cfDNA-Gen

Conditional Causal Transformer for Cell-Free DNA Sequence Generation

A 120M parameter transformer model that generates realistic synthetic cell-free DNA (cfDNA) sequences for NIPT simulation, benchmark development, and genomics research.

Model Description

cfDNA-Gen is trained on real cell-free DNA data and generates synthetic sequences with controllable properties:

• Fragment length: Control the length of generated fragments (typically 50-250bp)
• GC content: Target specific GC content (typical cfDNA: ~42%)
• Fetal fraction: Simulate different fetal fractions for NIPT applications (0-40%)

The model captures realistic patterns found in cfDNA including:

• Bimodal fragment length distribution (fetal ~144bp, maternal ~167bp)
• Nucleosome-associated 10bp periodicity
• Position-specific nucleotide preferences
• Characteristic end motifs

Architecture

• Parameters: 132M
• Layers: 14 transformer blocks
• Hidden dimension: 768
• Attention heads: 12
• FFN dimension: 3072
• Position encoding: RoPE (Rotary Position Embeddings)
• Activation: SwiGLU
• Normalization: RMSNorm
• Attention: Flash Attention (SDPA)

Usage

from cfdna_gen import CfDNAGenerator

# Load model
generator = CfDNAGenerator.from_pretrained("EabhaSeq/cfdna-gen")

# Generate sequences
sequences = generator.generate(
    n_sequences=100,
    fragment_lengths=165,
    target_gc=0.42,
    target_ff=0.10,
)

for seq in sequences[:5]:
    print(seq)

Installation

pip install cfdna-gen[hub]

Validation Results

Metric	Score
Overall Similarity	92.9%
Fragment Length Match	98.4%
GC Content Match	93.3%
Nucleotide Frequency	99.0%
Bimodal Peaks Detection	100%
Nucleosome Periodicity	100%

Use Cases

• NIPT Simulation: Generate synthetic samples with known conditions for algorithm development
• Benchmarking: Create standardized test datasets for cfDNA analysis pipelines
• Training Data: Augment real datasets for machine learning applications
• Method Development: Test new analysis methods on controlled synthetic data

Citation

@software{cfdna_gen,
  title={cfDNA-Gen: Conditional Causal Transformer for cfDNA Sequence Generation},
  author={Redelinghuys, Kyle},
  year={2025},
  url={https://github.com/eabhaseq/cfdna-gen}
}

License

This model is dual-licensed:

• Free for academic, research, educational, and personal non-commercial use under the PolyForm Noncommercial License 1.0.0.
• Commercial use requires a separate paid license. Contact kyle@eabhaseq.com for details.

See the LICENSE file for full terms.