A novel approach to the lumping of species in large chemical kinetic mechanisms is presented. Species with similar composition and functionalities are lumped into one single representative species. Simulations using the detailed scheme are used to gather statistical information on the distribution of the isomers within each lump group. These distributions are functions of space and time. Closure is performed in state space by approximating these distribution functions as the conditional averages depending on the independent state space variables of the lumped scheme. This approach is simplified further, so that the resulting chemical mechanisms can be used directly in standard chemistry packages. For this purpose, only the dependence of the isomer distributions on the temperature is retained, and optimal correcting factors are incorporated into the Arrhenius form of the rate coefficients of lumped reactions. Validation is performed using two comprehensive mechanisms for n-heptane and iso-octane oxidation. In all cases, a very good agreement is observed between the predictions obtained using the detailed and the lumped mechanisms. Effects of the lumping procedure on sensitivities of the kinetic scheme and on isomer concentrations are studied. Also, integration of this lumping approach into a multi-stage reduction strategy is discussed and illustrated.