Link Prediction in Multi-layer Networks and Its Application to Drug Design

Abstract

Search of valid drug candidates for a given target is a vital part of modern drug discovery. Since the problem was established, a number of approaches have been proposed that augment interaction networks with, typically, two compound/target similarity networks. In this work we propose a method capable of using an arbitrary number of similarity or interaction networks. We adapt an existing method for random walks on heterogeneous networks and show that adding additional networks improves prediction quality.

Link prediction via community detection in bipartite multi-layer graphs

“…The first group assumes 2 out of 3 possible layers to be similarity networks for drugs and targets respectively, and exploits similarity information to perform link prediction on the third bipartite layer [4,11]. The second models the behavior of a random-walker to perform link prediction in multi-layer graph using PageRank adaptations [6,7,21]. Such methods are dependent on fixing the similarity networks and while the approach was extended to any number of networks in [21], it pays for this flexibility with high computational cost.…”

“…The second models the behavior of a random-walker to perform link prediction in multi-layer graph using PageRank adaptations [6,7,21]. Such methods are dependent on fixing the similarity networks and while the approach was extended to any number of networks in [21], it pays for this flexibility with high computational cost. The last group of methods maps drugs, targets and their interactions into a combined feature space, and performs drug-target interaction prediction using distance functions or regression analysis [35,37].…”

“…Our problem setting was addressed in [21], using a random walk approach which is basically an extension of PageRank for any number of layers, and thus can be considered as a baseline in this work. JC CC and JC N C in combination with both spectral partitioning and the Louvain clearly outperform that baseline in terms of AUC: 0.84 vs 0.96 (Louvain) and 0.92 (spectral partitioning) for Enzyme, 0.8 vs 0.89 and 0.85 for GPCR, 0.76 vs 0.97 and 0.88 for IC, 0.63 vs 0.82 and 0.77 for NR, and 0.61 vs 0.91 and 0.86 for Kinase, respectively.…”