Generating realistic artificial human genomes using adversarial autoencoders.

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Publicly available human genomes offer research value but pose privacy risks due to potential exploitation.
• Existing AI models for synthetic genomic data face challenges with large datasets and computational demands.

Purpose of the Study:

• To develop an efficient AI-based method for generating high-fidelity synthetic human genomes.
• To ensure synthetic data preserves scientific merit while protecting individual donor anonymity.

Main Methods:

• A dimension reduction technique combining AI with knowledge of in vivo mutation association mechanisms.
• Genome segmentation based on chromosomal recombination hotspots to mimic mutation transmission.
• Training variational autoencoders with a Wasserstein GAN using data from the 1000 Genomes Project.

Main Results:

• The developed pipeline efficiently processes large genomic datasets without substantial computational resources.
• Generated synthetic populations are diverse, realistic, and exhibit human-like linkage disequilibrium.
• The synthesized genomes differ from reference samples, effectively anonymizing individual donors.

Conclusions:

• The AI-driven dimension reduction method offers an efficient solution for synthetic genome generation.
• This approach balances the need for scientifically valuable genomic data with robust patient privacy protection.
• The method successfully generates anonymized, realistic synthetic human genomes suitable for research.