Triple Matrix Factorization for Drug-Drug Interaction Prediction Using Fused Gromov-Wasserstein Distances

Determining whether two drugs interact is an expensive and time-consuming process, and computational tools can help identify putative drug candidates for further investigation. This paper focuses on predicting drug-drug interactions using a matrix factorization-based approach. In particular, we build upon a triple matrix factorization using Wasserstein-based distance metrics as our sole prediction feature. Of the methods we considered, the Fused Gromov-Wasserstein distance measure ultimately performs the best due to how it leverages both feature information at the atomic level as well as structural information, which captures the relational aspects among the atoms within the molecule.