevant parameters and computational tools make large candidate sets easy to enumerate. Sampling coverage is therefore not catching up; the gap is increasing.

DOE: Structured, Yet Still Constrained

Classical Design of Experiments addresses coverage by selecting a structured subset that extracts information about the response surface with fewer runs. For three factors, a central composite design needs roughly 15 to 20 experiments rather than 125. That reduction is substantial.

DOE is effective when the response surface is smooth, the important factors are known beforehand, and the task is characterization rather than optimization. Scientific discovery often violates all three conditions. Biological surfaces can be nonlinear and multimodal, the relevant factors may be unknown at project outset, and the practical objective is usually to find the best condition rather than map the complete landscape.

We are not saying DOE should be discarded. Where its assumptions hold, it remains an efficient way to allocate experimental budget. Its limitation is the static workflow it assumes: define the space, design the runs, execute them, and analyze the results. Discovery is often dynamic. New evidence can elevate a supposedly irrelevant factor, alter the space of interest, or produce a hypothesis that points to a region absent from the original design.

What Adaptive Search Changes

Active learning, particularly Bayesian optimization, changes search economics by using each result to choose the next experiment. Rather than fixing the entire design in advance, the system updates its response-surface model after every observation and selects the run with the greatest expected information gain or expected improvement over the current best result.

In controlled optimization benchmarks, Bayesian optimization reaches results within 10% of the global optimum using 3 to 5 times fewer experiments than random search or a full factorial design. With a 100-experiment budget in a 1000-point space, random search samples 10% of the space and finds the global optimum roughly 10% of the time. Bayesian optimization instead concentrates runs where predicted value is high and finds the optimum far more reliably. That difference can decide whether a project succeeds.

Those figures come from controlled benchmarks. Wet-lab work introduces more noise, longer cycle times, and response surfaces less orderly than benchmark functions. The gains remain meaningful, but are smaller than idealized results imply. We state that limitation plainly when labs ask what improvement to expect.

The Coverage Tension

Adaptive search exchanges comprehensive coverage for more efficient optimization. A Bayesian optimization loop is designed to find a good solution efficiently, not to explain the entire search space. Those objectives are distinct, and the distinction can matter.

When the objective is a global optimum, such as maximizing reaction efficiency, efficient optimization is appropriate. When the objective is to map how a protein class responds across a ligand family for SAR, comprehensive coverage may matter more, even if that allocation is less efficient at finding the single strongest result.

We design Discovery Loop's acquisition policy for both modes. Optimization directs budget toward regions predicted to have high value. Exploration spreads experiments more broadly to reduce uncertainty across the space. Most projects use both: early cycles explore to establish broad structure, then later cycles optimize once promising regions are reasonably clear.

None of this removes the underlying search-space constraint. The possible spaces remain larger than any reasonable experimental budget can cover comprehensively. Active search improves the expected quality of results obtained within that budget, not the budget's fundamental limit. That is useful, but it is not a complete answer to coverage. The system still operates within a finite number of experimental runs.