De-identified dataset, Interpretation Rulebook and Python analysis code: genotype-derived interpretive categories of drug response in 908 pharmacogenomic reports from Chinese patients with type 2 diabetes

This record contains the de-identified individual-level dataset, the Interpretation Rulebook in machine-readable form, and the complete Python analysis code for the study "Genotype-derived interpretive categories of drug response in 908 pharmacogenomic reports from Chinese patients with type 2 diabetes: a retrospective single-centre study". CONTENTS. data/Database_deidentified.csv is the de-identified dataset: 908 issued report records from 887 patients, with age, sex, the genotypes printed on each report, and the eight interpretive categories as issued by the testing laboratory. data/DATA_DICTIONARY.md describes it column by column. rulebook/ holds the Interpretation Rulebook and its precedence rules as CSV, with an evidence-level column and a provenance column marking which rules were taken from the laboratory's report appendix and which were reconstructed. scripts/ holds the analysis code; running scripts/00_run_all.py regenerates every table in derived/ and every figure in figures/ from data/ alone, in about 35 seconds. MANIFEST.csv gives the sha256 of every file. PRIVACY. No name, laboratory accession number, medical record number, test date or referring clinician is included. Records carry sequential identifiers R0001-R0908, and Patient_key (Q0001-Q0887) marks which records came from the same person without being resolvable to one; the linking key is held by the study team and is not published. The study was approved by the Medical Ethics Committee of Zaozhuang Mining Group Central Hospital (approval number 2022-Ethics Review-02, the certified English rendering of the Chinese original 2022-伦理审查-02; granted 27 April 2022), which waived the requirement for individual informed consent for this retrospective analysis of de-identified records. NOTES ON THE DATA. The category columns hold the interpretation as issued in the original clinical report, not a value recomputed from the genotype. Across the 908 records there are 5,983 category assignments, of which 5,940 (99.3%) agree with the rulebook; 30 records (3.3%) depart from it in some way, and every departure is classified by mechanism in derived/S6_departures_from_rulebook.csv. They were deliberately left as issued, because the issued interpretation is what the treating clinician would have seen. The cohort was tested on two successive panel versions, a four-locus panel (427 records) and a six-locus panel (481 records), so missingness at rs4149056, rs7756992 and rs2066865 is structural rather than random, and the two versions applied different rules to sulfonylurea efficacy. Six loci were assayed per patient; seven distinct loci appear across the two versions. ANALYSES. Concordance audit of all eight interpretive categories against the rulebook, with Cohen's kappa and a six-class mechanistic classification of every departure; genotyping quality control; genotype-to-category associations with explicit additive allele-dose coding and Firth penalised likelihood; a weighted polygenic score reported as a consistency index rather than as a measure of discrimination; and eight pre-specified gene-gene interaction tests with Bonferroni correction. REQUIREMENTS. Python 3.9 or later with pandas, numpy, scipy, statsmodels, scikit-learn and matplotlib. CHANGES FROM VERSION v2. The v2 scripts coded genotypes by alphabetical sort order of the genotype string, which inverted the additive dose direction at CDKAL1 rs7754840 and SLCO1B1 rs4149056. Two published results change as a consequence: the odds ratios at those two loci now run in the direction the rulebook assigns, and the CDKAL1 x CDKAL1 interaction on composite T2DM risk, previously reported as synergistic with OR 8.21 and P = 0.005, is sub-multiplicative at OR 0.027 (95% CI 0.0023 to 0.318, likelihood ratio P = 0.033) and meets no Bonferroni threshold; that direction is an arithmetic consequence of assigning the composite category by a Boolean OR over contributing loci, not evidence of epistasis. Anyone re-running the v2 scripts will obtain superseded numbers, so they are not redistributed here; they remain citable at 10.5281/zenodo.21634104. This version also adds the machine-readable rulebook, the derived tables, the figures and MANIFEST.csv; adds Record, Patient_key and Panel columns to the dataset and removes the non-unique Sample_ID column; and extends the concordance audit from four categories to all eight. No genotype or category value was altered. See CHANGELOG.md.