We use the Lempel-Ziv complexity method to investigate effects of amino acid classification on prediction of protein structural classes. First, we find that contributions of amino acid classification are differential for predicting protein structural classes and even the performances of some amino acid classification are better than that without using the amino acid classification. This inspires us to observe whether the combination of amino acid classification can improve the performance for predicting protein structural classes. Finally, we convert each Lempel-Ziv complexity distance matrix into a novel kernel matrix and then use Bayesian multiple kernel learning to combine all kernels. Our method is tested on four benchmark datasets and outperforms previous methods consistently. This suggests that our proposed method is valuable for predicting protein structural classes.