Discovery Science

Graph neural networks (GNNs) have been successfully applied in many structured data domains, with applications ranging from molecular property prediction to the analysis of social networks. Motivated by broad applicability GNNs, we propose family so-called RankGNNs, a combination Learning Rank (LtR) methods and GNNs. RankGNNs are trained set pair-wise preferences between graphs, suggesting that one them is preferred over other. One practical application this problem drug screening, where an expert wants find most promising molecules large collection candidates. We empirically demonstrate our proposed RankGNN approach either significantly outperforms or at least matches ranking performance naive point-wise baseline approach, which LtR solved via GNN-based graph regression.