Applying generative adversarial networks to generate artificial genotype data in livestock

Caballero Vargas, Edgar Giesus

Permalink

https://hdl.handle.net/2142/130182 Copy

Description

Title

Applying generative adversarial networks to generate artificial genotype data in livestock

Author(s)

Caballero Vargas, Edgar Giesus

Issue Date

2025-07-21

Director of Research (if dissertation) or Advisor (if thesis)

Bresolin, Tiago

Committee Member(s)

• Wheeler, Matthew B
• Roca, Alfred L

Department of Study

Animal Sciences

Discipline

Bioinformatics

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• generative artificial intelligence models
• synthetic genotypes
• SNP

Abstract

Genomics studies in livestock remain limited due to financial, ethical, or privacy considerations that restrict collaboration and data accessibility. To address these challenges, generative artificial intelligence models, such as Generative Adversarial Network (GAN), are being applied to generate biologically plausible synthetic data without compromising these standards. In this study, we trained a Principal Component (PC) Wasserstein GAN (PC-WGAN) with a gradient penalty to synthesize Single Nucleotide Polymorphism (SNP) in the principal component space. We used a simulated genotype dataset from 4,800 individuals and 37,540 SNP from chromosome 1, and retained 796 PC, which explained 90% of the total variance. These PC scores were used to generate synthetic PC scores and inverse transformed to SNP data. The validity of synthetic SNP data was assessed using both quantitative and visual approaches. Quantitatively, the Pearson correlation between the linkage disequilibrium (LD) values and minor allele frequency (MAF) distributions of the real and synthetic SNP datasets was calculated for comparison. Visually, the PC plots, LD decay curves, and MAF histograms were inspected to compare the distributions and structural patterns of the generated and real data. At 150 training epochs, the model effectively captured the major features of the real population, producing synthetic genotypes that overlapped with real genotypes in PC space and preserved long-range LD patterns and MAF distributions. While synthetic genotypes cannot replace real genomic data, our findings demonstrate that PC-WGAN produces biologically plausible artificial genotypes, offering a promising approach for future work in data augmentation, model benchmarking, and privacy-preserving livestock genomic research.

Graduation Semester

2025-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/130182

Copyright and License Information

Copyright 2025 Edgar Caballero Vargas

Owning Collections