PooledScreenAudit: auditing control-associated prediction in pooled variant-effect screens

Abstract

PooledScreenAudit 0.4.0 is an installable Python implementation for auditing control-associated prediction in pooled variant-effect screens. Starting from prepared variant–condition and residue-level matrices, it provides matched prediction with and without control features, rank-aware representation geometry, size-matched partition comparisons, and equal-feature predictability scoring. Outputs include split-level results, reports, execution parameters and checksums. This release adds a command-line interface, input validation and documentation around the unchanged numerical kernels of the earlier matrix analysis companion. It is a separate software distribution, not an update to PooledScreenID 0.1.2. Historical analysis materials and Source Data 1–5 are available in the analysis companion at https://doi.org/10.5281/zenodo.22707659. The deposit includes source code, tests, derived EGFR example inputs, an installable Python wheel, a README and SHA256 checksums. Python 3.9–3.12 is required, with NumPy restricted to the 1.26 series. Release verification used Python 3.9.6 on macOS and covered nine interface tests, numerical fixtures, EGFR predictability and geometry, 270 frozen EGFR model evaluations, SRC prediction and direction-subset analyses, and a fresh hash-checked retrieval and derivation of SRC inputs. Numerical comparisons with saved results passed an absolute tolerance of 1e-8. These checks establish same-implementation reproducibility, not independent scientific validation. Windows and Linux were not tested. Validation scope differs between commands. The generic screen workflow estimates direction bases from outer-training positions and holds them fixed during inner tuning. The frozen EGFR workflow re-estimates directions within inner-training folds, conditional on fixed upstream controls, responses and weights. Neither procedure represents fully nested reconstruction of the raw assay. Control association and representation overlap do not identify a causal technical mechanism; no universal separability threshold or patient-level treatment utility is claimed. Code is distributed under the accompanying MIT license. Derived experimental inputs retain their source attribution and terms, as documented in THIRD_PARTY_NOTICES.md. Raw third-party SRC counts are retrieved from their pinned source rather than redistributed. This deposit contains no private manuscript, cover letter or patient-level clinical data.