Given a mixture of samples from unobserved subpopulations with distinct underlying causal mechanisms, we give results on identification and discovery of causal graph with latent mixing variables.