Graph methods for predicting the function of chemical compounds

Finding chemical compounds that can be used to treat a certain disease has long been a focus of the biomedical research. Using traditional laboratory approaches, scientists have to test numerous chemical compounds in order to find a drugable compound. Computational methods can speed up this screening process. We compared various computational methods that predict the function of chemical compounds. The global similarity methods used graph kernels to enumerate and compare various structural components in compounds. The local similarity methods looked for common subgraphs that drugable compounds share. We evaluated these two categories of methods using five benchmark datasets from the National Cancer Institute database. Our results showed that the MinMax Tanimoto kernel was better than other graph kernels in predicting the function of chemical compounds. Our results also showed that local similarity methods outperformed the global similarity, which suggested that local substructures played a pivotal role in the function of chemical compounds. We explored different ensemble methods to integrate the predictions from global and local similarity methods. Our results showed that combining global and local similarity methods significantly improved the prediction performance, which indicated that the global and local similarity methods complemented each other in comparing structural similarity. Our ensemble method achieved better performance than other state-of-the-art methods in direct comparisons.