CUR - The Council on Undergraduate Research

Predicting Drug-Gene Interactions Via Graph Structure

Kayla Bennett and Jeff Kinne, Department of Mathematics and Computer Science, Indiana State University, 424 North 7th Street, Terre Haute IN 47809 Zachary Abrams, Department of Biomedical Informatics, Wexner Medical Center, The Ohio State University, 1585 Neil Ave, Columbus, OH 43210

Understanding how genes and drugs interact is critical to developing new therapies to treat and care for patients. Many drugs participate in interactions which have not yet been identified, which may hold clues to important novel therapies. To predict unknown interactions between known genes and druglike molecules, we are using a link prediction classifier as well as multiple publicly available datasets. The data was collected from the BioGRID and dgiDB databases. These two databases aggregate many currently known gene-gene and drug-gene interaction pairs. This provides us with a gold standard to train and test our link prediction classifier. We generate feature vectors representing links using the node2vec graph representation algorithm. Node2vec generates vectors to represent nodes in a network based on random walks about the network. We generate a network with two node types, drugs and genes, by drawing edges from BioGRID and dgiDB to seed the network. Our network consists of 6,271 drugs and 25,745 genes. We train a logistic regression model to classify gene-drug pairs as interactions or non-interactions. Our model has achieved an AUC score of .878, and our model predicts several potential interactions related to the central nervous system and the GABA receptor pathway. In the future, we plan to explore these novel drug-gene interactions more thoroughly, using wetlab experiments to validate our current computational findings.

Additional Abstract Information