FLONE: fully Lorentz network embedding for inferring novel drug targets

Abstract

MotivationTo predict novel drug targets, graph-based machine learning methods have been widely used to capture the relationships between drug, target, and disease entities in heterogeneous biological networks. However, most of the existing methods cannot explicitly consider disease types when predicting targets. More importantly, drug-disease-target (DDT) networks could exhibit multi-relational hierarchical sub-structures with the information of biological interactive functions, but these methods, especially those based on Euclidean space embedding cannot fully utilize such topology information, which might lead to sub-optimal prediction results. We hypothesized that, by properly importing hyperbolic space embedding specifically for modeling hierarchical DDT networks, graph-based machine learning algorithms could better capture the relationships between aforementioned entities, which ultimately improves the target prediction performance. ResultsTo test our hypothesis, we formulated the drug target prediction problem as a knowledge graph triple completion task explicitly considering disease types. To solve this problem, we proposed FLONE, a novel hyperbolic Lorentz space embedding-based method to capture the hierarchical structural information in the DDT network when inferring drug targets. We evaluated it on two different DDT networks, our experimental results showed that by introducing Lorentz space, one of the isomorphic models of hyperbolic space, FLONE generates more accurate candidate target predictions given the drug and disease than the Euclidean translation-based counterparts, which supports our hypothesis. Beyond these experiments, we also devised the hyperbolic encoders to fuse drug and target similarity information into FLONE, to make it capable to handle triples corresponding to previously unseen drugs and targets for more challenging prediction scenarios. Availability and implementationSource code and dataset information are at: https://github.com/arantir123/DDT_triple_prediction. Contacts.he@cs.bham.ac.uk