A Novel Generative Adversarial Network simulating the complementary structure of DNA genetic information

Abstract

To solve the problems of mode collapse and training instability in generative adversarial networks (GANs), a framework simulating the complementary structure of DNA is proposed, in which a complementary unit and a generalization unit are added. Four latent vectors representing four bases of A, T,C and G are obtained from the complementary unit. Through the combination of latent vectors, the generalization unit avoids the fitting of high-dimensional data distribution and obtains a more comprehensive vector space. Experimental results show that the problems of model collapse and training instability are effectively solved, compared with state-of-the-art VAE-GAN, the FID score increases 52.2%, indicating that the quality and diversity of images generated by the model are improved.