Considering　the　classification　of　compounds　as　a　nonlinear　problem,　the　use　of　kernel　methods　is　a　good　choice.　Graph　kernels　provide　a　nice　framework　combining　machine　learning　methods　with　graph　theory,　whereas　the　essence　of　graph　kernels　is　to　compare　the　substructures　of　two　graphs,　how　to　extract　the　substructures　is　a　question.　In　this　paper,　we　propose　a　novel　graph　kernel　based　on　matrix　named　the　local　block　kernel,　which　can　compare　the　similarity　of　partial　substructures　that　contain　any　number　of　vertexes.　The　paper　finally　tests　the　efficacy　of　this　novel　graph　kernel　in　comparison　with　a　number　of　published　mainstream　methods　and　results　with　two　datasets:　NCI1　and　NCI109　for　the　convenience　of　comparison.