Drug combinations are increasingly used to combat drug resistance in cancer and infections, but the current trial-and-error approach to choosing them often leads to ineffective or unsafe combinations that can worsen clinical outcomes. While several computational models have been developed to predict optimal drug combinations, they often assume homophily (i.e., similarity in drug-target and drug-drug interactions) and fail to capture the heterophilous nature of many biological relationships wherein nodes frequently connect to dissimilar nodes. This project proposes a paradigm shift by characterizing heterophily in drug-protein interaction networks and interrelated biological networks, and introducing a novel heterophily-aware graph neural network (GNN) that jointly models diverse biochemical graphs to improve synergy prediction and enable interpretable drug combination discovery.