Graph Autoencoders are a generalisation of Autoencoders in which the elements are structured as graphs. Since now, both models have been applied separately. This paper introduces AE+GAE, a new model for graph regression that combines both. Graph Autoencoders assume node attributes are related to their local structure. Nevertheless, in some applications, not all the node attributes have the property of being dependent of their local structure. Our method learns which attributes do have this property and which ones do not. This is done by feeding all the attributes to both models and combining their latent domain by a neural network. AE+GAE has been applied to predict the Energy, \(pIC_{50}\) and the binding affinity of drugs, which are represented as attributed graphs but could be used in other fields as well. The method demonstrates improved performance compared to other previously presented models.