Hierarchical and Centrality driven Polypharmacy Side Effect Prediction Model

Modelling polypharmacy side effect (POSE) prediction as a graph link prediction problem eases the adaptation of the convention graph theory and geometric learning approaches to predicting polypharmacy side effects. However, popular methods often deployed on POSE prediction concentrate on capturing the local neighborhood information neglecting other rich network information. Among the under-looked information includes the centrality role of nodes in the network and the network hierarchical information. In this work, we propose a novel architecture that preserves both the Hierarchical and node Centrality role during embedding learning for POSE prediction (HC-POSE). Firstly, we exploit the underlying network hier-archical information with k-core decomposition. Secondly, we preserve the node centrality and topology information with node strength-based features. Lastly, we propose an end-to-end archi-tecture that preserves all the crucial diverse network information in a unified framework. From the experimental results, HC-POSE showed a 3% improvement in POSE prediction accuracy over the best baseline.