Enhancing genomic data synthesis: A WGAN-GP approach for haplotype generation and evaluation using quasi Manhattan Wasserstein distance

Abstract

Abstract Background The utilization and translation of genomic data from large biobanks has revolutionized the field of biomedical research, drug development and precision medicine. Despite the advances in genetic epidemiology research, limited sample size of certain rare diseases and minority population remains a critical issue. As an alternative to collecting more samples, generating realistic synthetic human genomic data by mimicking the population structure can uplift the sample size of disease cohorts or minority groups. The recently proposed Generative Adversarial Networks (GANs) for generating artificial genomes still requires extensive hyperparameter tuning and often fails to converge. Results We utilized WGAN-GP on phased haplotype data sourced from individuals with type 1 diabetes (T1D: n=3,698) and a disease-free cohort (healthy: n=51,857) from the UK Biobank. Subsequently, synthetic datasets were generated, doubling the size of the original samples (TID': n=7,396, healthy': n=103,714). The input data comprised haplotype pairs with selected single nucleotide polymorphisms (SNPs), notably rs6679677 and rs2476601 in the PTPN22 gene, identified in previous genome-wide association studies (GWAS) as linked to increased T1D risk. We evaluated WGAN-GP’s ability to capture the complex multidimensional structure of the input data through Jensen–Shannon divergence, cosine distance, and a novel quasi Manhattan Wasserstein distance. Additionally, we demonstrated the two-dimensional principal component analysis (PCA) representation of real and synthetic data and showcased the allele frequencies between real and synthetic data. We introduced the Quasi Manhattan Wasserstein Distance (QMWD) and explored its potential in training. Conclusions The study highlights the potential of WGAN-GP in synthetic genomic data generation, addressing challenges faced by traditional GANs. QMWD, with O(n) efficiency, offers a promising avenue for improvement.