AbstractMolecular similarity measures are important for many cheminformatics applications like ligand‐based virtual screening and quantitative structure‐property relationships. Graph kernels are formal similarity measures defined directly on graphs, such as the (annotated) molecular structure graph. Graph kernels are positive semi‐definite functions, i.e., they correspond to inner products. This property makes them suitable for use with kernel‐based machine learning algorithms such as support vector machines and Gaussian processes. We review the major types of kernels between graphs (based on random walks, subgraphs, and optimal assignments, respectively), and discuss their advantages, limitations, and successful applications in cheminformatics.