Synthetic MSK-IMPACT 2017 cohort: audited clinical-variant tables and simulated FASTQ reads from a hybrid generative pipeline

No real patient data is included.                                                                                                            Contents:                                                                              (1) synthetic_msk_impact_2017_cohort.csv — 10,328 synthetic records in a 9-column clinical-variant schema (age, gender, pathological diagnosis, sample type,    TMB, MSI status, QC quality score, somatic variant list, clinical-audit stage).                                            (2) synthetic_msk_impact_2017_reads.tar — simulated sequencing data for 2,000 synthetic patients, each with 200,000 single-end 100 bp FASTQ reads (0.2× WES    depth, GRCh38 full analysis set + decoy + HLA reference), generated with synggen v1.6 using seed = 42 + patient_index for exact reproducibility.            (3) Technical-validation artifacts — statistical fidelity, TSTR downstream utility, privacy assessment, and high-coverage VAF recovery (Pearson r = 0.851) for   both the 2017-derived open cohort and an independent 50K-patient extension cohort.                                                                                                                             Pipeline: tabular generative model → rule/LLM clinical audit → biophysical read simulation.

Privacy: zero exact record matches to the source cohort;        distance-to-closest-record leakage 0.0; membership-inference AUC ≈ 0.68 is reported honestly as a caution for downstream ML reuse. Integrity: every FASTQ file    was individually audited (read count == 200,000) and MD5-manifested (MANIFEST.md5).