Neural networks on graphs enable learning over graph structures by generalizing the notion of convolution operation typically applied to image datasets to operations that can operate on arbitrary graphs. These neural networks can also be seen as an embedding methodology that distills high-dimensional information about each node’s neighborhood into a dense vector embedding without requiring manual feature engineering. We will focus on multi-relational graphs that may contain multiple type of nodes and edges and apply these methodologies on biomedical datasets.
As a first step the state-of-the-art on graph embeddings and graph neural networks on multi-relational graphs will be analyzed. Then, these approaches will be extended and modified to be used on different biomedical graphs (e.g. drugs, proteins, side-effects, pathways etc.) that may also contain external information (e.g. drugs description). The end goal will be to create an end-to-end system that can handle multi-relational graphs and infer links, node types or sub-structures, depending on the task at hand, by generating robust embeddings.