This paper introduces Decagon, a graph convolutional neural network for modeling polypharmacy side effects. Decagon constructs a multimodal graph of protein-protein interactions, drug-protein interactions, and drug-drug interactions (side effects), where each side effect is represented as a different edge type. The model uses a new multirelational link prediction approach to predict drug-drug interactions and their types. Decagon outperforms existing methods by up to 69% in predicting polypharmacy side effects, particularly for side effects with strong molecular bases. The model effectively shares parameters across edge types, leading to good performance on non-molecular side effects. Decagon's multimodal graph includes 964 side effect types and 715,612 protein-protein, 4,651,131 drug-drug, and 18,596 drug-protein edges. The model is trained on a dataset of 645 drugs and 19,085 proteins. Decagon's encoder produces node embeddings, while the decoder uses these embeddings to predict side effects. The model is evaluated on 964 side effect types, achieving high AUROC, AUPRC, and AP@50 scores. Decagon's predictions are validated against biomedical literature, showing strong support for several novel predictions. The model's ability to capture interdependence between side effects is demonstrated through t-SNE visualization of side effect embeddings. Decagon's approach provides a new method for predicting polypharmacy side effects, with potential applications in drug combination therapy development.This paper introduces Decagon, a graph convolutional neural network for modeling polypharmacy side effects. Decagon constructs a multimodal graph of protein-protein interactions, drug-protein interactions, and drug-drug interactions (side effects), where each side effect is represented as a different edge type. The model uses a new multirelational link prediction approach to predict drug-drug interactions and their types. Decagon outperforms existing methods by up to 69% in predicting polypharmacy side effects, particularly for side effects with strong molecular bases. The model effectively shares parameters across edge types, leading to good performance on non-molecular side effects. Decagon's multimodal graph includes 964 side effect types and 715,612 protein-protein, 4,651,131 drug-drug, and 18,596 drug-protein edges. The model is trained on a dataset of 645 drugs and 19,085 proteins. Decagon's encoder produces node embeddings, while the decoder uses these embeddings to predict side effects. The model is evaluated on 964 side effect types, achieving high AUROC, AUPRC, and AP@50 scores. Decagon's predictions are validated against biomedical literature, showing strong support for several novel predictions. The model's ability to capture interdependence between side effects is demonstrated through t-SNE visualization of side effect embeddings. Decagon's approach provides a new method for predicting polypharmacy side effects, with potential applications in drug combination therapy development.