KGLGANSynergy: knowledge graph-based local and global attention network for drug synergy prediction Authors: Chaokun Yan, Menglei Chu, Ge Zhang, Caili Fang, Nian Liu, Huimin Luo Published: 2025 ABSTRACT BackgroundDrug combination therapy offers an effective strategy for treating complex diseases by leveraging synergistic multi-target effects. However, conventional experimental approaches for identifying synergistic drug pairs are costly and inefficient. Graph representation learning has shown strong potential for modeling complex biomolecular interactions and is widely applied to drug synergy prediction. Despite this progress, existing knowledge graph–based methods primarily focus on node-level relational features while overlooking edge attributes and global semantic context, limiting predictive capability.MethodsWe propose KGLGANSynergy, a knowledge graph–based deep learning framework for drug synergy prediction. The model integrates a Local Graph Attention Network (LGAT) to learn fine-grained neighborhood features as local features, and a Global Graph Attention Network (GGAT) to capture global structural information in the drug knowledge graph as global features. These learned features are fused through a Mutual Cross Attention (MCA) mechanism to obtain representations of drugs and cell lines. The final synergy score is computed by combining the feature vectors of the two drugs and the cell line and applying a sigmoid function to generate the predicted probability.ResultsExperimental results on two benchmark datasets, DrugCombDB and Oncology-Screen, demonstrate that KGLGANSynergy achieves AUPR scores of 0.9360 and 0.9472, respectively, outperforming existing baselines. Case studies further validate the model’s ability to accurately identify novel synergistic drug combinations.ConclusionThis study provides a more accurate computational tool for predicting drug synergy and introduces a multi-level graph representation learning paradigm, offering new perspectives for biomedical data modeling and showing considerable potential for clinical applications. Introduction Drug combinations are a more effective strategy for treating complex diseases by disrupting disease progression through the synergistic action of multiple drugs targeting distinct molecular pathways [1]. For cancer therapy, monotherapy is often limited by issues such as drug resistance and suboptimal efficacy [2]. Mounting evidence suggests that multi-drug combination therapies can substantially improve therapeutic outcomes, mitigate resistance, and reduce toxicity through synergistic interactions, offering significant advantages in oncology [3, 4]. However, current clinical strategies for developing combination therapies are typically time-consuming, labor-intensive, and costly. While high-throughput screening (HTS) [5] enables the generation of large-scale experimental data at a relatively lower cost and within a shorter time frame, it remains impractical to exhaustively evaluate the entire drug combination space. Furthermore, cancer cell lines commonly used in HTS may not accurately reflect the biological complexity of in vivo tumor environments [6, 7]. These limitations underscore the urgent need for the development of efficient and cost-effective computational approaches to facilitate large-scale drug combination screening. With the rapid advancement of artificial intelligence, machine learning (ML) has been widely employed to tackle the challenges associated with drug combination prediction. Common ML approaches include forest-based models, matrix factorization techniques, and Bayesian frameworks [8]. Among these, forest-based models have been extensively explored. For example, Li et al. proposed a random forest–based method that integrates multi-source data to predict drug synergy [9]. Meng et al. extracted features from protein–protein interaction (PPI) networks using network embeddings, and combined drug and cancer cell line information to predict synergy using XGBoost [10]. More recently, these approaches have been extended to incorporate deep forest models for enhanced performance in synergy prediction [11–13]. Moreover, matrix factorization–based methods have also demonstrated promising results. Nafshi and Lezon employed probabilistic matrix factorization to infer the efficacy of untested drug combinations using only partially observed outcome data, applying matrix decomposition and optimization techniques [14]. Pinoli et al. proposed a non-negative matrix tri-factorization framework that integrates drug synergy measurements with cell line genomic profiles to predict context-specific synergy scores and drug–cell line interactions [15]. For Bayesian methods, Li et al. utilized a Bayesian network to integrate multidimensional drug features and applied probabilistic ensemble learning to predict combination outcomes [16]. Rønneberg et al. adopted a Bayesian framework based on Gaussian processes to model the dose–response behavior of drug pairs, explicitly capturing uncertainty and noise in experimental measurements to improve the prediction of synergistic and antagonistic interactions [17, 18]. While these machine learning approaches offer valuable insights, they often depend heavily on manual feature engineering and struggle to model complex nonlinear relationships. These limitations have spurred increasing interest in the development of deep learning–based methods for drug synergy prediction. Recently, deep learning and large language model (LLM)–based approaches have shown promising performance in drug combination prediction, contributing significantly to the advancement of computational drug development. For example, Preuer et al. introduced DeepSynergy [7], a model that integrates chemical features of drugs and genomic features of cell lines, represents them as one-dimensional vectors, and inputs them into a multilayer perceptron (MLP) to predict synergistic effects. This work marked a key breakthrough in applying deep learning to predict anticancer drug synergy. Beyond one-dimensional feature inputs, drug structural information and biological knowledge have also been incorporated into deep learning models [19]. Drug molecules are typically represented as graph structures, where nodes correspond to atoms and edges represent chemical bonds. Graph neural networks (GNNs) have been used to extract structural representations from molecular graphs [20–24]. Additionally, drug synergy prediction has been improved by constructing graphs based on protein–protein interaction (PPI) networks or by integrating relationships among drugs, proteins, and cell lines. GNNs are then applied to learn comprehensive feature representations of both drugs and cell lines within these heterogeneous graphs [25–29]. In parallel, large language models (LLMs) have demonstrated exceptional capabilities in natural language understanding and generation across a wide range of domains [30]. Xu et al. utilized a BERT model fine-tuned with SimCSE (Simple Contrastive Learning of Sentence Embeddings) to encode SMILES strings of drugs, generating more accurate molecular embeddings and thereby improving synergy prediction performance [31]. Li et al. reformulated few-shot drug synergy prediction as a natural language processing (NLP) task and leveraged pretrained LLMs to extract prior knowledge from unstructured text corpora. This enabled effective prediction of drug synergy in rare tissue types, where structured data and curated features are often limited [32]. Among the aforementioned deep learning methods, graph representation learning–based methods have been widely adopted due to their strong ability to model topological structures and capture complex biological relationships [33]. For instance, Yang et al. proposed GraphSynergy [28], which integrates the topological information of drug targets and cell line–associated proteins in protein–protein interaction (PPI) networks using graph convolutional networks (GCNs) combined with an attention mechanism. However, this model does not incorporate drug molecular structures or gene expression data. To address this limitation, Wang et al. introduced DeepDDS [20], which employs graph neural networks (GNNs) to learn structural features from molecular graphs of drugs and combines them with gene expression profiles of cell lines to predict drug synergy. Nonetheless, DeepDDS does not consider the biological correlations between drugs and cell lines, such as those involving proteins or tissues. Zhang et al. proposed KGANSynergy [29], which constructs a knowledge graph to represent relationships among drugs, cell lines, proteins, and tissues, and utilizes graph attention networks (GATs) to learn entity embeddings. However, this model relies on random initialization of entity features, which may lead to drug and cell line representations that deviate from their true biological properties. To mitigate this issue, Cheng et al. proposed HANSynergy [26], which initializes drug features using pretrained BERT embeddings of SMILES strings and cell line features using gene expression data. It employs a heterogeneous graph attention network to aggregate neighborhood information. Nevertheless, this approach does not incorporate drug–drug synergy graphs and is limited to single-view learning. Hao et al. introduced GNNSynergy [25], a multi-view GNN framework that constructs drug synergy graphs from a main view (representing specific cancer cell lines) and subviews (representing other cell lines from the same tissue). The model learns entity representations and fuses multi-view information via attention mechanisms, significantly enhancing prediction performance. Despite these advancements, current graph-based drug synergy prediction methods still face two major challenges. First, while models such as KGANSynergy and HANSynergy learn representations for both nodes and edge types, others like GraphSynergy, DeepDDS, and GNNSynergy focus exclusively on node-level features, completely ignoring edge attributes. This simplification may result in the loss of important local interaction information during representation learning. Second, most existing GNN-based models are limited by the local aggregation scope imposed by message-passing mechanisms, preventing them from effectively capturing graph-wide structural dependencies. As a result, node representations often lack global semantic context. These two limitations jointly restrict the ability of current models to accurately capture the complex interactions underlying drug synergy. To address the limitations of existing graph neural network–based methods for drug combination prediction, this study proposes a knowledge graph–driven deep learning framework termed Knowledge Graph Local and Global Attention Network (KGLGANSynergy). The framework consists of three core components designed to enhance prior models. The Local Graph Attention Network (LGAT) precisely captures neighborhood interactions in the knowledge graph by assigning a learnable feature vector to each edge, enabling fine-grained aggregation of local information and effectively addressing the challenge of missing edge features. The Global Graph Attention Network (GGAT) establishes graph-wide dependencies and learns global semantic representations for each entity, thereby compensating for the absence of global contextual information in previous approaches. To fuse the learned local and global representations, the framework introduces the Mutual Cross Attention (MCA) mechanism [34], which enables adaptive and complementary integration of multi-scale features. This hierarchical architecture not only mitigates the issue of inadequate edge feature modeling but also overcomes the limited receptive field inherent in traditional message passing networks through a coordinated local–global representation mechanism. As a result, the enriched feature representations significantly improve the accuracy and robustness of drug synergy prediction. The main contributions of this work are as follows:We propose a knowledge graph–based framework for drug combination prediction, which effectively captures both local and global information of entities within the knowledge graph, thereby improving predictive performance.We leverage Transformer-based models to learn entity representations from the knowledge graph by capturing long-range dependencies across nodes, enabling the integration of global semantic information for entity representation.We introduce a mutual cross-attention mechanism to adaptively fuse local and global features of knowledge graph nodes, which significantly enhances the accuracy of synergy prediction. We propose a knowledge graph–based framework for drug combination prediction, which effectively captures both local and global information of entities within the knowledge graph, thereby improving predictive performance. We propose a knowledge graph–based framework for drug combination prediction, which effectively captures both local and global information of entities within the knowledge graph, thereby improving predictive performance. We leverage Transformer-based models to learn entity representations from the knowledge graph by capturing long-range dependencies across nodes, enabling the integration of global semantic information for entity representation. We leverage Transformer-based models to learn entity representations from the knowledge graph by capturing long-range dependencies across nodes, enabling the integration of global semantic information for entity representation. We introduce a mutual cross-attention mechanism to adaptively fuse local and global features of knowledge graph nodes, which significantly enhances the accuracy of synergy prediction. We introduce a mutual cross-attention mechanism to adaptively fuse local and global features of knowledge graph nodes, which significantly enhances the accuracy of synergy prediction. Materials and methods Problem formulation We formulate the problem of drug–drug synergy (DDS) prediction based on knowledge graphs. Given a set of drugs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{d}$$\end{document} and a set of cell lines \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{c}$$\end{document}, the synergistic effects of drug combinations on specific cell lines can be represented by a three-dimensional binary tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y\in\{0,1\}^{|N_{d}|\times|N_{d}|\times|N_{c}|}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|N_d|$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|N_c|$$\end{document} denote the number of drugs and cell lines, respectively. For drugs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_i,d_j\in N_d$$\end{document} and cell line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c\in N_{c}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y(d_i,d_j,c)=1$$\end{document} indicates that the combination of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{j}$$\end{document} exhibits a synergistic effect on cell line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}, whereas \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y(d_i,d_j,c)=0$$\end{document} denotes an antagonistic interaction. In addition to the observed drug synergy data, we integrate auxiliary information related to drugs and cell lines through a knowledge graph. The knowledge graph \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{G}$$\end{document} is a heterogeneous graph composed of entities and relations, formally defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{G}=(\mathcal{E},\mathcal{R})$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{E}$$\end{document} is the set of entities including drugs, cell lines, proteins, and tissues, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{R}$$\end{document} is the set of relation types, such as drug–protein associations and cell line–protein associations. Each triple \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(s,r,t)$$\end{document} in the knowledge graph represents an association, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s,t\in\mathcal{E}$$\end{document} are the source and target entities, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r\in\mathcal{R}$$\end{document} is the relation type between them. The objective of this study is to predict the synergy of drug combinations on specific cell lines by leveraging the information in the knowledge graph. Specifically, we aim to learn a prediction function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{y}(d_i,d_j,c)=\Gamma(d_i,d_j,c|\theta,y,\mathcal{G})$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{y}(d_i,d_j,c)$$\end{document} denotes the predicted probability that drugs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_j$$\end{document} have a synergistic effect on cell line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Gamma$$\end{document} denotes the entire prediction model, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document} represents the model parameters, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y$$\end{document} is the observed drug synergy matrix, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{G}$$\end{document} is the knowledge graph. Dataset description To evaluate the applicability of KGLGANSynergy on both large-scale and small-scale datasets, we conducted experiments on two widely used benchmark datasets: DrugCombDB and Oncology-Screen.DrugCombDB [35] is a comprehensive database of drug combinations that integrates data from high-throughput screening (HTS) assays, external databases, and literature curated from PubMed. Drug synergy scores in this dataset are computed using the ZIP (Zero Interaction Potency) model. The full dataset contains 764 drugs and 76 cancer cell lines, comprising a total of 69,436 drug combinations.Oncology-Screen [36], curated by O’Neil et al., is derived from large-scale tumor cell line screening and is specifically designed to evaluate drug combination synergy in cancer contexts. This dataset includes 21 drugs and 29 cell lines, covering 4176 drug combinations. Additionally, Preuer et al. [7] processed the dataset by calculating Loewe additivity scores to quantify drug synergy. DrugCombDB [35] is a comprehensive database of drug combinations that integrates data from high-throughput screening (HTS) assays, external databases, and literature curated from PubMed. Drug synergy scores in this dataset are computed using the ZIP (Zero Interaction Potency) model. The full dataset contains 764 drugs and 76 cancer cell lines, comprising a total of 69,436 drug combinations. DrugCombDB [35] is a comprehensive database of drug combinations that integrates data from high-throughput screening (HTS) assays, external databases, and literature curated from PubMed. Drug synergy scores in this dataset are computed using the ZIP (Zero Interaction Potency) model. The full dataset contains 764 drugs and 76 cancer cell lines, comprising a total of 69,436 drug combinations. Oncology-Screen [36], curated by O’Neil et al., is derived from large-scale tumor cell line screening and is specifically designed to evaluate drug combination synergy in cancer contexts. This dataset includes 21 drugs and 29 cell lines, covering 4176 drug combinations. Additionally, Preuer et al. [7] processed the dataset by calculating Loewe additivity scores to quantify drug synergy. Oncology-Screen [36], curated by O’Neil et al., is derived from large-scale tumor cell line screening and is specifically designed to evaluate drug combination synergy in cancer contexts. This dataset includes 21 drugs and 29 cell lines, covering 4176 drug combinations. Additionally, Preuer et al. [7] processed the dataset by calculating Loewe additivity scores to quantify drug synergy. Each sample in the datasets includes a pair of drugs and a cell line, along with a continuous-valued synergy score. To adapt the datasets for our binary classification task, we followed the data partitioning strategy proposed by Zhang et al. [29]. Specifically, the synergy scores were divided using quartiles: samples in the upper quartile were labeled as synergistic, those in the lower quartile as antagonistic, and the remaining middle-range samples were excluded due to low confidence. The resulting dataset statistics are summarized in Table 1. The processed DrugCombDB dataset contains 17,404 synergistic drug pairs and 16,624 antagonistic pairs, and is primarily used to evaluate model performance on large-scale data. The Oncology-Screen dataset contains 1044 synergistic and 916 antagonistic pairs, and serves to assess performance on small-scale data.Table 1Detailed statistics of the processed datasetsDatasetsDrugsCell linesSynergyAntagonismDrugCombDB4757617 40416 624Oncology-Screen21291044916 Table 1Detailed statistics of the processed datasetsDatasetsDrugsCell linesSynergyAntagonismDrugCombDB4757617 40416 624Oncology-Screen21291044916 Detailed statistics of the processed datasets The knowledge graph (KG) data used in this study is based on the dataset curated by Zhang et al. [29], which integrates various types of biomedical information. Specifically, the knowledge graph includes the following components: Drug-protein interaction information collected from databases such as DrugBank, BindingDB, ChEMBL, and PharmGKB, providing curated information on known interactions between drugs and proteins; Cell line–protein associations were derived from the Cancer Cell Line Encyclopedia (CCLE), linking gene expression profiles of cell lines to corresponding proteins; Protein-protein interaction information extracted from the human protein interaction network, revealing relationships between interacting proteins; And Cell line–tissue associations, determined based on the tissue of origin for each cell line, used to represent the relationship between cell lines and their respective tissue types. Together, these entities and relations form the basis of the knowledge graph used in our model. Detailed statistics of the knowledge graph are provided in Table 2.Table 2Detailed statistics of the knowledge graphDrugCombDB_KGOncology-Screen_KGEntities16 82316 026Types of relations44KG triples250 256226 475 Table 2Detailed statistics of the knowledge graphDrugCombDB_KGOncology-Screen_KGEntities16 82316 026Types of relations44KG triples250 256226 475 Detailed statistics of the knowledge graph The framework design In this study, we design a deep learning framework named KGLGANSynergy. The overall architecture of KGLGANSynergy is illustrated in Fig. 1. The model consists of the following key components:Knowledge Graph Construction. We construct a drug knowledge graph using four types of biological associations: drug–protein, cell line–protein, cell line–tissue, and protein–protein interactions. Each entity and relation in the knowledge graph is randomly initialized as a vector representation.Local Graph Attention Network (LGAT). LGAT is designed upon the Edge Enhanced Local Attention (EELA) mechanism [37]. For a given entity, LGAT combines message passing and local attention to learn its vector representation from neighboring nodes in the knowledge graph. In this paper, LGAT concatenates the edge features with the source node features and combines them with the target node features, using an attention mechanism to perform local information aggregation. This design enables the simultaneous fusion of node and edge information in computation, thereby facilitating the learning of richer local representations for each node.Global Graph Attention Network (GGAT). GGAT leverages a Transformer-style attention mechanism [38] and applies it to the knowledge graph to learn global semantic representations of entities. By utilizing global attention, GGAT can capture dependencies between any pair of nodes in a single aggregation step, which is particularly advantageous when applied to large graphs. Through multi-layer global attention, GGAT gains a deeper understanding of the overall graph structure, thereby learning more informative global representations for each node.Mutual Cross-Attention (MCA) Feature Fusion. MCA is a parameter-free feature fusion mechanism that improves model performance without introducing learnable parameters [34]. In this work, we adopt MCA to support the fusion of local and global features in the knowledge graph. Specifically, the local features generated by LGAT and the global features produced by GGAT are used as the input to MCA. Through mutual attention computation between these two representations, MCA enables them to complement and reinforce each other, effectively capturing their mutual dependencies and generating the final feature representation for each entity.Drug Synergy Prediction Layer. The fused entity representations are used to predict whether a given drug pair exhibits synergistic effects on a specific cell line. The features of the two drugs and the cell line are first integrated using max pooling and inner product operations. The resulting feature vector is then passed through a sigmoid activation function to compute the final synergy prediction score.Fig. 1The architecture of KGLGANSynergy. (1) Knowledge graph construction: a knowledge graph is built based on interactions among drugs, cell lines, proteins, and tissues. (2) Local graph attention network (LGAT): local feature representations are generated by aggregating neighborhood information through message passing and attention mechanisms. (3) Global graph attention network (GGAT): attention mechanisms establish dynamic dependencies among all nodes in the graph, enabling each node to integrate global contextual information and output global feature representations. (4) Mutual cross attention network (MCA): a bidirectional attention mechanism adaptively fuses local and global features to produce the final representation of each node. (5) Drug synergy prediction: the features of drug and cell line entities are combined using max pooling and inner product operations, and a sigmoid function is applied to compute the predicted probability of synergy Knowledge Graph Construction. We construct a drug knowledge graph using four types of biological associations: drug–protein, cell line–protein, cell line–tissue, and protein–protein interactions. Each entity and relation in the knowledge graph is randomly initialized as a vector representation. Knowledge Graph Construction. We construct a drug knowledge graph using four types of biological associations: drug–protein, cell line–protein, cell line–tissue, and protein–protein interactions. Each entity and relation in the knowledge graph is randomly initialized as a vector representation. Local Graph Attention Network (LGAT). LGAT is designed upon the Edge Enhanced Local Attention (EELA) mechanism [37]. For a given entity, LGAT combines message passing and local attention to learn its vector representation from neighboring nodes in the knowledge graph. In this paper, LGAT concatenates the edge features with the source node features and combines them with the target node features, using an attention mechanism to perform local information aggregation. This design enables the simultaneous fusion of node and edge information in computation, thereby facilitating the learning of richer local representations for each node. Local Graph Attention Network (LGAT). LGAT is designed upon the Edge Enhanced Local Attention (EELA) mechanism [37]. For a given entity, LGAT combines message passing and local attention to learn its vector representation from neighboring nodes in the knowledge graph. In this paper, LGAT concatenates the edge features with the source node features and combines them with the target node features, using an attention mechanism to perform local information aggregation. This design enables the simultaneous fusion of node and edge information in computation, thereby facilitating the learning of richer local representations for each node. Global Graph Attention Network (GGAT). GGAT leverages a Transformer-style attention mechanism [38] and applies it to the knowledge graph to learn global semantic representations of entities. By utilizing global attention, GGAT can capture dependencies between any pair of nodes in a single aggregation step, which is particularly advantageous when applied to large graphs. Through multi-layer global attention, GGAT gains a deeper understanding of the overall graph structure, thereby learning more informative global representations for each node. Global Graph Attention Network (GGAT). GGAT leverages a Transformer-style attention mechanism [38] and applies it to the knowledge graph to learn global semantic representations of entities. By utilizing global attention, GGAT can capture dependencies between any pair of nodes in a single aggregation step, which is particularly advantageous when applied to large graphs. Through multi-layer global attention, GGAT gains a deeper understanding of the overall graph structure, thereby learning more informative global representations for each node. Mutual Cross-Attention (MCA) Feature Fusion. MCA is a parameter-free feature fusion mechanism that improves model performance without introducing learnable parameters [34]. In this work, we adopt MCA to support the fusion of local and global features in the knowledge graph. Specifically, the local features generated by LGAT and the global features produced by GGAT are used as the input to MCA. Through mutual attention computation between these two representations, MCA enables them to complement and reinforce each other, effectively capturing their mutual dependencies and generating the final feature representation for each entity. Mutual Cross-Attention (MCA) Feature Fusion. MCA is a parameter-free feature fusion mechanism that improves model performance without introducing learnable parameters [34]. In this work, we adopt MCA to support the fusion of local and global features in the knowledge graph. Specifically, the local features generated by LGAT and the global features produced by GGAT are used as the input to MCA. Through mutual attention computation between these two representations, MCA enables them to complement and reinforce each other, effectively capturing their mutual dependencies and generating the final feature representation for each entity. Drug Synergy Prediction Layer. The fused entity representations are used to predict whether a given drug pair exhibits synergistic effects on a specific cell line. The features of the two drugs and the cell line are first integrated using max pooling and inner product operations. The resulting feature vector is then passed through a sigmoid activation function to compute the final synergy prediction score. Drug Synergy Prediction Layer. The fused entity representations are used to predict whether a given drug pair exhibits synergistic effects on a specific cell line. The features of the two drugs and the cell line are first integrated using max pooling and inner product operations. The resulting feature vector is then passed through a sigmoid activation function to compute the final synergy prediction score. The architecture of KGLGANSynergy. (1) Knowledge graph construction: a knowledge graph is built based on interactions among drugs, cell lines, proteins, and tissues. (2) Local graph attention network (LGAT): local feature representations are generated by aggregating neighborhood information through message passing and attention mechanisms. (3) Global graph attention network (GGAT): attention mechanisms establish dynamic dependencies among all nodes in the graph, enabling each node to integrate global contextual information and output global feature representations. (4) Mutual cross attention network (MCA): a bidirectional attention mechanism adaptively fuses local and global features to produce the final representation of each node. (5) Drug synergy prediction: the features of drug and cell line entities are combined using max pooling and inner product operations, and a sigmoid function is applied to compute the predicted probability of synergy Local graph attention network In a knowledge graph, entities and their associations are of various types. For a given entity, its semantic meaning and feature representation should vary depending on the types of entities it is connected to. To accurately capture entity embeddings, we employ LGAT, a message-passing module based on attention mechanisms, designed to enhance the capacity for local information aggregation by taking into account the relationships among all neighboring entities. This module assigns attention weights to each neighbor in the entity’s neighborhood set, reflecting their relative importance. By incorporating both node features and edge features, LGAT enables more accurate modeling of the local structural information within the knowledge graph. As a result, it effectively learns meaningful embedding representations for drugs and cell lines. The model initializes the embeddings of entities and relations in the knowledge graph with randomly generated vectors, whose \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L2$$\end{document} norms are constrained to not exceed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1$$\end{document}. For a given entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s$$\end{document}, its set of neighbors is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{N}(s)=$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{t_1,t_2\cdots t_n\}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_i$$\end{document} represents a neighbor of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s$$\end{document}. The goal of local graph attention embedding propagation is to encode all neighboring entities. In this study, we use a knowledge graph triplet \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(s,r,t)$$\end{document} as an example, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document} denote entities (e.g., drugs, cell lines, proteins, tissues), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r$$\end{document} denotes relationship types (e.g., drug–protein, cell line–protein, cell line–tissue, protein–protein). To compute the attention-based local embedding of an entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}, the process can be intuitively understood as a self-attention mechanism applied at the neighborhood level. LGAT can obtain its representation containing enriched neighborhood information. It first learns attention weights as relevance scores to reflect the importance of each neighbor, allowing the model to selectively focus on the most relevant connections for the given entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}. Then the information from all neighbors is aggregated according to these learned attention weights to generate the local representation. Specifically, the embedding of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document} is first transformed into a query (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q$$\end{document}) vector, which represents the information that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document} is “looking for” from its neighbors. Each neighbor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document} is paired with its relation embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e_{s,t}$$\end{document}, and together they are transformed into key (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K$$\end{document}) and value (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V$$\end{document}) vectors, which describe the information that the neighbor can provide. The similarity between the query of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document} and the keys of its neighbors is then computed, passed through a LeakyReLU function, and normalized with a softmax function to obtain attention weights. These weights indicate how important each neighbor is for representing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}. Finally, the value vectors are aggregated according to these learned weights, producing the local representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^lX_t$$\end{document} for entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}. By applying the above neighborhood-level attention, LGAT allows each entity to adaptively focus on its most relevant neighbors and relations, rather than treating all connections equally. This design provides a more expressive and discriminative representation of local structures in the knowledge graph. The detailed computation is defined as follows: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ Q=f_1(x_t),\quad K,V=f_2([x_s\parallel e_{s,t}]),$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \hat{\pi}_{s,t}=\sigma\left(\frac{QK^T}{\sqrt{d}}\right),$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \pi_{s,t}=\frac{\exp(\hat{\pi}_{s,t})}{\sum_{k\in\mathcal{N}(t)}\exp(\hat{\pi}_{k,t})},$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {}^lX_t=\sum_{s\in\mathcal{N}(t)}\pi_{s,t}V,$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_1\in\mathbb{R}^{F\times F}$$\end{document} is the query projection matrix, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F$$\end{document} denoting the feature dimension. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_2\in\mathbb{R}^{2F\times F}$$\end{document} is the shared projection matrix used for both key and value vectors. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi_{s,t}$$\end{document} represents the attention score between entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s$$\end{document} and its neighbor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document} denotes the LeakyReLU activation function. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d$$\end{document} is the dimensionality of entity representations. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_s\in\mathbb{R}^{1\times F}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_t\in\mathbb{R}^{1\times F}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e_{s,t}\in\mathbb{R}^{1\times F}$$\end{document} denote the embedding vectors of the source entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s$$\end{document}, target entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}, and their relation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r$$\end{document}, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^lX_t\in\mathbb{R}^{1\times F}$$\end{document} is the aggregated local feature vector of entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{N}(t)$$\end{document} denotes the set of one-hop neighbors of entity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}. The operator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\cdot\|\cdot]$$\end{document} indicates concatenation along the feature dimension. To address the limitation that a single-layer LGAT can only aggregate information from first-order neighbors, we adopt a multi-layer LGAT architecture in this study. By stacking multiple attention layers, the model is able to capture higher-order neighborhood information within the knowledge graph, thereby enabling effective modeling of multi-hop relational dependencies.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ ^lX^{(L)}=LGAT\left(^lX^{(L-1)},E^{(L-1)},A\right),\quad L=1,2,...,\alpha,$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^lX^{(L)}\in\mathbb{R}^{N\times F}$$\end{document} denotes the local feature matrix of all entities at the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L$$\end{document}-th layer, with each row representing the feature vector of a node. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document} is the total number of entities. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E^{(L-1)}\in\mathbb{R}^{M\times F}$$\end{document} represents the edge feature matrix from the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(L-1)$$\end{document}-th layer, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M$$\end{document} is the total number of edges. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A\in\mathbb{R}^{2\times M}$$\end{document} is the sparse adjacency matrix of the knowledge graph. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} denotes the total number of LGAT layers, which controls the depth of information propagation. For drug combination prediction, LGAT constructs local feature representations for drugs and cell lines through multi-level neighborhood aggregation. For drug nodes, a single-layer LGAT dynamically aggregates information from their directly associated target proteins, assigning higher attention weights to the most relevant proteins to emphasize their influence. By stacking multiple LGAT layers, deeper hierarchical context can be captured through paths such as drug–protein–cell line–tissue, allowing drug nodes to incorporate contextual information from related cell lines and tissues, thereby enriching their representations. Similarly, for cell line nodes, a single-layer LGAT aggregates information from associated proteins and corresponding tissues. Deeper LGAT layers enable hierarchical aggregation along paths such as cell line–protein–drug, allowing cell lines to incorporate relevant features from related drugs. Through this hierarchical aggregation mechanism, LGAT incrementally enriches the feature representations of both drugs and cell lines within the knowledge graph, ultimately providing high-quality local representations that support accurate drug synergy prediction. Global graph attention network In knowledge graph representation learning, shallow graph neural network (GNN) methods aggregate node information via hierarchical message passing mechanisms. The receptive field of these models is limited by network depth. As illustrated in Fig. 2, taking the drug entity Nilotinib as an example, a single-layer GNN can only integrate information from its directly connected neighbors in the knowledge graph, such as P00519, P10721, and O15197. When using a two-layer GNN, the model can further incorporate second-order neighbors, such as CAOV3, P42684, and SW620, thus expanding the receptive field. However, knowledge graphs often contain long-range dependencies, such as the connection between Nilotinib and lung. Models like KGANSynergy and LGAT, which are based on message passing GNNs, rely on layer-wise propagation to aggregate information. When the number of layers becomes excessively deep, these models tend to introduce additional noise and suffer from performance degradation, which typically limits their architectures to a shallow depth in practice. Consequently, the multi-hop information they capture is confined to local neighborhoods, and long-range dependencies in knowledge graphs cannot be effectively modeled. These long-range relations are critical in knowledge graph representation learning. Due to their inherent architectural constraints, GNNs are limited in their ability to model potential associations between arbitrary entity pairs in the graph. As a result, they struggle to effectively capture the global structural information embedded in knowledge graphs. To address the inability of models such as KGANSynergy and LGAT to capture potential associations between arbitrary nodes in the knowledge graph, this study employs a Global Graph Attention Network (GGAT).Fig. 2Hierarchical partial subgraph of Nilotinib in the knowledge graph, illustrating first-order, second-order, and long-range multi-hop relationships Hierarchical partial subgraph of Nilotinib in the knowledge graph, illustrating first-order, second-order, and long-range multi-hop relationships GGAT is built upon the attention mechanism of the Transformer architecture, with modifications to the attention computation that enable the model to capture long-range dependencies across all nodes in a single step, without relying on multi-hop message passing. Specifically, the global feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{g}X^{(0)}$$\end{document} is first initialized for all entities using the same random embedding strategy as in LGAT. These embeddings are then projected into query (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q$$\end{document}), key (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K$$\end{document}), and value (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V$$\end{document}) vectors through separate linear transformations, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K$$\end{document} are normalized by their Frobenius norms to ensure numerical stability. The term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{K}^T\mathbf{1}$$\end{document} measures the global influence strength of each node as an “information source.” Based on this, a diagonal degree matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D$$\end{document} is constructed, where each diagonal entry incorporates a self-loop term (+1) to preserve a node’s own features and a global attention term that quantifies the total amount of information received from all other nodes in the graph. The global context is obtained by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{K}^T V$$\end{document}, which summarizes information across all nodes. Each query vector then interacts with this global context, yielding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V+\tfrac{1}{N}\widetilde{Q}(\widetilde{K}^T V)$$\end{document} that combines a node’s intrinsic information (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V$$\end{document}) with globally aggregated information. The normalization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D^{-1}[\cdots]$$\end{document} is subsequently applied to scale the aggregated output, preventing uncontrolled growth in feature magnitudes, while the residual connection with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{g}X^{(0)}$$\end{document} helps preserve original features and stabilize training. Unlike message-passing GNNs restricted to local neighborhoods, GGAT captures global dependencies in a single step, enabling each node to learn representations that integrate information from the entire knowledge graph. The detailed computation is defined as follows: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ Q=f_Q(^gX^{(0)}),\quad K=f_K(^gX^{(0)}),\quad V=f_V(^gX^{(0)}),$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \tilde{Q}=\frac{Q}{\parallel{Q}\parallel_\mathcal{F}},\quad \tilde{K}=\frac{K}{\parallel{K}\parallel_\mathcal{F}},$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ D=diag\left(1+\frac{1}{N}\tilde{Q}(\tilde{K}^T\mathbf{1})\right),$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ ^gX=D^{-1}\left[V+\frac{1}{N}\tilde{Q}(\tilde{K}^TV)\right]+{}^gX^{(0)},$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^gX^{(0)}\in\mathbb{R}^{N\times F}$$\end{document} represents the initial feature matrix of all entities in the knowledge graph, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document} denotes the total number of entities, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F$$\end{document} is the feature dimension, and the superscript \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g$$\end{document} indicates that the feature is global. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_Q,f_K,f_V\in$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb{R}^{F\times F}$$\end{document} are linear projection matrices used to generate the query, key, and value matrices, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\|\cdot\|_{\mathcal{F}}$$\end{document} denotes the Frobenius norm, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1$$\end{document} is an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document}-dimensional column vector of ones, and diag is an operation that converts an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document}-dimensional column vector into an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N\times N$$\end{document} diagonal matrix. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D$$\end{document} is a normalization term used to ensure numerical stability. Finally, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^gX\in\mathbb{R}^{N\times F}$$\end{document} represents the global feature matrix of all entities in the knowledge graph, in which each row encodes both the aggregated information from all other entities and the original entity information. A single-layer GGAT can model direct dependencies between any pair of nodes in the graph, but its aggregation capability is limited to establishing direct attention connections. To capture indirect information that requires transmission through intermediate nodes, we adopt a multi-layer GGAT architecture. This mechanism of information propagation via intermediate nodes substantially enhances each node’s ability to perceive its network environment, thereby enabling the model to gain a deeper understanding of the global structure of the graph. The global feature representation computed at the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L$$\end{document}-th layer is defined as follows: 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ ^gX^{(L)}=GGAT\left(^gX^{(L-1)}\right),\quad L=1,2,...,\beta,$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^gX^{(L)}\in\mathbb{R}^{N\times F}$$\end{document} denotes the global feature matrix of all entities at the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L$$\end{document}-th layer, with each row representing the global representation of an entity. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document} is the total number of entities, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F$$\end{document} is the feature dimension, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document} denotes the total number of GGAT layers used to control the depth of global information propagation In the drug combination prediction task, GGAT overcomes the limitations of traditional hierarchical message passing mechanisms by establishing dynamic dependencies across the entire graph. This allows drug and cell line nodes to adaptively capture feature information from any relevant entity in the knowledge graph, thereby compensating for the lack of global contextual information in their representations. Specifically, drug and cell line nodes in the knowledge graph attend to other related entities via attention mechanisms. The attention scores are used to weight and aggregate information from these connected entities, enabling the model to learn global feature representations of drugs and cell lines based on the full structure of the knowledge graph. Mutual cross-attention feature fusion Mutual Cross Attention (MCA) is a feature fusion mechanism designed based on the concept of cross-attention. In this study, MCA is employed to integrate the local and global features of drugs and cell lines learned by the LGAT and GGAT modules, respectively. Unlike traditional cross-attention mechanisms, MCA eliminates the need for learnable linear projection layers. MCA performs parameter-free computations to achieve adaptive fusion between the two types of features, allowing for efficient and effective integration without increasing model complexity. Conventional approaches, such as feature concatenation or linear fusion, typically combine features in a simple way and fail to capture deeper relevance between them. MCA performs bidirectional attention calculations, allowing local and global features to mutually guide each other in identifying their most informative components. This bidirectional mechanism captures feature complementarity more effectively than simple aggregation strategies, enabling the model to adaptively emphasize critical information. Through this parameter-free attention mechanism, MCA can achieve dynamic feature fusion while avoiding the additional complexity introduced by linear fusion methods. Specifically, MCA first performs a directional attention operation, where the local features attend to the global features to selectively aggregate complementary global information. This enables the local representations to be enhanced by the global semantic context. In the opposite direction, the global features attend to the local features, allowing the global representations to incorporate fine-grained local information. The outputs from both attention directions are then combined through element-wise summation to generate the fused representation. This fused representation integrates both local and global information and is used as the final node representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X$$\end{document} in the knowledge graph. The detailed computation is defined as follows:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{Atten}(Q,K,V)=\mathrm{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V,$$\end{document}12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ X = Atten\left(^lX^{(\alpha)},\ ^gX^{(\beta)},\ ^gX^{(\beta)}\right) + Atten\left(^gX^{(\beta)},\ ^lX^{(\alpha)},\ ^lX^{(\alpha)}\right),$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^lX^{(\alpha)}\in\mathbb{R}^{N\times F}$$\end{document} denotes the local feature matrix of knowledge graph entities output by the final layer of LGAT, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^gX^{(\beta)}\in\mathbb{R}^{N\times F}$$\end{document} represents the global feature matrix output by the final layer of GGAT. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X\in\mathbb{R}^{N\times F}$$\end{document} is the fused feature matrix, which serves as the final representation of entities in the knowledge graph after MCA-based fusion. The drug and cell line nodes obtain their local and global feature representations through the LGAT and GGAT modules, respectively. The MCA module adaptively fuses these two types of features via a mutual cross-attention mechanism, producing final representations that capture both local structural information and global semantic dependencies. These fused representations are then directly used as input features for the downstream drug synergy prediction layer. Drug synergy prediction The final fused feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X$$\end{document}, obtained through the MCA module, contains the integrated representations of all entities in the knowledge graph. The feature vector of each entity (e.g., a drug or a cell line) can be directly accessed by its index in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X.$$\end{document} Given a drug pair \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(d_i,d_j)$$\end{document} and a cell line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}, the prediction task is to determine whether the drug combination exhibits a synergistic effect on cell line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}. In the implementation, the feature vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i,x_j$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_c$$\end{document} corresponding to drug \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_i$$\end{document}, drug \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_j$$\end{document}, and cell line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document} are first extracted from the fused feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X.$$\end{document} Then, a max pooling operation is applied element-wise to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_j$$\end{document} to obtain the combined drug feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i,j}.$$\end{document} Next, the inner product between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i,j}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_c$$\end{document} is computed to yield a scalar synergy score. Finally, the synergy score is passed through a sigmoid function to produce the predicted probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}(d_i,d_j,c)\in[0,1]$$\end{document}, which indicates the likelihood that the drug combination \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(d_i,d_j)$$\end{document} exhibits a synergistic effect on cell line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c.$$\end{document} The detailed computation is defined as follows: 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ x_{i,j,c}=MaxPool(x_i,x_j)\odot x_c,$$\end{document}14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \hat{y}(d_i,d_j,c)=Sigmoid(x_{i,j,c}),$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i,x_j,x_c\in\mathbb{R}^{1\times F}$$\end{document} denote the feature vectors of drug \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_i$$\end{document}, drug \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_j$$\end{document}, and cell line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}, respectively. The operator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\odot$$\end{document} represents the inner product operation. This design offers three key advantages. First, the max pooling layer preserves the most salient dimensions of the drug feature vectors, thereby emphasizing the critical information that contributes to drug synergy. Second, the inner product operation enhances the similarity measurement between the drug combination and the cell line features, improving the model’s ability to capture meaningful relational patterns. Third, the sigmoid function produces an interpretable probabilistic output, enabling threshold-based classification for downstream tasks such as binary synergy prediction. Objective function The drug synergy prediction task in this study is formulated as a binary classification problem, for which we adopt the Binary Cross-Entropy Loss (BCELoss) as the training objective. To determine the optimal classification threshold, we employ a data-driven dynamic thresholding strategy based on Youden’s J statistic [39], which identifies the threshold that maximizes the combined performance of sensitivity and specificity. Specifically, the model’s predicted probabilities are evaluated on the training set across a range of thresholds, and the threshold that best balances the identification of positive and negative samples is selected as the final decision threshold. If the predicted synergy probabilities exceed this threshold, the drug pairs are considered synergistic and labeled as 1 (synergistic), while those below are labeled as 0 (antagonistic). Compared with fixed-threshold methods, this adaptive approach is more robust to variations in data distribution and helps mitigate classification bias resulting from class imbalance or task-specific requirements. Considering that parameters in different modules may require different degrees of regularization, we adopt a parameter grouping strategy for the Adam optimizer. Specifically, we apply different weight decay coefficients to the parameters of the LGAT and GGAT modules, respectively. This differentiated optimization scheme helps mitigate the potential conflicts that may arise from applying a single regularization setting to structurally heterogeneous components. The overall loss function of the model is defined as follows: 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \mathcal{L}_{\mathrm{BCELoss}}(y,\hat{y})=-\frac{1}{N}\sum_{i=1}^N[y_i\log(\hat{y}_i)+(1-y_i)\log(1-\hat{y}_i)],$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \mathcal{L}(\theta_1,\theta_2)=\mathcal{L}_{\mathrm{BCELoss}}(y,\hat{y})+\frac{\lambda_1}{2}\|\theta_1\|_2^2+\frac{\lambda_2}{2}\|\theta_2\|_2^2,$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y,\hat{y}\in\{0,1\}$$\end{document} represent the true label and the predicted value of each sample, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta_2$$\end{document} denote the model parameters of LGAT and GGAT, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_2$$\end{document} are the corresponding weight decay coefficients applied to LGAT and GGAT, respectively. Experiments and results In this section, we evaluate the performance of KGLGANSynergy by comparing it with existing baseline methods on two benchmark datasets introduced in section “Dataset description”. In addition, we conduct case studies and other experiments to further demonstrate the effectiveness of KGLGANSynergy in drug synergy prediction. Comparison with previous studies To validate the effectiveness of KGLGANSynergy, we compare it with the following representative baseline methods: DeepWalk, GCN, DeepSynergy, KGNN, GraphSynergy, KGANSynergy, MDNNSyn, and HERMES. A brief description of each method is provided below:DeepWalk [40] is a network embedding method that learns latent representations of vertices by performing random walks to capture local structural information. It maps nodes into a continuous vector space using techniques inspired by language models, generating effective features for graph-based tasks.GCN [41] learns node representations by aggregating information from their neighbors, effectively capturing the structural properties of graphs and dependencies among nodes.DeepSynergy [7] combines chemical and genomic information, and uses normalization strategies and a multilayer perceptron (MLP) to predict synergistic effects of anticancer drug combinations.KGNN [42] extracts drug-related and neighborhood information from a knowledge graph and uses graph neural networks to aggregate high-order structural and semantic relationships, enabling effective prediction of potential drug–drug interactions.GraphSynergy [28] leverages topological information from protein–protein interaction networks and applies attention mechanisms with graph convolutional networks to efficiently predict synergy in anticancer drug combinations.KGANSynergy [29] is a deep learning model based on knowledge graphs and graph attention networks. It integrates multi-source neighborhood information of drugs and cell lines to accurately predict drug synergy.MDNNSyn [43] is a multimodal deep learning framework that integrates biological associations between drugs and cell lines, as well as known synergy data. It employs a hypergraph neural network and semantic similarity strategies to extract multimodal features, which are fused through a gated neural network.HERMES [44] is a novel deep learning framework based on Hypergraph Neural Networks (HGNNs) and a gated residual mechanism. It predicts synergistic anticancer drug combinations by integrating heterogeneous information sources, including drug molecular structures, cell line gene expression profiles, and disease indications. DeepWalk [40] is a network embedding method that learns latent representations of vertices by performing random walks to capture local structural information. It maps nodes into a continuous vector space using techniques inspired by language models, generating effective features for graph-based tasks. DeepWalk [40] is a network embedding method that learns latent representations of vertices by performing random walks to capture local structural information. It maps nodes into a continuous vector space using techniques inspired by language models, generating effective features for graph-based tasks. GCN [41] learns node representations by aggregating information from their neighbors, effectively capturing the structural properties of graphs and dependencies among nodes. GCN [41] learns node representations by aggregating information from their neighbors, effectively capturing the structural properties of graphs and dependencies among nodes. DeepSynergy [7] combines chemical and genomic information, and uses normalization strategies and a multilayer perceptron (MLP) to predict synergistic effects of anticancer drug combinations. DeepSynergy [7] combines chemical and genomic information, and uses normalization strategies and a multilayer perceptron (MLP) to predict synergistic effects of anticancer drug combinations. KGNN [42] extracts drug-related and neighborhood information from a knowledge graph and uses graph neural networks to aggregate high-order structural and semantic relationships, enabling effective prediction of potential drug–drug interactions. KGNN [42] extracts drug-related and neighborhood information from a knowledge graph and uses graph neural networks to aggregate high-order structural and semantic relationships, enabling effective prediction of potential drug–drug interactions. GraphSynergy [28] leverages topological information from protein–protein interaction networks and applies attention mechanisms with graph convolutional networks to efficiently predict synergy in anticancer drug combinations. GraphSynergy [28] leverages topological information from protein–protein interaction networks and applies attention mechanisms with graph convolutional networks to efficiently predict synergy in anticancer drug combinations. KGANSynergy [29] is a deep learning model based on knowledge graphs and graph attention networks. It integrates multi-source neighborhood information of drugs and cell lines to accurately predict drug synergy. KGANSynergy [29] is a deep learning model based on knowledge graphs and graph attention networks. It integrates multi-source neighborhood information of drugs and cell lines to accurately predict drug synergy. MDNNSyn [43] is a multimodal deep learning framework that integrates biological associations between drugs and cell lines, as well as known synergy data. It employs a hypergraph neural network and semantic similarity strategies to extract multimodal features, which are fused through a gated neural network. MDNNSyn [43] is a multimodal deep learning framework that integrates biological associations between drugs and cell lines, as well as known synergy data. It employs a hypergraph neural network and semantic similarity strategies to extract multimodal features, which are fused through a gated neural network. HERMES [44] is a novel deep learning framework based on Hypergraph Neural Networks (HGNNs) and a gated residual mechanism. It predicts synergistic anticancer drug combinations by integrating heterogeneous information sources, including drug molecular structures, cell line gene expression profiles, and disease indications. HERMES [44] is a novel deep learning framework based on Hypergraph Neural Networks (HGNNs) and a gated residual mechanism. It predicts synergistic anticancer drug combinations by integrating heterogeneous information sources, including drug molecular structures, cell line gene expression profiles, and disease indications. We evaluate the model using two distinct five-fold cross-validation strategies: random split and leave-one-drug-pair-out split. In both strategies, the dataset is evenly divided into five folds, ensuring that the ratio of positive to negative samples remains similar to that of the original dataset. As shown in Fig. 3(A), the random split imposes no restrictions, allowing test samples to be drawn from any part of the dataset. In contrast, the leave-one-drug-pair-out split, illustrated in Fig. 3(B), is a stricter partitioning strategy performed at the drug-pair level. It ensures that drug combinations appearing in the test set do not simultaneously occur in the training set, although individual drugs may still appear in combination with other partners during training. For example, as shown in Fig. 3(B), drug combinations involving drug2 and drug5 are selected exclusively for the test set and are completely excluded from the training data. This design provides a more rigorous evaluation of the model’s generalization ability. Model performance is assessed using three widely adopted metrics: AUC (Area Under the ROC Curve), AUPR (Area Under the Precision–Recall Curve), and ACC (Accuracy). No separate validation set is used, and the model checkpoint is selected when the training loss fails to decrease for 15 consecutive epochs for the final evaluation.Fig. 3Data splitting strategies. (A) Random split; (B) Leave-one-drug-pair-out split Data splitting strategies. (A) Random split; (B) Leave-one-drug-pair-out split The experimental results under the leave-one-drug-pair-out setting are shown in Table 3. On the DrugCombDB dataset, KGLGANSynergy achieves an AUC of 0.9084 and an AUPR of 0.9099, outperforming the second-best method, KGANSynergy, by 1.33% and 1.78%, respectively. On the Oncology-Screen dataset, KGLGANSynergy achieves an AUC of 0.8986 and an AUPR of 0.9076, showing improvements of 0.75% and 0.93% over KGANSynergy. The results under the random split setting are presented in Table 4. On the DrugCombDB dataset, KGLGANSynergy achieves an AUC of 0.9340 and an AUPR of 0.9360, exceeding KGANSynergy by 1.52% and 1.63%, respectively. On the Oncology-Screen dataset, it obtains an AUC of 0.9449 and an AUPR of 0.9472, outperforming HERMES, the second-best method, by 0.69% and 0.78%. KGLGANSynergy consistently outperforms all baselines on both datasets, demonstrating its strong capability in drug synergy prediction. This effectiveness is attributed to the LGAT module, which emphasizes local neighborhood aggregation by assigning differentiated attention weights to neighboring nodes, thereby capturing local topological structures and fine-grained feature interactions. Meanwhile, the GGAT module further expands the receptive field to the entire graph, enabling the model to establish long-range dependencies and integrate global semantic information. Finally, the MCA module enhances the overall representation by effectively fusing local and global features in a complementary manner.Table 3Performance comparison of KGLGANSynergy and baselines under leave-one-drug -pair-out evaluationModelDrugCombDBOncology-ScreenAUC-ROCAUC-PRACCAUC-ROCAUC-PRACCDeepWalk0.71010.69230.66740.69870.67850.6496GCN0.69610.67970.65310.67480.70760.6534Deepsynergy0.74360.71920.68510.73600.72050.6695KGNN0.76100.76490.71460.72830.74780.6749GraphSynergy0.78820.77250.72160.78010.79450.6776KGANSynergy0.89510.89210.81740.89110.89830.8221MDNNSyn0.80460.80480.71160.82380.83270.7464HERMES0.81620.81830.72190.84810.85310.7786KGLGANSynergy0.90840.90990.83500.89860.90760.8337The bold values mean the best result for each performance metricTable 4Performance comparison of KGLGANSynergy and baselines under random split evaluationModelDrugCombDBOncology-ScreenAUC-ROCAUC-PRACCAUC-ROCAUC-PRACCDeepWalk0.75210.72390.71060.76650.72920.7002GCN0.73350.71580.69890.74280.79090.6824Deepsynergy0.79620.75830.71220.80840.78290.7257KGNN0.80130.79990.78320.80430.81050.7939GraphSynergy0.83510.81600.75450.84500.84970.7635KGANSynergy0.91880.91970.84470.92390.93140.8577MDNNSyn0.88570.88990.80050.92730.93250.8666HERMES0.89220.89590.80720.93800.93940.8779KGLGANSynergy0.93400.93600.86340.94490.94720.8883The bold values mean the best result for each performance metric Table 3Performance comparison of KGLGANSynergy and baselines under leave-one-drug -pair-out evaluationModelDrugCombDBOncology-ScreenAUC-ROCAUC-PRACCAUC-ROCAUC-PRACCDeepWalk0.71010.69230.66740.69870.67850.6496GCN0.69610.67970.65310.67480.70760.6534Deepsynergy0.74360.71920.68510.73600.72050.6695KGNN0.76100.76490.71460.72830.74780.6749GraphSynergy0.78820.77250.72160.78010.79450.6776KGANSynergy0.89510.89210.81740.89110.89830.8221MDNNSyn0.80460.80480.71160.82380.83270.7464HERMES0.81620.81830.72190.84810.85310.7786KGLGANSynergy0.90840.90990.83500.89860.90760.8337The bold values mean the best result for each performance metric Performance comparison of KGLGANSynergy and baselines under leave-one-drug -pair-out evaluation The bold values mean the best result for each performance metric Table 4Performance comparison of KGLGANSynergy and baselines under random split evaluationModelDrugCombDBOncology-ScreenAUC-ROCAUC-PRACCAUC-ROCAUC-PRACCDeepWalk0.75210.72390.71060.76650.72920.7002GCN0.73350.71580.69890.74280.79090.6824Deepsynergy0.79620.75830.71220.80840.78290.7257KGNN0.80130.79990.78320.80430.81050.7939GraphSynergy0.83510.81600.75450.84500.84970.7635KGANSynergy0.91880.91970.84470.92390.93140.8577MDNNSyn0.88570.88990.80050.92730.93250.8666HERMES0.89220.89590.80720.93800.93940.8779KGLGANSynergy0.93400.93600.86340.94490.94720.8883The bold values mean the best result for each performance metric Performance comparison of KGLGANSynergy and baselines under random split evaluation The bold values mean the best result for each performance metric Parameter sensitivity analysis The number of layers directly influences the receptive depth of the model. In this section, we evaluate the performance of KGLGANSynergy on the DrugCombDB and Oncology-Screen datasets by varying the number of layers in the LGAT and GGAT modules. Experiments are conducted under both the leave-one-drug-pair-out and random split settings. The results are presented in Fig. 4, where subfigures (A) and (B) correspond to the leave-one-drug-pair-out setting, and (C) and (D) to the random split setting. The number of layers is varied within the range {1, 2, 3, 4}. The results indicate that the optimal configuration varies across datasets. On the DrugCombDB dataset, the best performance is achieved when both LGAT and GGAT are set to two layers, whereas on the Oncology-Screen dataset, the optimal results are obtained with a single-layer LGAT and a two-layer GGAT. Compared to the DrugCombDB dataset, the Oncology-Screen dataset is more sensitive to parameter variations. For the LGAT module, a single layer captures information only from direct neighbors, while a multi-layer design enables the aggregation of indirect neighborhood information. For the GGAT module, although a single layer can theoretically aggregate information from all nodes in the knowledge graph, it is limited in its ability to capture indirect dependencies, thereby constraining the model’s expressive capacity. A multi-layer GGAT alleviates this limitation by enabling the integration of higher-order global dependencies and enriching representation learning. However, as the number of layers of LGAT or GGAT increases, additional noise is introduced into the aggregated features, which may degrade performance. Therefore, it is necessary to adopt an appropriate number of layers to balance information aggregation and noise.Fig. 4Layer analysis on the DrugCombDB and Oncology-Screen datasets. (A) and (B) correspond to the leave-one-drug-pair-out setting; (C) and (D) correspond to the random split setting Layer analysis on the DrugCombDB and Oncology-Screen datasets. (A) and (B) correspond to the leave-one-drug-pair-out setting; (C) and (D) correspond to the random split setting Ablation study In this section, we perform a series of ablation experiments on the DrugCombDB and Oncology-Screen datasets, focusing on both feature learning and feature fusion. These experiments aim to assess the contribution of each module in KGLGANSynergy to the overall model performance. The specific components that were ablated are as follows:KGLGANSynergy/-l: This variant removes the LGAT module and uses only the global features extracted by GGAT as the node representations. These global features are directly fed into the prediction module for drug synergy prediction.KGLGANSynergy/-g: This variant removes the GGAT module and relies solely on the local features generated by LGAT for synergy prediction.KGLGANSynergy/-m: This variant replaces the MCA feature fusion module with a traditional feature concatenation followed by a linear layer to fuse the local and global features. KGLGANSynergy/-l: This variant removes the LGAT module and uses only the global features extracted by GGAT as the node representations. These global features are directly fed into the prediction module for drug synergy prediction. KGLGANSynergy/-l: This variant removes the LGAT module and uses only the global features extracted by GGAT as the node representations. These global features are directly fed into the prediction module for drug synergy prediction. KGLGANSynergy/-g: This variant removes the GGAT module and relies solely on the local features generated by LGAT for synergy prediction. KGLGANSynergy/-g: This variant removes the GGAT module and relies solely on the local features generated by LGAT for synergy prediction. KGLGANSynergy/-m: This variant replaces the MCA feature fusion module with a traditional feature concatenation followed by a linear layer to fuse the local and global features. KGLGANSynergy/-m: This variant replaces the MCA feature fusion module with a traditional feature concatenation followed by a linear layer to fuse the local and global features. The ablation results are summarized in Tables 5 and 6, where all model variants demonstrate varying degrees of performance degradation, underscoring the necessity of each component in the overall architecture. Specifically, KGLGANSynergy/-l, which removes the LGAT module, exhibits the most substantial performance drop. This finding highlights the critical role of local feature learning in drug synergy prediction and suggests that neighborhood-level information offers stronger representational capacity than long-range dependencies in this context. In contrast, KGLGANSynergy/-g, which excludes the GGAT module, shows a more moderate decline, indicating that global features, while beneficial, primarily serve as a complement to local representations. Notably, the performance of KGLGANSynergy/-m reveals that the MCA-based feature fusion significantly outperforms traditional fusion methods, confirming its effectiveness in integrating local and global features. These results also reinforce the view that global features can compensate for the limited receptive field of local features, leading to more comprehensive and expressive entity representations.Table 5Ablation results of KGLGANSynergy on the DrugCombDB dataset under leave-one- drug-pair-out and random split settingsModelLeave-one-drug-pair-out splitRandom splitAUC-ROCAUC-PRACCAUC-ROCAUC-PRACCKGLGANSynergy/-l0.83790.84170.76160.84180.84410.7652KGLGANSynergy/-g0.89920.89970.82650.92580.92750.8561KGLGANSynergy/-m0.90150.90250.82710.92840.92900.8569KGLGANSynergy0.90840.90990.83500.93400.93600.8634The bold values mean the best result for each performance metricTable 6Ablation results of KGLGANSynergy on the Oncology-Screen dataset under leave-one- drug-pair-out and random split settingsModelLeave-one-drug-pair-out splitRandom splitAUC-ROCAUC-PRACCAUC-ROCAUC-PRACCKGLGANSynergy/-l0.79380.81580.75050.81090.82930.7559KGLGANSynergy/-g0.87790.88720.81310.93530.93700.8745KGLGANSynergy/-m0.89500.90120.83000.94150.94210.8811KGLGANSynergy0.89860.90760.83370.94490.94720.8883The bold values mean the best result for each performance metric Table 5Ablation results of KGLGANSynergy on the DrugCombDB dataset under leave-one- drug-pair-out and random split settingsModelLeave-one-drug-pair-out splitRandom splitAUC-ROCAUC-PRACCAUC-ROCAUC-PRACCKGLGANSynergy/-l0.83790.84170.76160.84180.84410.7652KGLGANSynergy/-g0.89920.89970.82650.92580.92750.8561KGLGANSynergy/-m0.90150.90250.82710.92840.92900.8569KGLGANSynergy0.90840.90990.83500.93400.93600.8634The bold values mean the best result for each performance metric Ablation results of KGLGANSynergy on the DrugCombDB dataset under leave-one- drug-pair-out and random split settings The bold values mean the best result for each performance metric Table 6Ablation results of KGLGANSynergy on the Oncology-Screen dataset under leave-one- drug-pair-out and random split settingsModelLeave-one-drug-pair-out splitRandom splitAUC-ROCAUC-PRACCAUC-ROCAUC-PRACCKGLGANSynergy/-l0.79380.81580.75050.81090.82930.7559KGLGANSynergy/-g0.87790.88720.81310.93530.93700.8745KGLGANSynergy/-m0.89500.90120.83000.94150.94210.8811KGLGANSynergy0.89860.90760.83370.94490.94720.8883The bold values mean the best result for each performance metric Ablation results of KGLGANSynergy on the Oncology-Screen dataset under leave-one- drug-pair-out and random split settings The bold values mean the best result for each performance metric Case study To further assess the model’s ability to identify unknown synergistic drug–cell line combinations, we conducted a case study on the DrugCombDB dataset. The latest global statistics from the World Health Organization (WHO) in 2024 report cancer incidence and mortality across 36 cancers in 185 countries worldwide [45]. In this study, common cancers refer to those ranking in the upper half of all cancers in terms of incidence rates, as they impose a relatively higher disease burden compared to other cancers. We searched the medical literature in PubMed related to common cancers (e.g., breast, lung, and ovarian cancers) and selected 10 publications reporting synergistic drug combinations. The PubMed IDs (PMIDs) of these studies are listed in Table 7. These publications involve 12 drugs and 7 cell lines, from which we derived 462 candidate drug pair–cell line combinations through combinatorial enumeration. To evaluate novel predictions, we removed any overlapping combinations from the DrugCombDB dataset and used the remaining data to train the model. The trained KGLGANSynergy model was then applied to predict synergy scores for the candidate combinations, which were ranked in descending order. Remarkably, 6 out of the top 10 predictions were confirmed by published biomedical studies, demonstrating the model’s strong potential for real-world drug synergy discovery. Descriptions of the validated cases are provided below:For drug-resistant breast cancer cell line MCF-7/ADR, a co-delivered nanocrystal formulation of paclitaxel and lapatinib significantly enhanced intracellular uptake of paclitaxel by optimizing the drug ratio, activating pro-apoptotic mechanisms, and overcoming drug resistance. This combination demonstrated notable synergistic antitumor effects in both in vitro and potentially in vivo models [46].By formulating docetaxel as an inhalable aerosol and combining it with the COX-2 inhibitor celecoxib, the treatment exhibited enhanced in vitro cytotoxicity against human non-small cell lung cancer (NSCLC) A549 cells, induced apoptosis, and upregulated the expression of PPAR-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document}, providing a novel strategy for inhalation-based lung cancer therapy [47].In patients with non-small cell lung cancer (NSCLC), the combination of docetaxel and tamoxifen effectively reversed P-glycoprotein–mediated multidrug resistance, increased intracellular accumulation of chemotherapeutic agents, and subsequently improved tumor response rates, progression-free survival, and overall survival, offering significant clinical benefits for advanced NSCLC treatment [48].In human ovarian cancer cell lines, the combination of ruxolitinib and paclitaxel significantly enhanced the inhibition of the JAK/STAT3 signaling pathway by ruxolitinib and promoted the expression of apoptosis-related proteins, resulting in marked therapeutic effects in both in vitro and in vivo ovarian cancer models [49].The combination of docetaxel and imatinib significantly increased cytotoxicity against Philadelphia chromosome–positive chronic myeloid leukemia (CML) K562 cells. This was achieved by upregulating caspase-3 activity and decreasing mitochondrial membrane potential, thereby inducing apoptosis and enhancing chemosensitivity to imatinib [50].In K562 cells, co-treatment with amifostine and imatinib significantly amplified imatinib-induced apoptotic signaling. This effect was mediated by upregulating genes such as Rho6, PPP2R5E, PPM1E, and BTF, and downregulating anti-apoptotic genes such as API5, TUBB2, and TLK1, demonstrating potent anti-leukemic activity in in vitro models of chronic myeloid leukemia [51].Table 7Supporting literature for unknown drug–cell line combinations identified by KGLGANSynergyCell lineTissueDrug 1Drug 2PMIDTrueT47DBreastFulvestrantBortezomib21292820MCF7BreastLetrozoleCelecoxib28643236MCF7BreastPaclitaxelLapatinib32019194\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}MDAMB231BreastCelecoxibPaclitaxel37548777A549LungCelecoxibDocetaxel16536899\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}A549LungTamoxifenDocetaxel26882453\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SKOV3OvaryPaclitaxelSorafenib37111691OVCAR8OvaryRuxolitinibPaclitaxel29849942\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}K562Peripheral bloodImatinibDocetaxel19490758\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}K562Peripheral bloodImatinibAmifostine17009037\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document} For drug-resistant breast cancer cell line MCF-7/ADR, a co-delivered nanocrystal formulation of paclitaxel and lapatinib significantly enhanced intracellular uptake of paclitaxel by optimizing the drug ratio, activating pro-apoptotic mechanisms, and overcoming drug resistance. This combination demonstrated notable synergistic antitumor effects in both in vitro and potentially in vivo models [46]. For drug-resistant breast cancer cell line MCF-7/ADR, a co-delivered nanocrystal formulation of paclitaxel and lapatinib significantly enhanced intracellular uptake of paclitaxel by optimizing the drug ratio, activating pro-apoptotic mechanisms, and overcoming drug resistance. This combination demonstrated notable synergistic antitumor effects in both in vitro and potentially in vivo models [46]. By formulating docetaxel as an inhalable aerosol and combining it with the COX-2 inhibitor celecoxib, the treatment exhibited enhanced in vitro cytotoxicity against human non-small cell lung cancer (NSCLC) A549 cells, induced apoptosis, and upregulated the expression of PPAR-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document}, providing a novel strategy for inhalation-based lung cancer therapy [47]. By formulating docetaxel as an inhalable aerosol and combining it with the COX-2 inhibitor celecoxib, the treatment exhibited enhanced in vitro cytotoxicity against human non-small cell lung cancer (NSCLC) A549 cells, induced apoptosis, and upregulated the expression of PPAR-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document}, providing a novel strategy for inhalation-based lung cancer therapy [47]. In patients with non-small cell lung cancer (NSCLC), the combination of docetaxel and tamoxifen effectively reversed P-glycoprotein–mediated multidrug resistance, increased intracellular accumulation of chemotherapeutic agents, and subsequently improved tumor response rates, progression-free survival, and overall survival, offering significant clinical benefits for advanced NSCLC treatment [48]. In patients with non-small cell lung cancer (NSCLC), the combination of docetaxel and tamoxifen effectively reversed P-glycoprotein–mediated multidrug resistance, increased intracellular accumulation of chemotherapeutic agents, and subsequently improved tumor response rates, progression-free survival, and overall survival, offering significant clinical benefits for advanced NSCLC treatment [48]. In human ovarian cancer cell lines, the combination of ruxolitinib and paclitaxel significantly enhanced the inhibition of the JAK/STAT3 signaling pathway by ruxolitinib and promoted the expression of apoptosis-related proteins, resulting in marked therapeutic effects in both in vitro and in vivo ovarian cancer models [49]. In human ovarian cancer cell lines, the combination of ruxolitinib and paclitaxel significantly enhanced the inhibition of the JAK/STAT3 signaling pathway by ruxolitinib and promoted the expression of apoptosis-related proteins, resulting in marked therapeutic effects in both in vitro and in vivo ovarian cancer models [49]. The combination of docetaxel and imatinib significantly increased cytotoxicity against Philadelphia chromosome–positive chronic myeloid leukemia (CML) K562 cells. This was achieved by upregulating caspase-3 activity and decreasing mitochondrial membrane potential, thereby inducing apoptosis and enhancing chemosensitivity to imatinib [50]. The combination of docetaxel and imatinib significantly increased cytotoxicity against Philadelphia chromosome–positive chronic myeloid leukemia (CML) K562 cells. This was achieved by upregulating caspase-3 activity and decreasing mitochondrial membrane potential, thereby inducing apoptosis and enhancing chemosensitivity to imatinib [50]. In K562 cells, co-treatment with amifostine and imatinib significantly amplified imatinib-induced apoptotic signaling. This effect was mediated by upregulating genes such as Rho6, PPP2R5E, PPM1E, and BTF, and downregulating anti-apoptotic genes such as API5, TUBB2, and TLK1, demonstrating potent anti-leukemic activity in in vitro models of chronic myeloid leukemia [51]. In K562 cells, co-treatment with amifostine and imatinib significantly amplified imatinib-induced apoptotic signaling. This effect was mediated by upregulating genes such as Rho6, PPP2R5E, PPM1E, and BTF, and downregulating anti-apoptotic genes such as API5, TUBB2, and TLK1, demonstrating potent anti-leukemic activity in in vitro models of chronic myeloid leukemia [51]. Table 7Supporting literature for unknown drug–cell line combinations identified by KGLGANSynergyCell lineTissueDrug 1Drug 2PMIDTrueT47DBreastFulvestrantBortezomib21292820MCF7BreastLetrozoleCelecoxib28643236MCF7BreastPaclitaxelLapatinib32019194\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}MDAMB231BreastCelecoxibPaclitaxel37548777A549LungCelecoxibDocetaxel16536899\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}A549LungTamoxifenDocetaxel26882453\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SKOV3OvaryPaclitaxelSorafenib37111691OVCAR8OvaryRuxolitinibPaclitaxel29849942\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}K562Peripheral bloodImatinibDocetaxel19490758\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}K562Peripheral bloodImatinibAmifostine17009037\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document} Supporting literature for unknown drug–cell line combinations identified by KGLGANSynergy In summary, the above experimental findings collectively highlight the potential of KGLGANSynergy to predict novel and clinically relevant synergistic drug combinations for future research and therapeutic development. Conclusion In this study, we proposed KGLGANSynergy, a knowledge graph–based deep learning framework designed to address the limitations of existing graph-based approaches in drug synergy prediction. The model introduces three core components that contribute to its effectiveness: The Local Graph Attention Network (LGAT) leverages fine-grained edge feature modeling and a neighborhood aggregation mechanism to precisely capture local interaction patterns between drugs and cell lines. The Global Graph Attention Network (GGAT) overcomes the limitations of traditional message-passing schemes by establishing graph-wide dependencies and integrating long-range semantic information. The Mutual Cross Attention (MCA) module enables adaptive fusion of local and global features in a novel and effective manner. This architecture effectively addresses two major limitations of existing methods: insufficient edge feature learning and limited ability to capture long-range dependencies. Experimental results demonstrate that KGLGANSynergy consistently outperforms existing baseline models on the DrugCombDB and Oncology-Screen datasets. Furthermore, the case study confirms the model’s capability to identify unknown yet biologically meaningful drug combinations. Overall, this work provides a more powerful computational tool for drug synergy prediction and holds broad potential for applications in biomedical research and drug development. While knowledge graphs are effective in representing pairwise interactions between biological entities such as drugs, their inherent graph structure limits their ability to directly model higher-order associations, such as those between synergistic drug pairs and specific cell lines. In contrast, hypergraphs are naturally suited for representing complex interactions among multiple entities. They enable joint modeling of a drug pair and its associated cell line as a unified relational unit. Based on this insight, future work could explore dual-view representation learning by integrating knowledge graphs and hypergraphs. Specifically, incorporating a hypergraph attention mechanism could help quantify the synergistic impact of different drug combinations on cell lines, providing a more powerful framework for predicting effective therapies. This direction holds promise for overcoming the representational limitations of traditional binary relation graphs and may introduce a novel computational paradigm for combinatorial drug screening and precision therapy design.