GLA-Synergy: An Interpretable Global-Local Adaptive Framework for Drug Synergy Prediction in Cancer Treatment

Effective anticancer drug combinations are crucial for advancing cancer treatment, yet predicting drug synergy remains challenging due to the complexity of biological interactions. Existing methods struggle to integrate multimodal features and to model the hierarchical interactions between drugs and cell lines, limiting their ability to uncover the underlying mechanisms of synergy. To address these issues, we propose global-local adaptive synergy model (GLA-Synergy), a novel deep learning framework designed to enhance both the accuracy and the interpretability of synergy prediction. GLA-Synergy employs a multimodule architecture to extract global features from both drug molecules and cell lines and introduces an improved linear attention mechanism to capture local pairwise interactions. For drug representation, the model utilizes a graph convolutional network to extract molecular structural information combined with an adaptive bidirectional attention mechanism to construct global representations of drug pairs. For cell line encoding, a 1D convolutional neural network is integrated with an adaptive channel attention module, enabling the capture of both global semantic information and local details via cross-layer fusion. In the interaction learning module, a dual bilinear attention network (Dual-BAN) is employed to perform local multilevel fusion of drug and cell features, followed by a fully connected neural network for synergy prediction. The core innovation of GLA-Synergy lies in its progressive global-local fusion framework, which effectively captures key interactions between drugs and cellular contexts. Experimental results demonstrate that GLA-Synergy consistently outperforms existing methods across multiple benchmark data sets, providing an efficient and interpretable tool for the discovery of synergistic anticancer therapies.