Automatic analysis of changes in the H-1 NMR spectrum of a mixture and their interpretation in terms of chemical reactions taking place have a diversity of possible applications, from the monitoring of reaction processes or degradation of chemicals to metabonomics. Classification of photochemical and metabolic reactions by Kohonen self-organizing maps and random forests is demonstrated, taking as input the difference between the H-1 NMR spectra of the products and the reactants. The chemical shifts of the reactants and products were fuzzified to obtain a crude representation of the spectra. With a dataset of 911 metabolic reactions catalyzed by transferases (EC number 2.x.x.x), classification according to subclass (second digit of the EC number) could be achieved with up to 84% of accuracy. Experiments with a dataset of 189 photochemical reactions, manually assigned to seven classes, yielded 86-93% of correct classifications for an independent test set of 42 reactions, and the models were further validated with a test set combining experimental and simulated chemical shifts. The results support our proposal of linking databases of chemical reactions to NMR data for automatic reaction classification and show the usefulness of the predictions obtained by the SPINUS program for the estimation of missing NMR experimental data.