A dual granular balanced deep forest model for effective drug combination prediction

The treatment of complex diseases often benefits from combination therapies, yet identifying synergistic drug pairs remains challenging due to the vast search space and the extreme imbalance between synergistic and non-synergistic outcomes in available data. Here, we present a deep-forest-based framework designed to improve synergy prediction under highly skewed class distributions by prioritizing informative and uncertain training examples during learning. Experiments demonstrate that our method consistently achieves favorable results relative to a broad set of representative canonical, imbalanced learning, and drug-specific prediction models. Beyond predictive accuracy, we provide model interpretability analyses to highlight chemical substructures and cell-line-specific genetic signals associated with synergy, and we further validate top-ranked predictions through literature- and database-supported case studies. Together, these results suggest a practical and interpretable approach for accelerating the discovery of biologically plausible synergistic drug combinations.