Code, Summary Results, and Figures for Evaluating Chemical–Target Pairing in Retrospective Regulatory Organ Ranking

This repository accompanies a retrospective regulatory toxicology benchmark evaluating how observed chemical–target pairing contributes to organ-ranking performance. The benchmark draws on 501 U.S. EPA Office of Pesticide Programs (OPP) studies and 214 National Toxicology Program (NTP) studies, using 14 shared organs, eight network scoring functions, and HPA- and GTEx-derived tissue representations. Within-study concordance compares treatment-related pathology-positive organs with officially tested-negative organs; untested organs are excluded from the negative class. The primary analysis evaluates an overlap-based organ score against fixed-margin chemical–target randomization, preserving chemical target-set sizes and target popularity. Across four regulatory-source/tissue-representation settings, the primary concordances were 0.670457 (OPP/HPA), 0.644340 (OPP/GTEx), 0.580324 (NTP/HPA), and 0.576452 (NTP/GTEx). The prespecified upper-tail permutation tests did not detect an advantage for observed chemical–target assignments in any setting. These tests do not establish equivalence or the absence of chemical-specific biological information. Version 1.1 adds exploratory randomization analyses of the highest-ranked STRING bridge-based scoring functions. The bridge scores likewise showed no observed-pairing upper-tail advantage under the tested degree-preserving double-edge-swap and Curveball randomization procedures. These comparisons were selected after examining score rankings and are not confirmatory primary tests. Sensitivity analyses examined trading budgets, independent random seeds, and chemical-level clustering. The finite-budget null distributions exhibited some method sensitivity: for NTP/GTEx primary overlap, the reported upper-tail p value was 0.1229 under the original double-edge-swap implementation and 0.0620 with 1,500 Curveball randomizations at 50 trades per chemical. Neither result fell below 0.05; neither demonstrates exact stationary sampling of the fixed-margin null. The archive provides the original verification scripts, aggregate results, retained per-randomization summary statistics, updated main Figures 1–4, supplementary sensitivity figures, editable sources for revised Figures 2 and 3, versioned provenance records, and instructions to recalculate publicly archived summaries. Version 1.1 replaces Figures 2 and 3 with corrected versions and retains Figures 1 and 4. The project identifier NTX-BENCH-01 is retained in the documentation for traceability. Complete source-level reproduction requires independently obtained, authorized, version-matched external inputs. The archive does not redistribute underlying regulatory pathology records, raw ToxCast data, original HPA or GTEx expression matrices, STRING source networks, or the full historical source-data extraction pipeline. Relevant sources include ToxRefDB v3.0, ToxCast invitroDB v4.3, HPA v25.1, GTEx Analysis V10, and STRING v12.0. The provided code and results support verification of the deposited summaries but are not a fully self-contained re-creation of every upstream data-processing stage. This record is a new version of the repository first published at https://doi.org/10.5281/zenodo.23776204. The all-versions DOI is https://doi.org/10.5281/zenodo.23776203. Cite the version-specific DOI assigned to version 1.1 after its publication.