cdtest: A Dependence-Aware Feature-Randomization Test for Relating Gene-Level Chromatin Features to Context-Dependent Perturbation Responses

The same CRISPR interference knockdown changes expression differently in different cell lines. Testing whether a gene-level feature such as chromatin accessibility explains these differences is difficult because neighbouring genes share regulatory elements and copy-number segments, so genes are not independent; in simulation, standard gene-level tests reached false-positive rates of up to 25.8% at a nominal 5% level. cdtest is a Python package that keeps the response fixed and permutes the feature in genome-ordered blocks, with chromosome-arm adjustment, noise-corrected response differences, test-inversion confidence intervals and planted-signal power benchmarks. On real outcomes it held the nominal level under a decoy-free null (4.9% at 5%), and all confirmatory results can be recomputed in under a minute from deposited intermediates. Applied to essential-gene screens in four human cell lines, it detected planted effects of 0.5% partial variance explained with at least 80% power in a matched pair, found no detectable accessibility effect in that pair, and flagged an association that depended on the choice of signal track.