Cancer is a multifaceted disease caused by dynamic interaction between genetic mutations and environmental factors. Understanding the genetic mutations underlying the development and progression of cancer is the stepstone for developing effective treatments and therapies. However, these mutations occurred in only a small fraction of cancer patients and it is extremely difficult to associate with cancer. Here, we propose MutNet, a heterogeneous network embedding method which integrate biomolecular network with cancer genomics data. Using pan cancer genomic data from The Cancer Genome Atlas program and public protein-protein interaction and pathway data, MutNet identifies rarely mutated cancer genes often overlooked by conventional genetic studies. In addition, the unified vector representation of biological entities allows us to reveal the tumor type specific cancer genes, cancer gene modules, and potential relationships among different tumor types. Our heterogeneous network embedding method holds the promise for the underlying mechanisms of cancer and potential therapeutic targets.