Many relationships among data in several areas (such as computer vision, molecular chemistry and pattern recognition) can be represented by graphs. In the machine learning setting, it is an important learning task to classify graph-structural data correctly. Typically, the established techniques for this setting proceed via graph kernels and neural-network classification. In this work, we explore end-to-end learning for graphs: the objective is to operate on the graph representations directly. The key idea of our approach is to use standard tools for graph canonization. We test the performance of this approach on several datasets arising from bioinformatics. In general, we find that the graph canonization, as such, does not improve the accuracy of the classification. A possible reason for this behavior is that the neural network ends up overfitting to the given adjacency matrix representation.