Deep convolutional and conditional neural networks for large-scale genomic data generation.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Generative models are increasingly used for genomic data analysis and generation.
• Previous work showed Generative Adversarial Networks (GANs) and Restricted Boltzmann Machines (RBMs) can create artificial genomes (AGs).
• Scalability remains a challenge for genome-wide data due to its large feature space.

Purpose of the Study:

• To develop scalable generative models for high-SNP-number artificial genome generation.
• To assess the quality of generated haplotypes and evaluate privacy leakage.
• To enable ethical research using genomic data surrogates.

Main Methods:

• Implementation of a novel convolutional Wasserstein GAN (WGAN).
• Development of a novel conditional RBM (CRBM) framework.
• Comparative analyses of haplotype quality and privacy leakage.

Main Results:

• The novel WGAN and CRBM models effectively generate AGs with high SNP numbers.
• These models capture complex genomic correlations and produce diverse, plausible haplotypes.
• Generated artificial genome segments show minimal privacy leakage from training data.

Conclusions:

• Scalable generative neural networks can produce high-quality artificial genomes with preserved characteristics.
• These methods offer a promising approach for protecting genetic privacy in genomic data.
• Large-scale artificial genome databases can facilitate ethical genomic research.