Data augmentation with generative models improves detection of Non-B DNA structures

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Non-B DNA structures, or flipons, are crucial for cellular functions.
• Current experimental methods for flipon detection are limited and cannot capture whole-genome sets.
• Accurate whole-genome annotation of non-B DNA relies on deep learning models, necessitating high-quality training data.

Purpose of the Study:

• To assess the performance of diffusion models in generating synthetic non-B DNA structures for data augmentation.
• To compare diffusion models against other generative models (WGAN, VQ-VAE) for this task.
• To evaluate the impact of data augmentation using synthetic non-B DNA structures on classifier performance.

Main Methods:

• Utilized denoising diffusion probabilistic and implicit models (DDPM and DDIM).
• Compared diffusion models with Wasserstein generative adversarial network (WGAN) and vector quantised variational autoencoder (VQ-VAE).
• Employed a data augmentation strategy combining synthetic and real biological data.

Main Results:

• Data augmentation using generated synthetic non-B DNA structures significantly improved classifier performance.
• Diffusion models generally outperformed other generative models in generating synthetic non-B DNA structures.
• Analysis revealed trade-offs among diffusion models concerning sample quality, diversity, and sampling speed.

Conclusions:

• Diffusion models are effective for generating synthetic non-B DNA structures, enhancing genomic annotation.
• While diffusion models excel, WGAN and VQ-VAE offer alternative trade-offs in the generative trilemma (quality, diversity, speed).
• Further research can optimize generative models for comprehensive non-B DNA annotation.