Studies on neurobiology of autism predominantly focus on describing topological properties of brain connectivity, offering little insight into mechanisms that give rise to associated alterations. Recently, generative network modeling has been proposed to uncover those mechanisms with parameterized wiring rules. However, most models utilize predefined wiring rules, which may not be the best characterization. Here we propose a novel deep learning based generative modeling to infer underlying mechanisms that drive atypical network growth. We demonstrate high similarity of generated network to observed structural networks, with their latent features predicting autism diagnosis. Given the neurobiological interpretability of embedding, our approach shows potential in providing connectomic biomarkers for autism.