A computer implemented method of particular, although not exclusive application to analysing a plurality of molecules which comprises computing a kernel function for each pair of the plurality of molecules, the kernel function being representative of the number of features present in both molecules of the pairs and using the kernel function in a kernel based learning algorithm to model the relationship between the features and a property of the molecules. The method is also applicable to predicting a numerical value representing a characteristic of a molecule and, more generally, modelling instances of data in a database. A particular, although again not exclusive application, is the prediction of toxicity of a molecule.