Agraph convolution-based classification model for identifying anticancer metabolites from traditional vietnamese herbal medicine database
Nguyen Anh Vu, Pham Truong Duy, Le Thi Ly · 2018
Vietnam has been well known as a source of abundantly diverse herbal medicines for thousands of years, which serves a variety of purposes in drug development in attempts to address health issues, such as cancer. As claimed by a chemoinformatics-related principle that structurally similar chemical compounds will very likely have similar biological activity, this study employs molecular graph convolution, a machine learning architecture for extracting features from small molecules as undirected graphs, to predict anticancer ability of Vietnamese herbal medicines based on their metabolites' structures. In addition to molecular graph convolution, extended connectivity fingerprint (ECFP), a traditional featurizer for exploiting details of molecules, is also performed in order to make performance comparison. Finally, we successfully constructed a graph convolution-based neural network with high predictive accuracy on both training and validation set, suggesting that the model is reliable in detecting anticancer activity.