Learning inherent genetic patterns and trait associations with deep generative models for discrete genotype...

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Deep generative models offer novel methods for realistic genomic data simulation.
• Existing models primarily focused on gene expression or haplotype data.
• Generating discrete genotype data presents unique challenges.

Purpose of the Study:

• To explore deep generative models for genotype data simulation.
• To evaluate models in unconditioned and phenotype-conditioned settings.
• To adapt models for the discrete nature of genotype data.

Main Methods:

• Developed and evaluated Variational Autoencoders (VAEs), Diffusion Models, and Generative Adversarial Networks (GANs).
• Adapted models for discrete genotype data characteristics.
• Conducted experiments on large-scale cow and human genotype datasets.

Main Results:

• Models effectively captured genetic patterns in genotype data.
• Genotype-phenotype associations were successfully preserved.
• Performance was assessed using deep learning and quantitative genetics metrics.

Conclusions:

• Deep generative models can accurately reproduce genotype data characteristics.
• These models facilitate privacy-preserving genomic data sharing.
• Findings offer guidance for genotype simulation research and applications.