mims-harvard/decagon

Embeds nodes in multimodal graphs using relation-specific graph convolutions, enabling predictions across multiple edge types simultaneously. Implements multiple edge decoders (innerproduct, distmult, bilinear, dedicom) and loss functions (hinge, cross-entropy) to accommodate different prediction tasks. Built on TensorFlow with support for highly multi-relational settings, demonstrated on drug-drug interaction prediction using protein-protein interactions and drug-protein targets as auxiliary graph structure.

469 stars. No commits in the last 6 months.