A machine learning framework for predicting synergistic and antagonistic drug combinatorial efficacy

Synergistic drug combinations could achieve increase of therapeutic efficacy and reduce risk of drug resistance and toxicity via targeting multiple genes, signaling pathways or mutations simultaneously. Comparatively antagonistic drug combinations decrease the therapeutic efficacy of co-prescribed drugs and even occasionally cause severe adverse effects. The existing computational methods generally treat drug synergy as binary classification with drug antagonism neglected, and the training data are so small as to riskily lead to model overfitting. In this study, we propose a self-contained machine learning framework to predict drug synergy, antagonism and additivity simultaneously. This framework trains a three-class l2-regularized logistic regression model on large-scale drug-drug interactions with therapeutic efficacy from DrugBank. Moreover, the proposed framework represents drug pairs via simple profiles of drug-targeted genes and cellular processes so as to well explain the molecular mechanism behind drug-drug interactions and reduce data complexity. Cross validation and independent test show that the proposed framework achieves encouraging performance and outperforms the existing computational methods. We further employ the proposed framework to predict the synergy and antagonism of known cancer-associated drug combinations and analyse the molecular mechanisms via GO and pathways enrichment analyses. The predictions and unravelled mechanisms potentially provide insights into for drug combinatorial therapies of cancers.