COLLECTIVE PAIRWISE CLASSIFICATION FOR MULTI-WAY ANALYSIS OF DISEASE AND DRUG DATA.

Interactions between drugs, drug targets or diseases can be predicted on the basis of molecular, clinical and genomic features by, for example, exploiting similarity of disease pathways, chemical structures, activities across cell lines or clinical manifestations of diseases. A successful way to better understand complex interactions in biomedical systems is to employ collective relational learning approaches that can jointly model diverse relationships present in multiplex data. We propose a novel collective pairwise classification approach for multi-way data analysis. Our model leverages the superiority of latent factor models and classifies relationships in a large relational data domain using a pairwise ranking loss. In contrast to current approaches, our method estimates probabilities, such that probabilities for existing relationships are higher than for assumed-to-be-negative relationships. Although our method bears correspondence with the maximization of non-differentiable area under the ROC curve, we were able to design a learning algorithm that scales well on multi-relational data encoding interactions between thousands of entities.We use the new method to infer relationships from multiplex drug data and to predict connections between clinical manifestations of diseases and their underlying molecular signatures. Our method achieves promising predictive performance when compared to state-of-the-art alternative approaches and can make "category-jumping" predictions about diseases from genomic and clinical data generated far outside the molecular context.