Combination therapy, which administers multiple drugs either simultaneously or sequentially, aims to enhance therapeutic efficacy while minimizing side effects and drug resistance common in monotherapies for complex diseases. Identifying drug combinations that exhibit synergistic effects is therefore critical to both biomedical research and clinical practice. In this study, we introduce TensoGraph, a novel tensor-transformer model that integrates gene expression, drug structure, and physiochemical fingerprints within heterogeneous graphs to capture global-local interactions for drug synergy prediction. Specifically, we construct cell line-specific heterogeneous drug combination graphs where edges represent synergistic, additive, or antagonistic interactions. To capture global patterns, we apply Tucker tensor decomposition on the multi-relational interaction tensor to extract low-dimensional drug embeddings that reflect holistic interaction profiles. Concurrently, a heterogeneous graph transformer network is employed to learn local structural representations via automatic meta-path discovery. The global and local features are further integrated with drug molecular structure and physicochemical fingerprints to construct comprehensive drug representations. Extensive experiments on multiple benchmark datasets demonstrate that TensoGraph significantly outperforms state-of-the-art baselines in predicting synergistic drug combinations, with improved biological interpretability. These results underscore the effectiveness of tensor-based global-local modeling for capturing complex drug interaction mechanisms and facilitating the discovery of personalized combination therapies.
Wang et al. (Sun,) studied this question.