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

Abstract

Non-B DNA structures, or flipons, are important functional elements that regulate a large spectrum of cellular programs. Experimental technologies for flipon detection are limited to the subsets that are active at the time of an experiment and cannot capture whole-genome functional set. Thus, the task of generating reliable whole-genome annotations of non-B DNA structures is put on deep learning models, however their quality depends on the available experimental data for training. The data augmentation approach as the combination of synthetic and real data is widely used in various fields. Deep generative models demonstrated promising results in data augmentation improving classifiers’ performance. Here we aimed at testing performance of diffusion models in comparison to other generative models in generating synthetic non-B DNA structures for data augmentation approach. We tested denoising diffusion probabilistic and implicit models (DDPM and DDIM), Wasserstein generative adversarial network (WGAN), vector quantised variational autoencoder (VQ-VAE) and showed that data augmentation improves the quality of classifiers. Diffusion models overall show the best results, but when considering three criteria of generative trilemma - quality of generated samples, diversity and sampling speed, we conclude that trade-off is possible between generative diffusion model and other architectures such as WGAN and VQ-VAE.