Calibrating gene-set nulls for imaging transcriptomics with planted truths

Background. Imaging-transcriptomic studies guard map–gene-set correlations with spatial and gene nulls, the strongest common gene null being coexpression-matched sampling. Neither has been calibrated with planted positives in the tissue-class and coherent-set regime where published claims sit. Methods. On Allen Human Brain Atlas expression, we planted maps and gene sets of known truth and measured false-alarm and catch rates under a frozen, amended preregistration. Results. Joint spatial surrogates were anticonservative where tissue classes differed in mean (false-alarm rate .078–.120); class-stratified surrogates were nominal. Coexpression-matched gene nulls were anticonservative for coherent sets (.196–.210; within a brain background, .399–.442), also in Cao et al.'s own regime, above the fitted coherence crossing for published-like sets. Ecology's fourth-corner max test, with constrained map-side and gene-side legs (α = .025), stayed ≤ .019 in all 390 verdict strata (gene-side leg alone ≤ .022 in 372 more), calibrated on every informative planted stratum within stated scope limits, but caught .585–.666 of strong gene-specific plants. On a preregistered register of published autism claims, the interim result (two of three release batches) is 0 of 16 population units surviving; the third batch's 196 units have not run and are counted neither way, and 9 rows (Dear et al., 3; Vairam, Roy et al., 6) are deferred. The one tested unit could not have survived (gene-side null uninformative; power unmeasured in its class). Conclusions. Gene-set nulls need planted-truth calibration in the coherence regime of the sets they test; a max test holds its level there, at measured cost in power. This record holds the preprint (PDF) and its archive: code, frozen preregistration and amendments, claims register, result files, figure and table scripts, and the manuscript source. Third-party brain maps are not redistributed. No individual-level or controlled-access data are included. Code is MIT-licensed; text, protocol, results and figures are CC BY 4.0.