This paper presents the application of a new machine learning approach called Graph Machines to the prediction of the adsorption enthalpies of linear and branched alkanes on various zeolites. In this approach, the molecules are considered as structured data and are represented by graphs. For each individual of the data set, a mathematical function (graph machine) is built, which structure reflects the one of the molecule under consideration. This approach differs from classical quantitative structure-activity relationship (QSAR) methods where molecules are generally described using vectors composed of descriptors. Since no molecular descriptors are used, the collection, computation and selection of these descriptors, which is often a major issue in QSAR applications, is no longer required. The models developed using this approach allowed to satisfactorily predict adsorption enthalpies of all zeolites, even using a very limited training set of 10 molecules. The efficiency of such models, allowing the modelling of complex adsorption behaviour like zeolite ZSM-22 using only few experimental data, illustrates the potential of this new approach in the screening of zeolites for catalytic or adsorption based separation applications. (C) 2010 Published by Elsevier B.V.