Synthetic observations from deep generative models and binary omics data with limited sample size.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Deep generative models can model joint data distributions for tasks like noise removal and privacy preservation.
• Their performance with limited sample sizes, particularly for binary omics data like single nucleotide polymorphisms (SNPs), remains under-explored.

Purpose of the Study:

• To evaluate the effectiveness of variational autoencoders (VAEs), deep Boltzmann machines (DBMs), and generative adversarial networks (GANs) for modeling binary omics data with limited sample sizes.
• To assess the ability of these models to recover pairwise relationships, specifically odds ratios (ORs), in SNP data.

Main Methods:

• Investigated three deep generative models: VAEs, DBMs, and GANs.
• Utilized both simulated and real single nucleotide polymorphism (SNP) data.
• Assessed model performance based on the recovery of pairwise odds ratios (ORs).
• Included conditional modeling, such as gene expression conditional on SNPs.

Main Results:

• DBMs demonstrated ability to recover structure for up to 300 variables but tended to overestimate ORs without careful tuning.
• VAEs accurately captured the direction and relative strength of pairwise relations but underestimated ORs.
• GANs yielded stable results only with larger sample sizes and strong pairwise associations.

Conclusions:

• DBMs and VAEs are suitable for binary omics data analysis, even with small sample sizes, unlike GANs.
• These findings support the use of DBMs and VAEs for generating synthetic omics data in various applications.