A pre-registered prospective test of MEK inhibition in systemic sclerosis skin: criterion not met, and half the test statistic was a tied block

A single prospective, pre-registered test of an L1000 connectivity pipeline, run against its own author's interest and reported as a negative.

The pre-registration was committed with a SHA-256 hash before any disease signature or any score existed. It named two systemic sclerosis skin cohorts in advance, excluded by design the series that generated the hypothesis, and fixed a single criterion: a permutation p below 0.05 on the drug's mean percentile across the primary queries.

The criterion was not met, p = 0.354. The more consequential finding is why. The connectivity statistic returned exactly zero for 45.6 and 51.2 percent of the library in the two cohorts, which is what it returns when a query's up and down gene sets enrich in the same direction. That is a refusal to measure rather than a measurement of no effect. In one cohort the tied block holds 15 of the 18 compounds declared in advance, including all four reference compounds, so dasatinib, bosutinib, nintedanib and mycophenolate mofetil return the same number as one another. A ranking that cannot separate its own controls is not a ranking anyone should select candidates from.

A secondary analysis was reported, found to be confounded three minutes later, tested directly, and withdrawn in full the same day. The record of that withdrawal is deposited with everything else.

The deposit contains the pre-registration, five phase notebooks, the frozen inputs, the per-phase result records and a SHA-256 manifest over the whole tree. Each phase refuses to run unless the previous phase's output hashes to the recorded value.

Prepared with AI assistance. No AI system is listed as an author.