AuricGrid-Laboratory/E2.P2-Drug_Response_Remodelling_Cohort

Auric Grid Synthetic Drug Response Cohort for Advanced NSCLC

100% SYNTHETIC PATIENT-LEVEL DATA — NO REAL PATIENT RECORDS — RESULTS ARE SIMULATIONS, NOT CLINICAL EVIDENCE

• Dataset Title: Auric Grid Synthetic Drug Response Cohort for Advanced Non-Small Cell Lung Cancer
• Dataset ID: AG-NSCLC-DRM-001
• Organization: Auric Grid Laboratory
• Data type: Synthetic patient-level longitudinal data
• Patients: 100,000 synthetic patients
• Real patient data: None. No PHI, no identifiers, no copied records.

Intended uses

Machine-learning development, statistical modelling, drug-response simulation, clinical-trial methodology research, biomarker-modelling method development, software/pipeline testing, education, and reproducibility benchmarking.

Prohibited interpretation

This dataset must not be used to claim, imply, or support:

• real-world drug efficacy of AG-DRUG-A, AG-DRUG-B, AG-DRUG-C, or AG-SOC
• real-world drug safety conclusions
• real-world biomarker-response effectiveness
• clinical treatment recommendations of any kind

All treatment identifiers are fictional. All relationships between variables are synthetic constructs generated from documented, seeded statistical rules — see METADATA/ground_truth_parameters.md.

Provenance

synthetic_data_version: v2.0
generation_seed: 20260814
generation_date: 2026-08-14
pipeline_script_hash: 81105c2c28ffa159831d42ea034fd5047fd59ee60a360ebaa533684ebd759489

Directory structure

AG-NSCLC-DRM-001/
├── README.md
├── LICENSE.md
├── DATA_DICTIONARY.csv
├── DATA/                     10 relational CSVs, 100,000 patients
├── ML_SPLITS/                 deterministic patient-level 70/15/15 train/val/test
├── BENCHMARK_SCENARIOS/       baseline, missing_data, treatment_imbalance,
│                              biomarker_imbalance, temporal_shift, data_leakage
├── VALIDATION/                quality_report.md, distribution/missingness reports,
│                              ground_truth_recovery_report.md
├── ANALYSIS_EXAMPLES/         descriptive, response-prediction, survival, biomarker,
│                              subgroup analysis starter notebooks/scripts (see note below)
├── PIPELINE/                  generate_cohort.py, validate_cohort.py, model_specifications.md
└── METADATA/                  generation_methodology.md, ground_truth_parameters.md,
                                random_seed.txt, version.txt, changelog.md

Quick start

import pandas as pd
patients = pd.read_csv("DATA/patients.csv")
survival = pd.read_csv("DATA/survival_outcomes.csv")
splits = pd.read_csv("ML_SPLITS/patient_split_assignment.csv")

Data tables (DATA/)

File	Grain	Rows
patients.csv	1 row / patient	100,000
disease_characteristics.csv	1 row / patient	100,000
biomarkers.csv	1 row / patient / biomarker (9 markers)	900,000
baseline_labs.csv	1 row / patient / lab, MCAR-dropped rows omitted	~1,439,931
treatments.csv	1 row / patient	100,000
longitudinal_observations.csv	1 row / patient / visit	~578,340
tumor_response.csv	1 row / patient / post-baseline visit	~484,348
adverse_events.csv	1 row / adverse event	~179,913
progression.csv	1 row / patient	100,000
survival_outcomes.csv	1 row / patient	100,000

Full field-level definitions: DATA_DICTIONARY.csv.

Ground truth & validation

• METADATA/ground_truth_parameters.md — every coefficient/hazard ratio used to generate response, PFS, and OS, kept separate from DATA/ for benchmarking.
• VALIDATION/quality_report.md — structural, temporal, clinical-logic, and statistical checks (all passed on this v2.0 run).
• VALIDATION/ground_truth_recovery_report.md — confirms a standard logistic regression and group-wise PFS comparison recover the correct direction of the built-in effects.

Regenerating this dataset

python3 PIPELINE/generate_cohort.py     # builds DATA/, ML_SPLITS/, BENCHMARK_SCENARIOS/, ground truth
python3 PIPELINE/validate_cohort.py     # builds VALIDATION/

Fully reproducible from generation_seed: 20260814.

Note on ANALYSIS_EXAMPLES/

This v2.0 release ships the benchmark-ready data, splits, scenarios, and validation reports. ANALYSIS_EXAMPLES/ contains placeholder READMEs pointing to the relevant DATA/ and ML_SPLITS/ files for each analysis type; worked example code was out of scope for this run — see METADATA/changelog.md for what's deferred to a future version.