Graph neural networks exploit relational inductive biases for data that come in the form of a graph. However, in many cases the graph is not available. Can we still use them?