Sampling from Disentangled Representations of Single-Cell Data Using Generative Adversarial Networks

Deep generative models, including variational autoencoders (VAEs) and generative adversarial networks (GANs), have achieved remarkable successes in generating and manipulating highdimensional images. VAEs excel at learning disentangled image representations, while GANs excel at generating realistic images. Here, we systematically assess disentanglement and generation performance on single-cell gene expression data and find that these strengths and weaknesses of VAEs and GANs apply to single-cell gene expression data in a similar way. We also develop MichiGAN1, a novel neural network that combines the strengths of VAEs and GANs to sample from disentangled representations without sacrificing data generation quality. We learn disentangled representations of two large singlecell RNA-seq datasets [13, 68] and use MichiGAN to sample from these representations. MichiGAN allows us to manipulate semantically distinct aspects of cellular identity and predict single-cell gene expression response to drug treatment.