Validation
Methodology & Results
Train on synthetic, test on real. A classifier trained only on our synthetic data reached AUC 0.87 for T21 on one published dataset (Lun et al. 2014: 26 unique samples, 6 T21 and 20 euploid; p = 0.002). That is a TSTR result on this cohort, not a claim that these files beat clinical NIPT.
1. Validation Overview
A classifier trained only on our synthetic data reached AUC 0.87 for T21 on one published dataset (Lun et al. 2014: 26 unique samples, 6 T21 and 20 euploid; p = 0.002). This is the only validation result we claim.
2. Train-on-Synthetic, Test-on-Real
The definitive test: train a classifier using only synthetic data, then evaluate it on real clinical samples with confirmed karyotypes. No real data is used during training.
Methodology
Generate Synthetic Training Data
50 euploid + 3 T21 samples at 8% fetal fraction, 2M read pairs each. Production orders use ReferenceConditionedCfDNACausalLM / ModelSequenceGenerator → hg38 → injection.
Align and Extract Features
bwa mem alignment to GRCh38, extract 22 autosomal chromosome read fractions from idxstats.
Train Logistic Regression
Logistic regression classifier trained on 22 autosomal read fractions from synthetic data only. No real data seen during training.
Test on Real Clinical Samples
26 real clinical cfDNA samples (Lun et al. 2014, PRJNA215135): 6 unique T21 patients (karyotyped), 20 euploid. 3 TMR replicates excluded to prevent data leakage.
A sample is classified as T21 if z_chr21 > 3.0. The reference panel (mean and standard deviation) comes entirely from synthetic euploid samples.
Result
A classifier trained only on our synthetic data reached AUC 0.87 for T21 on one published dataset (Lun et al. 2014: 26 unique samples, 6 T21 and 20 euploid; p = 0.002).
Fetal Fraction Sweep
Synthetic training data was generated at 5 fetal fraction levels to cover the clinically relevant range:
| Fetal Fraction | Samples | Fragments | Notes |
|---|---|---|---|
| 4% | 3 T21 + 3 euploid | 100,000 | Near clinical lower limit of detection |
| 6% | 3 T21 + 3 euploid | 100,000 | Low but detectable range |
| 8% | 3 T21 + 3 euploid | 100,000 | Typical clinical FF |
| 10% | 3 T21 + 3 euploid | 100,000 | Moderate FF |
| 12% | 3 T21 + 3 euploid | 100,000 | Higher FF, strong signal |
3. Methodology Notes
Reference Data
Calibration uses 9 euploid cfDNA samples from PRJNA756388, aligned to GRCh38. Validation uses 26 unique samples from Lun et al. 2014 (PRJNA215135): 6 karyotype-confirmed T21 and 20 euploid, after excluding TMR technical replicates. Sequenced on HiSeq 2000 (50bp single-end, ~30M reads).
Synthetic Generation Pipeline
Production samples use ReferenceConditionedCfDNACausalLM / ModelSequenceGenerator → hg38 → injection so reads stay on-locus after BWA. Key parameters:
- Target fetal fraction (1-25%)
- Karyotype condition (see the conditions page for what is currently available)
- Fragment count per sample (2M read pairs for validation)
- Transition-biased error profiles (Ti/Tv ≈ 2.0)
Alignment Pipeline
Generated sequences are written as paired-end 150bp FASTQ, aligned with bwa mem to GRCh38, sorted and indexed with samtools. Chromosome fractions extracted from idxstats. The full pipeline (generate → FASTQ → BAM → idxstats → analysis) mirrors clinical NIPT workflows.
Limitations
Current Limitations
1. Single cohort validation: TSTR tested against one published dataset (26 unique samples, 6 T21 and 20 euploid). Results may vary with larger or more diverse cohorts.
2. Coverage baseline offset: Synthetic and real chromosome fractions have a ~0.2pp systematic offset. Reference panel z-scoring eliminates this, but it means raw fractions are not directly interchangeable.
3. Single ancestry: Current validation uses primarily European-ancestry samples. Multi-ancestry validation is planned.
Reproducibility
All validation scripts, data, and results are available upon request. Contact kyle@eabhaseq.com for access to validation code and datasets.
Back to Synthetic Data ProductsReady to Get Started?
Contact us to discuss your data needs or request access to our validation datasets.