This contribution presents a novel method for predicting the mechanical response of arbitrary microstructures based on their microstructures and their stress-strain behavior. The approach use two-point statistics of the microstructures, in addition, the Finite element (FE) computations of the stiffness in two directions are used. The approach learn the relationship between the computed statistics and the specific stiffness by first deriving a low dimensional representation of the 3D statistics, this representation also allows clustering and visualization of the distribution of microstructure data. The mapping between the low dimensional representation and the stiffness is computed using a non-linear regression method with radial basis functions. The predicted stiffness for a given two-point statistical distribution is compared with a Deep Learning approach. Both approaches are compared in term of accuracy, computational effort and required amount of training data. It is demonstrated that the proposed method can predict the mechanical properties of arbitrary microstructure designs with adequate accuracy and much reduced computational effort.