bsc-simulated-nsclc-sample

A sample of 500 simulated non-small cell lung cancer (NSCLC) patients on immunotherapy, with serial circulating tumor DNA (ctDNA) draws and imaging over two years. It shows the data format used to evaluate the BSC Engine.

Simulated data only. No row is a real patient, and the data contains no protected health information. Nothing here is suitable for clinical use.

What's included and what isn't

• Included: patient baseline characteristics and every visit observation (blood draws and scans).
• Not included: the ground truth (true resistance-onset day and progression status), the simulator, and the engine. Those are part of the BSC Engine, which Scalar Logic Group, LLC is offering for acquisition (info@scalarlogicgroup.com).

The patients were drawn at random from one of the two 10,000-patient reference cohorts behind the v8 results (preprint v8) and given new IDs.

Files

patients.csv — one row per patient (500 rows)

Column	Description
patient_id	Sample ID, BSC-SIM-0001 to BSC-SIM-0500
age_at_diagnosis	Age in years
pd_l1_expression_pct	PD-L1 tumor proportion score, %
treatment_regimen	Immunotherapy regimen

visits.csv — one row per visit (32,228 rows)

Column	Description
patient_id	Links to patients.csv
day	Days since treatment start (0–728)
visit_type	blood (every 14 days; about 5% of draws are missing, as in real follow-up) or imaging (every 56 days)
ctdna_vaf	Blood visits: mean variant allele fraction across 16 tracked variants. 0 means not detected (below a limit of detection of 0.01% VAF)
exhaustion_marker	Blood visits: simulated T-cell exhaustion marker, scaled 0–1
sod_rel	Imaging visits: sum of target-lesion diameters relative to baseline (1.0 = baseline)

How it was generated

• Tumor dynamics. A mechanistic simulator models competing tumor clones and immune exhaustion with a power-law memory kernel, so resistance builds up gradually rather than switching on at once.
• Assay noise. Measured ctDNA follows a proportional noise model: CV of 25%, 12%, 8% and 4% at 0.1%, 0.5%, 1% and 5% VAF, 16 tracked variants, and 4% draw-to-draw pre-analytical variation. Baseline VAF has a median of 1.2%, and 12.4% of patients are non-detectable at baseline. These noise parameters have not yet been confirmed with vendor validation data.
• Imaging. Scans record lesion size with measurement noise. Progression is typically called by a RECIST-style rule: at least 1.2 × nadir and at least 5% absolute.

Appropriate uses

• Testing pipelines, schemas and ingestion code for serial ctDNA and imaging data
• Exploring the data format before discussing the BSC Engine
• Teaching longitudinal liquid-biopsy data

Not appropriate for

• Benchmarking or validating a model (there's no ground truth here)
• Any clinical or diagnostic use

License

CC BY-NC 4.0: free for non-commercial use with credit to Scalar Logic Group, LLC. For commercial use, contact info@scalarlogicgroup.com.

Related

• Corrected preprint (v8): doi.org/10.5281/zenodo.23348706