Available Now

How we
tested it

Locked claims only: AUC 0.87 on real karyotyped T21, and T21/T18/T13 at 8% FF on a purely synthetic WisecondorX reference panel. Labelled FASTQ for caller testing.

0.87AUC vs real T21Train synthetic, test on karyotyped samples
8%WisecondorX FFT21, T18, T13 on a purely synthetic euploid panel
p = 0.002Permutation test1,000 resamples
26Unique clinical samples6 T21 + 20 euploid, Lun et al. 2014

Validated against 26 real clinical cfDNA samples (6 T21 patients, 20 euploid) from Lun et al. 2014, PRJNA215135.

Detecting Real T21 With Synthetic Training Data

We trained a classifier entirely on synthetic cfDNA — zero real patient data in training. Then tested it blind on 26 real clinical samples from an independent dataset (Lun et al. 2014, PRJNA215135), including 6 karyotype-confirmed trisomy 21 cases.

Per-Patient Results

Four of six real T21 samples sit below a z-score of 3.0 on a standard reference (NIPD-66 is a known ~1–2% FF anomaly). The table is a TSTR result, not a product claim that these files beat clinical NIPT.

PatientZ-Score (real ref)Standard NIPTTSTR Detection
NIPD-035.45DetectedAbove all euploid
NIPD-603.43BorderlineAbove all euploid
NIPD-072.47FAILAbove all euploid
NIPD-501.57FAILAbove all euploid
NIPD-041.29FAILBelow 2 euploid
NIPD-660.86FAIL (anomaly)Below 14 euploid

Summary Metrics

MetricValueNotes
TSTR AUC0.8676 T21 vs 20 euploid, trained entirely on synthetic
Adjusted AUC0.980Excluding NIPD-66 (genuine anomaly, FF ~1-2%)
Augmented AUC0.942Real + synthetic outperforms real-only (0.883) by +5.8pp
Permutation p-value0.0022/1000 random label shuffles matched observed AUC
Low-z T21 rankingSee tableTSTR ranking on this cohort, not a claim that files beat clinical NIPT

Key Finding

NIPD-66 is a genuine anomaly — chr21 fraction only 0.027pp above the euploid mean, implying fetal fraction ~1-2%. This sample would fail any chromosome-fraction NIPT method. Excluding it, the adjusted AUC is 0.980 and 4 of 5 T21 samples rank above all 20 euploid samples.

Detection on a synthetic panel

Fetal fraction was 8%. Fifty synthetic euploid samples form the reference panel — not a clinical panel. Do not read this as 100% on the public 1M ZIP or at 2% FF.

ConditionSamplesZ-Score RangeDetection
Trisomy 2134.7 - 6.5Rank 1 (all)
Trisomy 1836.2 - 8.2Rank 1 (all)
Trisomy 1337.5 - 9.0Rank 1-2
Euploid controls50-1.2 - 1.1No false calls on this panel

Synthetic Data Rescues Low-Data Regimes

When real training samples are scarce — the reality for rare aneuploidies — adding synthetic data dramatically improves classifier performance. 100 random stratified splits per condition.

N Real SamplesReal-Only AUCAugmented AUCImprovementWin Rate
30.6610.858+0.19788%
50.6910.832+0.14183%
70.7250.887+0.16387%
100.7680.877+0.10975%
200.8980.944+0.04626%

Practical Impact

With only 3 real training samples, synthetic augmentation improves AUC by +19.7 points (0.661 → 0.858) and wins 88% of random splits. Augmented models also have consistently lower variance — synthetic data stabilises training.

4M T21 on the live engine

Same generator as paid. 4M fragments at 8% fetal fraction.

Paired-end FASTQ (ZIP)Live-engine artifactTrisomy 21
Download Sample

Choose Your Data Package

Self-serve aneuploidy starter pack at $399, deeper four-file packs, or catalogue/generate on the dashboard.

Catalogue & generate

Conditions beyond the starter pack, self-serve on the dashboard

Self-serve

  • Same generation path as the starter pack
  • Catalogue reuse or generate a new file
  • Questions: kyle@eabhaseq.com

How we tested it

Our synthetic cfDNA has been validated end-to-end: generated sequences are aligned to GRCh38, then tested against real karyotype-confirmed clinical samples.

01

Alignment & Realism

Insert size peaks at 166bp (mono-nucleosome). Alignment yield depends on the generator path; do not treat a single MAPQ percentage as a product spec.

See alignment notes
02

Train Synthetic, Test Real

Classifier trained entirely on synthetic data detects real karyotyped T21 samples. Tested on 26 unique clinical samples from PRJNA215135 (Lun et al. 2014).

AUC = 0.87, p = 0.002
03

WisecondorX on a synthetic panel

WisecondorX analysis detects all synthetic trisomies (T21, T18, T13) at 8% fetal fraction. Augmenting real data with synthetic improves AUC by up to +19.7 points.

T21/T18/T13 at 8% FF, purely synthetic panel
04

Statistical Rigour

Permutation testing (1,000 resamples) and bootstrap CIs confirm results are not due to chance. Low-data augmentation tested across 100 random splits per condition.

95% CI: [0.60, 1.00]

What You Can Build

Algorithm Development

Train and validate NIPT detection algorithms with labelled synthetic data. Test edge cases like low fetal fraction that are rare in real datasets.

Synthetic reference panels

WisecondorX detected T21/T18/T13 at 8% FF using a synthetic euploid reference panel — not a clinical panel or a physical control.

Pipeline verification

Drop FASTQ into the analysis step. This does not exercise extraction or the sequencer.

Privacy-Compliant Research

Conduct research without patient data concerns. Synthetic data contains no identifiable information.

Education & Training

Train clinical scientists and bioinformaticians with realistic data. Perfect for courses and workshops.

Benchmark Creation

Create standardised benchmarks with known ground truth for comparing NIPT methods across laboratories.

What's Included

ParameterStandard DatasetCustom Generation
Fragments per sample4M (starter)4M–32M
Fragment length range50-250 bp50-250 bp
Fetal fraction8% default1% - 25%
ConditionsT21, T18, T13 + euploid controls107 including SCAs, microdeletions, microduplications, monogenic, oncology
Output formatpaired-end FASTQpaired-end FASTQ
Minimum order1 pack1 sample
MetadataJSON per sample (FF, condition, params)Full provenance tracking
Ground truth labelsYesYes
DeliverySecure download linkSecure download

Order the files, or start with the free sample

Self-serve $399 starter pack. Catalogue and generate on the dashboard.