Abstract Simulation is a critical component of experimental design and evaluation of analysis methods in microbiome association studies. However, statistically modeling the microbiome data is challenging since that the complex structure in the real data is difficult to be fully represented by statistical models. To address this challenge, we designed a novel simulation framework for microbiome data using a generative adversarial network (GAN), called MB-GAN, by utilizing methodology advancements from the deep learning community. MB-GAN can automatically learn from a given dataset and compute simulated datasets that are indistinguishable from it. When MB-GAN was applied to a case-control microbiome study of 396 samples, we demonstrated that the simulated data and the original data had similar first-order and second-order properties, including sparsity, diversities, and taxa-taxa correlations. These advantages are suitable for further microbiome methodology development where high fidelity microbiome data are needed.