This paper presents a novel approach devoted to the design of feed forward hybrid neural models. Graph genetic programming techniques are used to provide a flexible construction of partially interconnected neural structures with heterogeneous layers built as combinations of local and global neurons. By exploiting the inner modularity and the parallelism of the neural architectures, the approach suggests the encryption of the potential mathematical models as directed acyclic graphs and defines a minimally sufficient set of functions which guarantees that any combination of primitives encodes a valid neural model. The exploration capabilities of the algorithm are heightened by means of customized crossovers and mutations, which act both at the structural and the parametric level of the encrypted individuals, for producing offspring compliant with the neural networks' formalism. As the parameters of the models become the parameters of the primitive functions, the genetic operators are extended to manage the inner configuration of the functional nodes in the involved hierarchical individuals. The applicability of the proposed design algorithm is discussed on the identification of an industrial nonlinear plant.