The new global ensemble prediction system GME-EFS is set up using the National German Meteorological Service's global weather forecasting model GME which is based on an icosahedral grid structure. Ensemble initialization in the GME-EFS is implemented through the "breeding of growing modes" technique described by Toth and Kalnay (1993, 1997). Results from a simple bred vector initialization experiment exhibit small performance in terms of uncertainty estimation which is due to similarities amongst the bred vectors. To better utilize the subspace of growing modes for the initial ensemble perturbations, a singular value decomposition is applied to the bred vectors in order to orthogonalize them based on their similarity. Results of ensemble simulations initialized with such orthogonalized bred vector perturbations lead to a significant enhancement of ensemble spread skill and therefore forecast uncertainty estimation.