Design an experiment with the statistics settled first

What this involves

A well-planned experiment starts with a hypothesis formulated to be falsifiable: stating direction, magnitude, a specific comparator, and the explicit result that would cause you to abandon the hypothesis. Rather than running a fixed schedule of procedures, the protocol is structured in stages so that resource-intensive experiments are only reached by samples that pass early computational and low-cost filters.

Settling the statistics in advance means fixing one primary outcome and specifying the smallest effect size worth detecting before calculating the required biological sample size. Setting sample size by habit, such as testing triplicate wells and assuming statistical power, leads to studies that can only detect massive effect sizes while conflating technical variance with biological replication.

On this platform, the research agent evaluates designs, runs power calculations and randomisation scripts, screens molecules and sequences with dedicated computational tools, and sets out complete analysis protocols without requiring local software installation.

How it goes

1. 1Define the falsifiable hypothesis and decision criteriaFormulate the hypothesis with an explicit direction, magnitude, and comparator, define the precise result that falsifies it, and verify that different experimental outcomes will dictate different downstream decisions.
2. 2Filter candidates through low-cost computational triageOrder early stages by cost and discrimination using molecule_properties for property filters, search_papers for liabilities, and analyze_sequence, lookup_structure, or fold_protein to check targets before committing to assays.
3. 3Calculate sample size from the minimum relevant effect sizeSpecify a single primary outcome and the smallest meaningful effect size, then compute sample size n for biological replicates using TTestIndPower from statsmodels.stats.power via run_python.
4. 4Assign controls and generate randomised allocationsDefine vehicle, positive, specificity, and batch controls alongside assay-specific interference controls, then produce reproducible allocation schedules with allocate_experiment using unique unit identifiers and explicit seeds.
5. 5Preregister the statistical analysis planFix the statistical test, outlier definitions, and multiple testing corrections using scipy and statsmodels.stats.multitest, establishing whether to apply Holm for confirmatory tests or Benjamini–Hochberg for discovery screens.

Start it here

Ask in plain language. The agent works on a Linux machine of its own — it installs and runs what the job needs, keeps every file, and reads the results back with the numbers explained. Nothing to set up, and a long run continues after you close the page.

What to watch for

• Never determine the target effect size from pilot data or the very data being tested, as small pilots severely overestimate effect sizes.
• Do not treat technical replicates as biological replicates; average technical repeats across instrument wells first so biological sample size n reflects truly independent units.
• Do not place controls along microplate edges where evaporation generates edge effects, and always randomise position and processing order.
• Never calculate or report post-hoc observed power after an experiment, as it is purely a function of the obtained p-value and provides no new information.
• Avoid running Shapiro–Wilk normality tests to pick between parametric and non-parametric tests, and do not use plain Bonferroni corrections when Holm strictly dominates them.

Questions