Interpretable graph deep learning framework for drug synergy prediction by integrating functional and clinical similarities Authors: Jiyin Lai, Jiashuo Wu, Yalan He, Yongbao Zhang, Bin Li, Tingyu Shi, Junwei Han Published: 2026 ABSTRACT Tumor heterogeneity and drug resistance limit single-agent therapies, making combination treatments essential. However, traditional screening methods are costly and inefficient. Here, we present DSimSynergy, a graph deep learning framework for predicting drug synergy. It constructs drug similarity networks from biological process and clinical applications, then learns drug representations through graph convolution on these networks. Subsequently, it combines them with graph attention representations of drug molecular fingerprints and cell line gene expressions to predict synergy scores for drug combinations. Comprehensive benchmarking on multiple independent datasets demonstrates that DSimSynergy consistently outperforms state-of-the-art methods. Model interpretability analysis revealed key genes and pathways underlying drug synergy, while validation on clinical patient and cohort data demonstrated good clinical translational potential and discovered the molecular mechanisms by which drugs generate synergistic effects through “pathway complementary networks”. DSimSynergy efficiently identifies synergistic combinations, reducing experimental costs while elucidating biological mechanisms to overcome resistance and guide personalized treatment. Introduction With the continuous advancement in oncology research, our understanding of tumor development mechanisms has become increasingly profound, culminating in the successful development of numerous anticancer drugs and a series of computational methods have been developed to predict anticancer drugs1. However, tumor heterogeneity and the emergence of drug resistance during treatment have progressively diminished the efficacy of single-agent therapies targeting individual molecular targets2. Consequently, drug combination therapy has emerged as a pivotal strategy in cancer treatment. This approach simultaneously targets multiple molecular targets and signaling pathways, eliciting synergistic biological interactions that significantly enhance therapeutic efficacy, minimize adverse effects, and effectively circumvent drug resistance3. The cornerstone of combination therapy lies in achieving synergistic effects, where combined drugs produce therapeutic benefits exceeding the sum of individual effects. However, many drug combinations prove ineffective or even antagonistic. Such inappropriate combinations not only fail to achieve anticipated therapeutic outcomes but may also compromise patient health. Consequently, accurately predicting synergistic combinations for individual patients represents a major challenge in personalized cancer therapy4. Traditional screening methods, which rely primarily on clinical trials, are time-consuming, inefficient, and prohibitively expensive5, and are therefore considerably limited in their ability to explore the vast space of possible drug combinations. As a result, it is imperative to develop robust and rapid computational approaches to facilitate the identification of effective combination therapies. Recently, numerous computational approaches have been developed for drug combination synergy prediction. These approaches can be broadly categorized into two types: (i) high-throughput screening (HTS)-based methods, which primarily investigate anti-cancer drugs based on their therapeutic efficacy measured at the cell line level, and (ii) knowledge graph (KG)-based methods, which leverage structured biomedical knowledge to explore a more diverse range of drug combinations6. Compared with KG-based methods, HTS-based methods benefit from standardized experimental data with quantitative synergy scores, enabling more accurate regression tasks and direct validation against in vitro outcomes. Several HTS-based methods have been developed to predict drug synergy. For example, Sidorov et al. used chemical features such as fingerprints or descriptors to build cell line-specific models, predicting combination scores with extreme gradient boosting (XGBoost)7. Li et al. employed logistic regression to identify statistically significant gene and pathway features in drug interactions for predicting drug combination synergy8. However, these traditional machine learning approaches face notable limitations: training machine learning models with superior predictive performance requires manual feature selection and construction, while their relatively low computational efficiency hinders large-scale drug combination prediction within short timeframes9. The rapid development of deep learning technologies has provided novel solutions to these challenges. It can autonomously learn and optimize input raw features, thereby reducing dependence on manual feature selection and enabling faster and more effective learning of complex nonlinear associations within data9. DeepSynergy10 was a typical deep learning method, which employed normalization to address data heterogeneity and leveraged convolutional neural networks to capture representations of drug chemical structures and cell line genomic expression data. DeepDDS11 leveraged graph convolution and attention mechanisms to learn the representations of drugs represented as molecular graphs, capturing richer structural features. AttenSynergy12 utilized attention-based graph neural networks to learn drug pair interactions and enhance key feature representations. HyperGraphSynergy13 constructed hypergraph models with drugs and cell lines as nodes, learning embedded features through hypergraph neural networks. JointSyn14 utilized synergy data, drug features, and cell line features as inputs, extracting features related to drug synergy from two distinct perspectives and combining graph attention networks to learn embedded features. Nevertheless, existing methods predominantly optimize drug chemical structures and cell line genomics while inadequately incorporating the complex biological processes involved in drug combinations as well as their clinical relevance. Actually, drug synergy is indeed associated with biological process similarity and clinical similarity between drugs15. Furthermore, most graph neural networks (GNN) based methods operate on individual drug molecular structure graphs, capturing only intra-drug molecular information and failing to model inter-drug relationships. Motivated by the established correlations among drug biological processes, clinical similarity, and combination therapy outcomes, we present DSimSynergy, a novel HTS-based framework that leverages drug functional and clinical similarity networks to predict drug synergy in cancer cell lines. DSimSynergy constructs a drug similarity network based on biological process (BP) and a drug similarity network based on the Anatomical Therapeutic Chemical (ATC)16 to characterize at biological functional and clinical level among drugs. DSimSynergy employs drug molecular fingerprints as initial node embeddings and utilizes graph convolutional networks (GCN)17 to learn neighborhood information from similarity networks for optimizing drug representations. The model simultaneously extracts features from drug molecular structures via graph attention networks (GAT)18 and from cell line gene expression through fully connected neural networks (FCNN). For each drug-drug-cell line triplet, DSimSynergy integrates features from drug similarity, molecular structure, and cell line features to predict synergy scores. We comprehensively validated DSimSynergy using internal and multiple external datasets. Our results demonstrate that the model exhibits robust predictive performance, consistently outperforms eight state-of-the-art methods. Its interpretability provides mechanistic insights into drug synergy, offering a theoretical framework for understanding drug interactions and resistance mechanisms. Furthermore, the model delivers reliable predictions of drug synergy from both single-cell and bulk sequencing data of cancer patients, presenting a valuable tool for guiding personalized therapeutic decisions. Results Overview of the DSimSynergy framework This study introduces DSimSynergy, a novel computational framework designed to predict drug synergy in cancer cell lines (Fig. 1). Building upon HTS-based cell-line synergy regression, DSimSynergy operates on drug-drug-cell line triplets and performs two distinct prediction tasks: binary synergy classification and continuous synergy score predictions (see “Materials and methods” for details). The framework comprises four modules: (1) constructing drug functional and clinical similarity networks using molecular fingerprints as initial drug node embeddings; (2) learning drug representations from similarity networks and molecular structure graphs; (3) cell line representation learning; and (4) predicting synergistic effects from drug-drug-cell line triplet representations. DSimSynergy could predict both drug synergy classifications and synergy scores. The model was trained on 66,273 triplets from the DrugCombDB dataset which were randomly divided into training, validation, and test sets (see “Materials and methods” for details). Details of hyperparameter selection, including learning rate and batch size, are provided in the Supplementary Note 1, with the final optimal configurations summarized in Supplementary Table 1.Fig. 1Schematic workflow of the DSimSynergy model.The model comprises four modules: A Construction of drug-drug similarity networks. Two independent drug similarity networks were constructed based on drug target information and ATC classification codes. The initial embeddings for drugs were derived from their SMILES representations encoded as one-hot vectors using the MACCS fingerprint. B Drug feature extraction. This component contains three drug feature learning modules: updating drug features through GCN on the similarity network based on biological process, updating drug features through GCN on the similarity network based on ATC, and learning associations between drug molecules through GAT with the resulting molecular representations reduced in dimensionality via global maximum pooling (GMP). C Cell line feature extraction module. Cell line gene expression data undergo feature optimization and dimensionality reduction through FCNN. D Drug combination synergy prediction module. The three types of drug features and cell line features are concatenated according to triplet relationships. The concatenated vectors are input into FCNN for prediction, outputting synergy scores (Loewe scores) or synergy classification results. Schematic workflow of the DSimSynergy model. The model comprises four modules: A Construction of drug-drug similarity networks. Two independent drug similarity networks were constructed based on drug target information and ATC classification codes. The initial embeddings for drugs were derived from their SMILES representations encoded as one-hot vectors using the MACCS fingerprint. B Drug feature extraction. This component contains three drug feature learning modules: updating drug features through GCN on the similarity network based on biological process, updating drug features through GCN on the similarity network based on ATC, and learning associations between drug molecules through GAT with the resulting molecular representations reduced in dimensionality via global maximum pooling (GMP). C Cell line feature extraction module. Cell line gene expression data undergo feature optimization and dimensionality reduction through FCNN. D Drug combination synergy prediction module. The three types of drug features and cell line features are concatenated according to triplet relationships. The concatenated vectors are input into FCNN for prediction, outputting synergy scores (Loewe scores) or synergy classification results. Evaluation of DSimSynergy and comparison with state-of-the-art methods To determine the optimal gene set for cell line representation learning, we compared model performance using the L1000 landmark genes (949 genes) versus all protein-coding genes (19,193 genes) from DepMap. The model based on the L1000 set achieved better and more stable performance across all evaluated metrics and was therefore adopted for all subsequent analyses (Supplementary Fig. 1; Supplementary Table 2). For the classification prediction on internal test set, DSimSynergy achieved AUROC, AUPR, ACC, and F1 scores of 0.96 ± 0.01, 0.77 ± 0.01, 0.98 ± 0.01, and 0.74 ± 0.02, respectively (Fig. 2A–D). We compared DSimSynergy against five existing deep learning methods (JointSyn14, HyperGraphSynergy13, AttenSynergy12, DeepDDS11 and DeepSynergy10) and three traditional machine learning methods, such as RF, Elastic Net and Support Vector Machine. DSimSynergy surpassing the second-best model by 3%-6% across all metrics (Fig. 2A–D; Supplementary Table 3). Notably, on independent external test datasets (DrugComb and AZ-DREAM), DSimSynergy’s performance advantage increased, with metrics leading the second-best model by 5–10% (Fig. 2A–D; Supplementary Tables 4 and 5). The performance margins over the two strongest baselines, JointSyn and HyperGraphSynergy, were relatively modest, which is expected given that these metrics were already approaching their upper bounds (AUROC > 0.95 and ACC > 0.97). For synergy score prediction on internal test set, DSimSynergy’s predictions showed a strong positive correlation with observed scores (PCC = 0.87 ± 0.01, RMSE = 8.32 ± 0.57). And these results outperformed JointSyn, HyperGraphSynergy, and DeepSynergy, which possess capabilities for synergy score prediction (Fig. 2E; Supplementary Table 3). Moreover, on two independent external test sets (DrugComb and AZ-DREAM), DSimSynergy also outperformed other models (Fig. 2F, G; Supplementary Tables 4 and 5).Fig. 2Predictive performance evaluation of DSimSynergy.A–D Performance comparison for classification prediction tasks. Performance comparison of DSimSynergy with eight other classification prediction models across three datasets. A AUROC metric comparison; B AUPR metric comparison; C ACC metric comparison; D F1 metric comparison. E–G Performance comparison for synergy score prediction tasks. Performance comparison of DSimSynergy with three other deep learning methods for synergy score prediction. E Performance comparison on the DrugCombDB internal test set; F Performance comparison on the DrugComb external test set; G Performance comparison on the AZ-DREAM external test set. Predictive performance evaluation of DSimSynergy. A–D Performance comparison for classification prediction tasks. Performance comparison of DSimSynergy with eight other classification prediction models across three datasets. A AUROC metric comparison; B AUPR metric comparison; C ACC metric comparison; D F1 metric comparison. E–G Performance comparison for synergy score prediction tasks. Performance comparison of DSimSynergy with three other deep learning methods for synergy score prediction. E Performance comparison on the DrugCombDB internal test set; F Performance comparison on the DrugComb external test set; G Performance comparison on the AZ-DREAM external test set. Furthermore, DSimSynergy rapidly converged during training for both binary synergy classification and synergy score prediction tasks (Supplementary Fig. 2A, B), with reduced training time per epoch compared to other deep learning methods (Supplementary Fig. 2C). DSimSynergy also demonstrated balanced predictive performance across cancer tissues (test set cell lines encompassed 14 tissue types), achieving PCC > 0.75 for all tissues except liver (Supplementary Fig. 2D). Lung tissue exhibited excellent performance with stable cross-validation results. For cancer tissues with relatively poor prediction performance, the possible reason lies in insufficient tissue sample data, leading to inadequate model training. Model ablation study and generalization capability assessment Our model comprises three drug feature subsets: drug features derived from a drug similarity network based on biological process, drug features derived from a drug similarity network based on ATC, drug features from drug molecular structures. To rigorously evaluate the contribution of each component, we conducted a comprehensive ablation study. For the remaining three drug feature subsets, we systematically removed each one individually. In addition, to disentangle the contribution of the similarity network architecture from the underlying biological features themselves, we designed two additional control conditions in which BP and ATC features were provided directly as raw input without being processed through the similarity networks (see Supplementary Note 2 for details). Five-fold cross-validation was performed for each new combination model. The complete model demonstrated optimal performance in both binary synergy classification prediction (Fig. 3A) and synergy score prediction tasks (Supplementary Fig. 3). For both tasks, paired t-test results showing p < 0.01 when compared to other combinations, indicating that each feature subset contributes substantially to model performance. Notably, all high-performing combinations included drug structures representation, which are also common characteristics shared by most deep learning models for drug synergy prediction. Furthermore, combining structural representations with similarity network features enhanced performance. Importantly, models utilizing GCN-refined drug features from similarity networks consistently outperformed those using raw BP or ATC as direct input. These results collectively demonstrate that the performance gains stem not merely from the incorporation of additional biological information and clinical information, but from the similarity network architecture’s capacity to learn and propagate relational drug–drug associations, thereby capturing latent drug features that are highly relevant to synergy prediction.Fig. 3Ablation study and generalization capability evaluation of DSimSynergy.A Ablation study performance. Performance of different module combinations in classification prediction. Black solid circles indicate that the model includes the feature module, while gray solid circles indicate that the model does not include the feature module. B–D Generalization capability evaluation on stratified datasets. Performance comparison of DSimSynergy with other methods under three data stratification strategies. Number of Wins represents the number of times the method outperformed other methods in performance metrics; red represents PairOut, blue represents CellOut, and yellow represents DrugOut; circle size represents the magnitude of performance scores. B Performance on the DrugCombDB test set; C Performance on the DrugComb external test set; D Performance on the AZ-DREAM external test set. Ablation study and generalization capability evaluation of DSimSynergy. A Ablation study performance. Performance of different module combinations in classification prediction. Black solid circles indicate that the model includes the feature module, while gray solid circles indicate that the model does not include the feature module. B–D Generalization capability evaluation on stratified datasets. Performance comparison of DSimSynergy with other methods under three data stratification strategies. Number of Wins represents the number of times the method outperformed other methods in performance metrics; red represents PairOut, blue represents CellOut, and yellow represents DrugOut; circle size represents the magnitude of performance scores. B Performance on the DrugCombDB test set; C Performance on the DrugComb external test set; D Performance on the AZ-DREAM external test set. We further evaluated generalization capability using three stratified data partitioning strategies: PairOut, CellOut and DrugOut (see Supplementary Note 3 for details). Model performance declined under all stratification strategies compared to random partitioning on internal test set. For PairOut, the model could still leverage shared drug information across different combinations, resulting in only modest performance decreases, with DSimSynergy surpassing the second-best method by 2–5% across metrics. Under CellOut, performance declined further, outperforming the second-best method by 2–5% across metrics. DrugOut posed the most challenging scenario, simulating a prediction task involving entirely unseen drugs that are absent from the training data, under which the performance of all methods declined substantially. Nevertheless, DSimSynergy consistently achieved the best performance under the DrugOut scenario, outperforming the second-best method by 3–7% across all evaluation metrics. (Fig. 3B; Supplementary Fig. 4A; Supplementary Table 6). Notably, DSimSynergy also achieved the highest performance across most metrics under all three stratification strategies on both external test sets. On the DrugComb external test set, DSimSynergy surpassed the second-best model by margins of 3–6% across metrics under PairOut and CellOut, and by 4–7% under the DrugOut condition. Similarly, on the AZ-DREAM dataset, DSimSynergy maintained a slight but consistent lead over competing methods across all stratification strategies (Fig. 3C, D; Supplementary Fig. 4B, C; Supplementary Table 6). These results demonstrate that the constructed drug similarity networks can effectively learn transferable latent drug feature representations, enabling the model to maintain competitive generalization performance even when encountering novel drugs not present in the training set. To further expand the applicability of our model, we applied Similarity ensemble approach (SEA)19 to predict targets for drugs that have ATC information and are present in DrugCombDB but lack annotated target data (see Supplementary Note 4 for details). This allowed us to incorporate and retain a larger portion of the synergy dataset. We retrained the model using the same framework on the expanded dataset and evaluated performance on three subsets separately: (i) all drug triplets (All), (ii) drug triplets with experimentally validated known targets (Known Target), and (iii) drug triplets with SEA-predicted targets (Pred Target). The model performs well across all drug triplets. For drug triplets with known targets, performance remained consistent with that of the model prior to dataset expansion. For drug triplets with predicted targets, the model still achieved competitive results, maintaining high AUROC and ACC (Supplementary Fig. 5; Supplementary Table 7; AUROC = 0.97 and ACC = 0.98), while the relatively larger fluctuations in F1 and AUPR likely reflect the inherent uncertainty of computationally inferred targets. These results suggest that our model can be extended using predicted drug target information, substantially improving its practical applicability in real-world scenarios. To further validate the robustness of DSimSynergy, we evaluated model performance under conditions of incomplete drug information. When similarity networks were constructed with half of the target or ATC information removed (see Supplementary Note 5 for details), the model maintained comparable and stable performance relative to the complete information baseline (Supplementary Fig. 6; Supplementary Table 8). Furthermore, we trained models on incrementally larger datasets (10% intervals). Results showed that as the dataset scale increased, the model’s predictive performance exhibited a stable upward trend (Supplementary Fig. 7). Collectively, these results demonstrate that DSimSynergy is robust to incomplete drug information and exhibits strong data scalability, highlighting its potential to leverage large-scale data for enhanced predictive capability. Analysis of feature optimization capability in drug similarity networks In the initial stage of model construction, we hypothesized that drug similarity in the constructed networks correlates with synergistic effects. To validate this hypothesis, we analyzed correlations between drug pair weights in similarity networks and their mean synergy scores across cell lines in the train set. Results revealed significant positive correlations for both the similarity network based on biological process (PCC = 0.21, p = 1.3 × 10−6) and the similarity network based on ATC (PCC = 0.23, p < 2.2 × 10−16) (Fig. 4A, B).Fig. 4Drug feature optimization capability in drug similarity networks.A–E Correlation analysis between drug similarity and mean synergy score. Scatter plots showing correlations between mean synergy score of drug combinations across all experimental cell lines and different feature conditions. The color gradient corresponds to the local density distribution of drug pairs. A Correlation between the drug pair weights from similarity network based on biological process and mean synergy score; B Correlation between the drug pair weights from similarity network based on ATC and mean synergy score; C Correlation between cosine similarity of MACCS-encoded vectors of drug pairs and mean synergy score; D After model training, correlation between cosine similarity of drug pair features from the similarity network based on biological process and mean synergy score; E After model training, correlation between cosine similarity of drug pair features from the similarity networks based on ATC and mean synergy score. F–I GSEA enrichment analysis of drug feature fusion. GSEA enrichment analysis for high and low cosine similarity groups based on cosine similarity of concatenated drug pair features. Red curves represent high cosine similarity groups, and blue curves represent low cosine similarity groups. F Enrichment of high mean synergy scores in the high cosine similarity group for drug combinations in the dataset; G Enrichment of low mean synergy scores in the low cosine similarity group for drug combinations in the dataset; H Enrichment of high mean prediction synergy scores in the high cosine similarity group for all anticancer drug combinations in the network; I Enrichment of low mean prediction synergy scores in the low cosine similarity group for all anticancer drug combinations in the network. Drug feature optimization capability in drug similarity networks. A–E Correlation analysis between drug similarity and mean synergy score. Scatter plots showing correlations between mean synergy score of drug combinations across all experimental cell lines and different feature conditions. The color gradient corresponds to the local density distribution of drug pairs. A Correlation between the drug pair weights from similarity network based on biological process and mean synergy score; B Correlation between the drug pair weights from similarity network based on ATC and mean synergy score; C Correlation between cosine similarity of MACCS-encoded vectors of drug pairs and mean synergy score; D After model training, correlation between cosine similarity of drug pair features from the similarity network based on biological process and mean synergy score; E After model training, correlation between cosine similarity of drug pair features from the similarity networks based on ATC and mean synergy score. F–I GSEA enrichment analysis of drug feature fusion. GSEA enrichment analysis for high and low cosine similarity groups based on cosine similarity of concatenated drug pair features. Red curves represent high cosine similarity groups, and blue curves represent low cosine similarity groups. F Enrichment of high mean synergy scores in the high cosine similarity group for drug combinations in the dataset; G Enrichment of low mean synergy scores in the low cosine similarity group for drug combinations in the dataset; H Enrichment of high mean prediction synergy scores in the high cosine similarity group for all anticancer drug combinations in the network; I Enrichment of low mean prediction synergy scores in the low cosine similarity group for all anticancer drug combinations in the network. Next, we investigated whether latent drug features could be discovered through similarity network graph convolution. We compared drug features encoded by molecular access system (MACCS) molecular fingerprints against those optimized using the two drug similarity networks. To this end, we computed the cosine similarity between the features of different drug pairs and correlated it with their mean synergy score across cell lines. The results demonstrated a marked improvement: while the original MACCS fingerprint features showed only a weak linear correlation (PCC = 0.16, p < 3.4 × 10−16, Fig. 4C), the features optimized by our graph convolutional learning framework extracted from the synergy score prediction task achieved correlation coefficients exceeding 0.35 (p < 2.2 × 10−16, Fig. 4D, E), represents an improvement of over twofold. These results demonstrate that our model effectively captures latent inter-drug associations, enhancing correspondence with observed synergistic effects and improving prediction accuracy. To validate that the model has effectively learned to integrate and fuse heterogeneous feature representations from the three optimization modules (Fig. 1B), we performed enrichment analysis based on the concept of Gene Set Enrichment Analysis (GSEA)20. After concatenating the optimized features and calculating drug pair cosine similarities in the training set, we categorized the top and bottom 20% as high-similarity and low-similarity groups. GSEA revealed significant enrichment of high mean synergy scores in the high-similarity group (normalized enrichment score, NES = 1.13, permutation test p = 5.2 × 10−3; Fig. 4F), and low mean synergy scores were enriched in the low-similarity group (NES = −1.43, p = 1.2 × 10−10; Fig. 4G). Applying the same analysis to all anticancer drugs combination confirmed the same pattern: significant enrichment of high mean prediction synergy scores in the high-similarity group (NES = 1.53, permutation test p = 6.4 × 10−5; Fig. 4H) and low mean prediction synergy scores in the low-similarity group (NES = −1.62, permutation test p = 1 × 10−10; Fig. 4I). These results validate our multi-module optimization strategy for fused drug features that capture critical associations underlying drug synergy. Interpretation of cell line features for drug synergy DSimSynergy demonstrated excellent performance after training and learning. To intuitively compare the model’s feature learning capability, we performed a comparative analysis of feature representations before and after training on the binary synergy classification task. We concatenated the drug-drug-cell line features that were progressively optimized by the model across different training epochs. The order of concatenation for the three drug features is as follows: drug features extracted from the drug similarity network based on biological process, followed by drug features extracted from the drug similarity network based on ATC, and finally drug features extracted from the graph attention network based on drug molecular structure. Subsequently, Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction was applied to observe the distribution of synergistic and antagonistic samples in the drug-drug-cell triplets. The visualization of triplet distribution changes during checkpoints at different training epochs revealed that our model continuously learns to distinguish synergistic from antagonistic triplets. This progressive differentiation underscores the model’s excellent feature optimization capabilities (Fig. 5A).Fig. 5Visualization of feature learning dynamics during model training and SHAP based interpretability analysis of DSimSynergy.A Drug feature UMAP dimensionality reduction visualization of synergistic and antagonistic drug-drug-cell line triplets during model training. Red dots represent synergistic drug-drug-cell triplets; blue dots represent antagonistic drug-drug-cell line triplets. B, C Correlation analysis between gene expression levels and SHAP values. Each dot represents a drug–drug–cell line triplet, with dot color reflecting the expression level of the y-axis gene in the cell line of that triplet. The color gradient from blue to red indicates increasing gene expression levels. The x-axis represents the magnitude of SHAP values, reflecting the contribution of varying gene expression levels to the model’s prediction output. The y-axis displays the top 20 genes ranked by positive and negative correlation respectively. B Top 20 genes with positive correlation between gene expression values and SHAP values; C Top 20 genes with negative correlation between gene expression values and SHAP values. D, E KEGG pathway enrichment analysis of gene sets with significant correlations. The “Category” represents functional classification of enriched pathways, and “Count” indicates the number of genes enriched in each pathway. D The top 10 KEGG enriched pathways for genes with positive correlation (PCC > 0.5) between gene expression values and SHAP values; E The top 10 KEGG enriched pathways for genes with negative correlation (PCC < −0.5) between gene expression values and SHAP values. Visualization of feature learning dynamics during model training and SHAP based interpretability analysis of DSimSynergy. A Drug feature UMAP dimensionality reduction visualization of synergistic and antagonistic drug-drug-cell line triplets during model training. Red dots represent synergistic drug-drug-cell triplets; blue dots represent antagonistic drug-drug-cell line triplets. B, C Correlation analysis between gene expression levels and SHAP values. Each dot represents a drug–drug–cell line triplet, with dot color reflecting the expression level of the y-axis gene in the cell line of that triplet. The color gradient from blue to red indicates increasing gene expression levels. The x-axis represents the magnitude of SHAP values, reflecting the contribution of varying gene expression levels to the model’s prediction output. The y-axis displays the top 20 genes ranked by positive and negative correlation respectively. B Top 20 genes with positive correlation between gene expression values and SHAP values; C Top 20 genes with negative correlation between gene expression values and SHAP values. D, E KEGG pathway enrichment analysis of gene sets with significant correlations. The “Category” represents functional classification of enriched pathways, and “Count” indicates the number of genes enriched in each pathway. D The top 10 KEGG enriched pathways for genes with positive correlation (PCC > 0.5) between gene expression values and SHAP values; E The top 10 KEGG enriched pathways for genes with negative correlation (PCC < −0.5) between gene expression values and SHAP values. We observed that identical drug combinations exhibit varying synergistic effects across cell lines. Since gene expression values are used to characterize cell lines in our model, the differences in synergistic effects predicted across cell lines depend on differences in gene expression levels between cell lines. Therefore, we employed SHAP (Shapley Additive Explanations) methodology21 to interpret cell line features within our synergy score prediction model (see Supplementary Note 6 for details), aiming to identify key genes whose expression changes could determine the sensitivity of cells to combinatorial drug treatment. We sought to identify key genes whose expression changes determine cellular sensitivity to combinatorial drug treatment. Specifically, following the approach described by Janizek22, we computed the correlation between gene expression values and their corresponding SHAP values to identify genes whose expression levels systematically influence model predictions of synergistic effects, thereby revealing the underlying biological mechanisms. We displayed the top 20 genes with positive correlations and negative correlations between gene expression values and SHAP values (Fig. 5B, C; Supplementary Fig. 8A, B). Among highly positively correlated genes (Supplementary Fig. 8C, D), PPP2R3C belongs to the PP2A tumor suppressor complex and enhances dephosphorylation of oncogenic kinases, inhibiting cancer development23, MSH6 maintains genomic stability by repairing DNA replication errors and represents a classic tumor suppressor gene24. Among highly negatively correlated genes (Supplementary Fig. 8E, F), GHR overexpression activates the JAK2/STAT signaling to promote cancer growth25; EPHA3 overexpression promotes gastric cancer invasion26. These results demonstrate that our model effectively quantifies the contribution of gene expression to its synergy predictions and uncovers potential biological rationale underlying the model’s predictive behavior. Specifically, expression of genes associated with immune regulation and tumor suppression functions often exhibit positive correlations with SHAP values, thereby increasing the model’s predicted synergy score. Conversely, expression of genes that promote cancer development and progression more often display negative correlations with SHAP values, thereby decreasing the model’s predicted synergy score. This indicates that the model automatically extracts and integrates key molecular mechanisms underlying tumor progression and drug response as part of its predictive logic. To understand how the model predicts drug synergy at the system level, we performed KEGG pathway enrichment analysis using hypergeometric test-based ORA on genes with positive correlations greater than 0.5 and negative correlations less than −0.5. Top pathways enriched in positively correlated genes (Fig. 5D; Supplementary Table 9), cellular senescence pathway is closely associated with cell cycle arrest and tumor suppressive effects, capable of preventing unlimited cancer cell proliferation27; high expression or activation of the p53 signaling pathway directly leads to cancer cell apoptosis or senescence28. Activation of these pathways typically inhibits cancer development, suggesting they enhance synergistic effects during combination therapy. Top pathways enriched in negatively correlated genes (Fig. 5E; Supplementary Table 9), the cell cycle represents a core mechanism in various cancers, where highly expressed related factors drive unlimited proliferation29; the cAMP signaling pathway promotes tumorigenesis through abnormal hyperactivation and crosstalk with other pathways, contributing to invasion, metastasis, and drug resistance30. Overactivation of these pathways may promote cancer progression and resistance, counteracting combination therapy effects and diminishing synergy. Synergistic mechanism for specific drug combination Having established the key gene selection and pathway preferences identified by the model for predicting synergistic and antagonistic outcomes across different cell lines in the context of overall drug combinations, we further investigated specific drug combinations to gain a deeper understanding of the model’s underlying mechanisms. Specifically, we sought to understand why the same drug combination can produce different synergistic effects across different cell lines. Since the proportion of synergistic cell lines is very small for most drug combinations, directly analyzing such imbalanced combinations would compromise the reliability of SHAP based interpretation. To address the potential confounding effect of class imbalance between synergistic and antagonistic samples, drug combinations were screened and selected based on their synergy scores, which were predicted by the DSimSynergy model. We selected drug combinations for which the ratio of synergistic to antagonistic cell lines was approximately 1:1 across all cell lines. From these filtered combinations, we ranked them by prediction balance and selected the top 10 drug pairs (Fig. 6A). Among these, the top-ranked combination, Binimetinib–Brigatinib, was chosen as a representative case for subsequent in-depth analysis.Fig. 6Interpretability analysis of specific drug combinations.A Synergistic effect prediction of drug combinations across different cell lines. Display of the top 10 drug combinations with relatively balanced synergistic and antagonistic samples and their predicted synergy score distributions across cell lines. B, C Pathway enrichment analysis of genes associated with Binimetinib and Brigatinib combination synergistic effects. C The top 20 KEGG pathway enrichment for gene sets with positive correlation (PCC > 0.5) between cell line gene expression values and SHAP values; D The top 20 KEGG pathway enrichment for gene sets with negative correlation (PCC < −0.5) between cell line gene expression values and SHAP values. D, E Attention weight visualization of drug molecular substructures, where darker circles indicate higher attention scores (only atoms with scores greater than 1 are shown). D Attention scores of atoms in Binimetinib molecules; E Attention scores of atoms in Brigatinib molecules. Interpretability analysis of specific drug combinations. A Synergistic effect prediction of drug combinations across different cell lines. Display of the top 10 drug combinations with relatively balanced synergistic and antagonistic samples and their predicted synergy score distributions across cell lines. B, C Pathway enrichment analysis of genes associated with Binimetinib and Brigatinib combination synergistic effects. C The top 20 KEGG pathway enrichment for gene sets with positive correlation (PCC > 0.5) between cell line gene expression values and SHAP values; D The top 20 KEGG pathway enrichment for gene sets with negative correlation (PCC < −0.5) between cell line gene expression values and SHAP values. D, E Attention weight visualization of drug molecular substructures, where darker circles indicate higher attention scores (only atoms with scores greater than 1 are shown). D Attention scores of atoms in Binimetinib molecules; E Attention scores of atoms in Brigatinib molecules. To determine how cell line expression levels influence drug synergy for this specific pair. We correlated gene expression levels with SHAP values, identifying genes with significant positive or negative correlations (Supplementary Fig. 9). Since SHAP values reflect feature contributions to synergy predictions, these genes expression levels indirectly reveal their positive and negative associations with the synergistic effects of this drug combination. KEGG pathway enrichment analysis using hypergeometric test-based over-representation analysis (ORA) of genes with correlations >0.5 or <−0.5 revealed distinct patterns. After excluding overlapping pathways, most pathways enriched in positively correlated genes (Fig. 6B; Supplementary Table 10) are extensively linked to cancer inhibition. For example, p53 signaling, which induces DNA repair, cell cycle arrest, or apoptosis31; and mismatch repair, nucleotide excision repair, and DNA replication pathways, which maintain genomic stability and suppress tumorigenesis32–34. In contrast, pathways enriched in negatively correlated genes (Fig. 6C; Supplementary Table 10) relate predominantly to cancer development, such as Ras and Rap1 signaling35 and EGFR tyrosine kinase inhibitor resistance. Notably, Brigatinib itself is an EGFR inhibitor, and high expression of EGFR resistance pathways typically weakens EGFR inhibitor efficacy, which is highly consistent with its concentrated appearance in pathways enriched in negatively correlated genes. To further validate this finding, we performed the same interpretability analysis on an additional drug combination, Encorafenib–Binimetinib, which shares Binimetinib with the previously analyzed pair Binimetinib–Brigatinib. Consistent results were observed: genes positively correlated between expression values and SHAP values were enriched in tumor-suppressive pathways, notably the p53 signaling pathway28 and cellular senescence27 (Supplementary Fig. 10A), both of which recurrently appeared across both drug combinations; while genes negatively correlated between expression values and SHAP values were enriched in tumor-promoting pathways, such as the Ras signaling pathway35, PI3K-Akt signaling pathway36, and ErbB signaling pathway37 (Supplementary Fig. 10B). At the same time, the two compared drug combinations also exhibited uniquely enriched pathways, highlighting the specificity of the synergistic mechanisms underlying different drug combinations. Taken together, these results demonstrate that during prediction, the model can capture key gene expression patterns closely related to drug interactions and influence the final prediction results by assigning corresponding SHAP values. By further mapping these key genes to biological pathways, the model can identify both shared tumor-related pathways common across drug combinations and unique synergy-associated pathways specific to individual drug combinations, providing a comprehensive and biologically interpretable view of the mechanisms underlying drug synergy prediction. We further utilized attention modules to analyze key molecular substructures closely associated with synergistic effect in Binimetinib-Brigatinib drug combinations. Results showed that Binimetinib primarily focuses on aromatic amine and nitrogen-containing heterocyclic structures in drug synergy, which were assigned high attention weights (Fig. 6D). This aligns with literature demonstrating that these moieties specifically bind target proteins through hydrogen bonding and hydrophobic interactions, effectively inhibiting kinase activity38. For Brigatinib, attention analysis revealed that this compound relies more on phosphate groups and diarylamine pyrimidine core scaffolds for its action in the synergistic effect (Fig. 6E). Among these, phosphate groups serve as hydrogen bond acceptors, significantly enhancing the binding affinity between drug molecules and target proteins39. Notably, the identified key substructures closely correlate with the pharmacodynamic mechanisms of their respective drugs, validating the effectiveness of structure-activity relationship analysis in predicting drug synergy. Applications of DSimSynergy in single-cell and TCGA data Our results support DSimSynergy’s effectiveness in predicting drug synergy effects in cell lines. We assessed its ability to predict synergy in cancer patients (see Supplementary Note 7 for details). To explore this application, we evaluated predictions using scRNA-seq data. From the drug combination screening performed on head and neck cancer patients (HN120, HN137, HN148, HN159, HN160)40, we selected the three combinations that overlapped with our model’s scope for subsequent validation. We defined cells predicted to exhibit synergy as undergoing cell death. This allowed us to calculate the proportion of cell death for comparison with experimental observations. DsimSynergy’s predictions significantly correlated with observed responses across all 15 conditions (5 patients × 3 drug combinations) (Fig. 7A, PCC = 0.53, p = 0.04). Notably, performance comparable to CaDRReS-Sc, an established method for predicting single-cell drug combination responses40. We observed pronounced variation in predicted treatment responses across different tumor subclone clusters and patient samples (Fig. 7B). This inter- and intra-tumor heterogeneity, reflecting differing drug sensitivities among subclones, is consistent with findings reported in previous studies using the same single-cell dataset40. Our model can capture and identify such heterogeneity, demonstrating its potential applicability in the context of tumor subclone-specific drug response analysis.Fig. 7Applicability validation on single-cell and TCGA datasets and discovery of pathway complementarity mechanisms.A Single-cell data prediction validation. Scatter plot showing correlation between actual and predicted cell death proportions under different drug combination treatments in patient single-cell samples. B Cell subpopulation heterogeneity analysis. Heatmap of cell death proportions in different cell subpopulations from patient single-cell samples under different drug combinations. C TCGA lung cancer cohort synergy prediction. Top 10 anticancer drug combinations ranked by predicted synergistic effects in TCGA lung cancer patient samples. D Drug synergy network diagram. Different colors represent different indication tissue types of drugs; node size indicates the number of drug combinations producing synergistic effects centered on that drug; edge thickness represents the number of synergistic effects between drug pairs. E Pathway intersection analysis of synergistic drug targets. KEGG pathway enrichment intersection between Docetaxel targets and targets of drugs with high synergistic effect counts. Different colors represent different drugs. F Pathway complementary network diagram. Network showing pathways enriched by synergistic drugs with one-hop and two-hop interactions with the EGFR tyrosine kinase inhibitor resistance pathway. Distinct colors in sector regions denote different drugs. Line colors indicate network proximity: red represents one-hop neighboring nodes (distance = 1), black represents two-hop neighboring nodes (distance = 2). (pathway names in Supplementary Table 11). Applicability validation on single-cell and TCGA datasets and discovery of pathway complementarity mechanisms. A Single-cell data prediction validation. Scatter plot showing correlation between actual and predicted cell death proportions under different drug combination treatments in patient single-cell samples. B Cell subpopulation heterogeneity analysis. Heatmap of cell death proportions in different cell subpopulations from patient single-cell samples under different drug combinations. C TCGA lung cancer cohort synergy prediction. Top 10 anticancer drug combinations ranked by predicted synergistic effects in TCGA lung cancer patient samples. D Drug synergy network diagram. Different colors represent different indication tissue types of drugs; node size indicates the number of drug combinations producing synergistic effects centered on that drug; edge thickness represents the number of synergistic effects between drug pairs. E Pathway intersection analysis of synergistic drug targets. KEGG pathway enrichment intersection between Docetaxel targets and targets of drugs with high synergistic effect counts. Different colors represent different drugs. F Pathway complementary network diagram. Network showing pathways enriched by synergistic drugs with one-hop and two-hop interactions with the EGFR tyrosine kinase inhibitor resistance pathway. Distinct colors in sector regions denote different drugs. Line colors indicate network proximity: red represents one-hop neighboring nodes (distance = 1), black represents two-hop neighboring nodes (distance = 2). (pathway names in Supplementary Table 11). We next applied the model to TCGA lung cancer samples, as lung tissue exhibited optimal predictive performance (Supplementary Fig. 2D). For each drug-drug-sample triplet, combinations were classified as synergistic or antagonistic. Notably, among the top ten drug combinations ranked by synergistic effect count, most have been validated in clinical studies or experiments, demonstrating their efficacy in combination therapy (Fig. 7C; see Supplementary Note 8 for details). This confirm the clinical reliability of our model predictions, indicating that the drug synergy mechanisms revealed by the model are highly consistent with actual clinical observations. We extracted combinations demonstrating synergy in >200 samples and constructed a drug-drug network with nodes representing drugs annotated with DrugBank indication data and edges weighted by synergy count (Fig. 7D). Analysis revealed that Docetaxel not only appeared most frequently among the top ten drug combinations but also exhibited the highest synergistic count with other drugs in the network (represented by node size), establishing it as a pivotal hub drug for lung cancer combination therapy. Interestingly, while most network drugs target lung tissue, Docetaxel (a lung cancer drug) exhibits highest synergy with agents primarily targeting non-lung diseases. Conversely, Dasatinib, which ranks second in synergistic connectivity but is not a lung cancer drug, interacts most synergistically with lung disease-indicated drugs. Based on this phenomenon, we hypothesize that drug combinations may produce synergistic effects through complementary mechanisms, enhancing efficacy while circumventing resistance development. KEGG pathway enrichment was performed on the targets of Docetaxel and its high-synergy drugs. The analysis revealed that seven of these drugs share significantly enriched pathways with Docetaxel (Fig. 7E). Notably, EGFR tyrosine kinase inhibitor resistance pathway was significantly enriched in nearly all drugs analyzed and emerged as the central hub in the pathway intersection, highlighting it as a potential common mechanism. Importantly, this pathway is a well-established key regulator in lung cancer pathogenesis41. To elucidate molecular network mechanisms underlying drug synergy, we constructed one-hop and two-hop pathway interaction networks for the EGFR tyrosine kinase inhibitor resistance pathway (hsa01521), each pathway is labeled with its corresponding drug (Fig. 7F and pathway names in Supplementary Table 11). Network analysis revealed an intriguing phenomenon: pathways with one-hop and two-hop interactions with this pathway were almost entirely significantly enriched by other synergistic drugs, while the hub drug Docetaxel itself did not enrich these related pathways. Importantly, drug appearance frequency in one-hop pathways positively correlated with the number of synergistic effects produced with Docetaxel (PCC = 0.86), suggesting that pathway network connectivity density directly influences synergistic strength. To validate the universality of this pathway complementary mechanism, we analyzed Dasatinib, which ranked second in synergistic effect count. Results showed that drugs producing high synergy with Dasatinib predominantly co-enriched the PI3K-Akt signaling pathway (hsa04151), a key pathway (Supplementary Fig. 11A). Pathway network analysis revealed that Dasatinib and its synergistic drugs act on different one-hop and two-hop pathways but ultimately co-influence PI3K-Akt signaling through pathway cascades (Supplementary Fig. 11B and pathway names in Supplementary Table 12). These findings indicate that drug combinations predicted by our model to be highly synergistic typically exhibit pathway complementarity. Specifically, these drugs form a “pathway complementarity network” by targeting distinct but interconnected signaling pathways. This network allows for the coordinated regulation of key disease-related pathways, thereby enhancing therapeutic efficacy. This mechanism is consistent with existing research findings, and related studies have demonstrated that drug combinations can enhance therapeutic efficacy by acting on different but interconnected pathways3,42. Discussion This study presents DSimSynergy, a deep learning model that achieves high-precision prediction of drug synergistic effects by integrating multi-level drug and cell line feature information. Its core innovation includes: (1) adding two additional representations (ATC codes and biological process annotations), (2) constructing similarity networks from these representations, (3) processing these networks using GCNs. Through GCNs for learning drug similarity network, GATs for extracting drug molecular structure features, and fully connected neural networks for optimizing cell line gene expression features, this multi-sub-module collaborative optimization strategy enables the model to capture latent feature associations between drugs and cell lines from different dimensions. DSimSynergy outperforms state-of-the-art deep learning and conventional machine learning methods on both internal and external test sets. Ablation experiments confirm the important contributions of each sub-module to model performance improvement, particularly the introduction of drug similarity networks, which significantly enhances predictive capability. Stratified data validation demonstrates that although performance decreases in unknown drug prediction scenarios, the model maintains relative advantages, showing good generalization ability. Drug synergistic effect prediction is also closely related to drug features, and understanding the learning mechanisms of drug features is equally important. Drug similarity network hypothesis validation demonstrates significant positive correlations between drug similarity and synergistic effects. The correlation between drug feature similarity optimized through graph convolution and synergistic effects improved more than twofold compared to original features, further validating the effectiveness of our similarity networks. Although deep learning models demonstrate excellent performance in drug synergy prediction, their “black box” nature often raises concerns. To overcome this limitation, various interpretable artificial intelligence methods have been widely applied in recent years, such as backpropagation-based methods, input perturbation methods, and SHAP value21 calculation. This study reveals key genes and biological pathways associated with drug synergistic effects through SHAP analysis. Genes related to immune regulation and tumor suppression functions tend to promote synergistic effect prediction, while oncogenes tend to weaken synergistic effects. KEGG pathway enrichment analysis further confirms this biological pattern, providing systematic molecular explanations for drug synergy mechanisms. Additionally, we gained deeper understanding of model prediction mechanisms through specific drug combination case studies. Clinical translation application validation demonstrates that DSimSynergy possesses good clinical application potential. Single-cell data prediction results show significant correlation with actual observed responses and reveal heterogeneity in treatment responses among different patient cell subpopulations. TCGA lung cancer data analysis identifies a drug synergy network centered on Docetaxel and elucidates the molecular mechanisms of “pathway complementary networks” through pathway interaction network analysis, providing theoretical basis for overcoming acquired drug resistance. Therefore, it may effectively circumvent acquired drug resistance problems common in single-pathway targeted therapy through multi-pathway synergistic intervention, providing a novel guidance for precision combination therapy. In summary, this study presents a high-performance and interpretable model for predicting drug synergy. It effectively addresses the critical bottlenecks of time and cost associated with traditional screening methods, thereby providing a powerful computational tool for advancing personalized cancer combination therapy. The research results not only achieve methodological innovation but also provide new perspectives for understanding the biological mechanisms of drug synergistic effects, potentially accelerating the clinical application process of personalized drug combination selection in precision medicine. Future research can be further extended to more cancer types and drug combinations, incorporating additional omics data to improve prediction accuracy and clinical applicability. Materials and methods Drug synergy data Drug synergy data were obtained from DrugCombDB (2019)43, a comprehensive database integrating drug combination information from multiple sources. DrugCombDB contains 6,891,566 combination experiments encompassing quantitative dose-response data for 2887 drugs across 124 human cancer cell lines. The database calculates synergy scores (Loewe scores) for each triplet using the Loewe additivity model by fitting single-drug dose-response curves. We downloaded triplet data with pre-calculated Loewe scores and averaged scores for replicate experiments of identical drug combinations, yielding 408,350 unique triplets, each consisting of two drugs and one cancer cell line. We further screened the synergy data by retaining only drugs with both target and ATC code information in DrugBank. Prior to cell line filtering, the dataset comprised 178,425 drug-drug-cell triplets, after matching cell lines present in the DepMap Portal database (24Q4 version) with available expression profiles, this was reduced to 83,162 triplets, involving 521 drugs and 83 cell lines. Additionally, we employed two independent datasets for model test: the DrugComb44 and the AZ-DREAM45. All datasets underwent data processing procedures consistent with the DrugCombDB dataset. For DrugComb, which shares experimental data with DrugCombDB, we removed duplicate triplet-Loewe pairs, yielding 66,273 triplets (307 drugs, 108 cell lines). The AZ-DREAM dataset contained 611 triplets (19 drugs, 52 cell lines). For the classification task, drug combinations were categorized based on their Loewe synergy scores. Following the study of Wang et al.11, combinations with scores greater than 10 were classified as synergistic, while those with scores less than 0 were classified as antagonistic. This threshold selection accounts for the uncertainty introduced by noise near zero. This classification strategy was guided by two primary considerations: first, it enables a more intuitive presentation of whether a given triplet exhibits synergistic or antagonistic effects; second, it facilitates the prioritization of drug combinations with stronger and more robust synergistic effects, thereby supporting their advancement to subsequent stages of clinical research. Drug feature data In this study, the drug-related data, including Simplified Molecular Input Line Entry System (SMILES) structures, target information, and Anatomical Therapeutic Chemical (ATC) codes, were sourced from the DrugBank database (v5.0)46. We screened and retained approved drugs that simultaneously contained all three types of information, totaling 1604 drugs for subsequent similarity network construction and analytical studies. RDKit20,47 was employed to convert drug SMILES sequences into corresponding molecular graph representations, where nodes represent atoms in the drug molecular structure and edges represent chemical bonds connecting these atoms. Additionally, SMILES sequences were processed through MACCS fingerprint featurization, with each drug represented as a unique one-hot vector serving as embeddings for drug similarity network graphs. Cell line gene expression data Gene expression data for cell lines were obtained from the DepMap Portal database (24Q4 version). Protein-coding gene expression data were standardized using Transcripts Per Million (TPM) and subsequently transformed using log2(TPM + 1). Since excessive numbers of expressed genes in cell lines would significantly reduce model training efficiency and to alleviate dimensional imbalance issues between drug features and cell line feature vectors, we adopt the 978 landmark genes measured by the L1000 platform from the LINCS project48, which are believed to capture approximately 80% of the information in the entire transcriptome. By taking the intersection of these landmark genes with cell line expressed genes, we ultimately obtained cell line expression data containing 949 genes, which served as input features for subsequent models. DSimSynergy architecture We propose DSimSynergy, a novel deep learning model for predicting drug synergistic effects (Fig. 1). The workflow of DSimSynergy can be summarized into four modules: (1) drug similarity networks construction module, (2) drug feature representation module, (3) cell line feature representation module, and (4) drug synergy prediction module. The detailed methodology of DSimSynergy is described below. Construction of drug-drug similarity networks The synergistic effects of drug combinations are associated with the similarity of drugs in biological processes and clinical characteristics profiles15. Therefore, we constructed two drug similarity networks to characterize two types of drug similarity: one based on BP and the other on the ATC. These two networks respectively characterize the relationships between drugs from the perspectives of functional and clinical level, providing core modules for systematically understanding and predicting drug synergy. We first constructed a drug-drug functional similarity network based on shared BPs between drugs49,50. To build this network, we obtained target gene sets of drugs from the DrugBank database and acquired gene sets of biological processes (GO-BP) from the Gene Ontology database (GO)51. The association strength between a drug and a GO-BP is quantified by calculating the Jaccard similarity coefficient between drug target genes (DT) and the GO-BP gene set (G), which serves as the weight between the drug and the GO-BP. The weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{DT\cdot \; G}}$$\end{document}WDT· G is calculated 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}$${W}_{{DT\cdot \; G}}=\frac{{|DT}\bigcap G|}{{|DT}\bigcup G|}$$\end{document}WDT· G=|DT⋂G||DT⋃G|The similarity score (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{G}$$\end{document}SG) between drugi and drugj, is calculated as follows:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{G}\left(i,j\right)=\mathop{\sum }\limits_{k=1}^{n}{W}_{{{DT}}_{i}{G}_{k}}{W}_{D{T}_{j}{G}_{k}}$$\end{document}SGi,j=∑k=1nWDTiGkWDTjGkwhere DTi and DTj represents target genes of drugi and drugj, n represent the total number of shared GO-BP between two drugs, Gk denotes the k-th GO-BP. We next constructed a drug-drug clinical similarity network based on ATC codes. The ATC is a five-level hierarchical classification system, whose code comprises seven characters organized into five hierarchical levels: the first level (character 1) represents the anatomical group; the second level (characters 2–3) represents the therapeutic group; the third level (character 4) represents the pharmacological group; the fourth level (character 5) represents to the chemical group; and the fifth level (characters 6–7) represents the chemical substance group. Evaluating drug similarity based on the ATC codes enables a systematic characterization of pharmacological relationships among drugs relevant to clinical decision-making. Specifically, we used the hierarchical structure of ATC codes as the core quantitative framework. Given that a drug may maintain multiple ATC codes, therefore we employed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{ATC}}_{i}^{k}$$\end{document}ATCik to denote the set of characters at the k-th hierarchical level across all ATC codes of drugi and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{ATC}}_{j}^{k}$$\end{document}ATCjk to denote the set of characters at the k-th hierarchical level across all ATC codes of drugj. The calculation of the association strength \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{k}\left(i,j\right)$$\end{document}Ski,j between drugi and drugj at the k-th level is as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{k}\left(i,j\right)=\frac{|{{ATC}}_{i}^{k}\cap {{ATC}}_{j}^{k}|}{|{{ATC}}_{i}^{k}\cup {{ATC}}_{j}^{k}|}$$\end{document}Ski,j=|ATCik∩ATCjk||ATCik∪ATCjk|Finally, the overall clinical similarity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{{atc}}$$\end{document}Satc) between drugi and drugj, is defined as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{{atc}}\left(i,j\right)=\frac{{\sum }_{k=1}^{5}{S}_{k}(i,j)}{5}$$\end{document}Satci,j=∑k=15Sk(i,j)5 Drug feature based on drug similarity network with GCN The drug similarity network contains N nodes representing the number of screened drugs, with node features consisting of one-hot vectors of length F obtained through MACCS fingerprint featurization of SMILES, resulting in a node feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X\in {{\mathbb{R}}}^{N\times F}$$\end{document}X∈RN×F and a node adjacency matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A\in {{\mathbb{R}}}^{N\times N}$$\end{document}A∈RN×N. These two matrices serve as inputs to the multi-layer GCN (Fig. 1B), with the iterative propagation computation process defined as follows:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}^{(l+1)}=\sigma ({\widetilde{D}}^{-\frac{1}{2}}\widetilde{A}{\widetilde{D}}^{-\frac{1}{2}}{H}^{\left(l\right)}{W}^{\left(l\right)})$$\end{document}H(l+1)=σ(D~−12A~D~−12HlWl)where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{A}=A+I$$\end{document}A~=A+I(I is the identity matrix) represents the adjacency matrix with added self-connections, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{D}$$\end{document}D~ is the degree matrix of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widetilde{A}$$\end{document}A~, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}^{(l)}$$\end{document}H(l) represents the node feature matrix at the l-th layer, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}^{(l)}$$\end{document}W(l) represents the learnable weight matrix, σσ is the activation function LeakyReLU, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}^{(l+1)}$$\end{document}H(l+1) is the activation matrix at the (l + 1)-th layer, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}^{(0)}=X$$\end{document}H(0)=X. After multi-layer iterative computation, we ultimately obtain the drug feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{{drug}}\in {{\mathbb{R}}}^{N\times {F}^{{\prime} }}$$\end{document}Hdrug∈RN×F′ based on the drug similarity network, where N represents the number of drugs and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}^{{\prime} }$$\end{document}F′ represents the output feature vector length. Drug feature based on drug structure graph with GAT We employed the RDKit47 package in Python to convert drug SMILES into molecular graphs, where nodes represent atoms and edges represent chemical bonds. Specifically, a given drug molecular graph is defined as G=(V,E)G=(V,E), where V is a set composed of N nodes and E is a set of edges represented by the node adjacency matrix. In the molecular graph, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}\in V$$\end{document}xi∈V represents the i-th atom and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${e}_{{ij}}\in E$$\end{document}eij∈E represents the chemical bond between the i-th and j-th atoms. We utilized DeepChem52, a molecular featurizer specifically designed for GNN, to generate fixed-length one-hot vectors as initial features for each node, containing information such as atomic symbol, atomic mass, number of neighboring atoms, number of neighboring hydrogen atoms, and whether the atom is in an aromatic structure. GAT was employed to learn from drug molecular graphs, enabling information propagation between graph nodes and their adjacent nodes (Fig. 1B). GAT proposes an architecture based on multi-head attention mechanisms, aggregating information from neighboring nodes through self-attention mechanisms to learn latent features of nodes in the graph. GAT is constructed from multiple graph attention layers. The node output features are calculated 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}$${h}_{i}^{{\prime} }={\mathrm{||}}_{m=1}^{M}\left({\alpha }_{i,j}^{m}W{h}_{i}+\mathop{\sum }\limits_{j\in N(i)}{\alpha }_{i,j}^{m}W{h}_{i}\right)$$\end{document}hi′=||m=1Mαi,jmWhi+∑j∈N(i)αi,jmWhiwhere ∥∥ denotes the concatenation of outputs from multiple attention heads, M is the number of attention heads, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W\in {R}^{{F}^{{\prime} }\times F}$$\end{document}W∈RF′×F is the weight matrix, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{i}\in {{\mathbb{R}}}^{F}$$\end{document}hi∈RF represents the feature vector of node i, and j∈ N(i)j∈N(i) represents the node j adjacent to node i. The attention coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\alpha }_{i,j}$$\end{document}αi,j between input node i and its first-order neighbor node j in the molecular graph is calculated as follows:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${e}_{i,j}={LeakyReLU}\left({a}^{T}\left[W{h}_{i}{||W}{h}_{j}\right]\right)$$\end{document}ei,j=LeakyReLUaTWhi||Whj8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\alpha }_{i,j}={softmax}\left({e}_{i,j}\right)=\frac{\exp \left({e}_{i,j}\right)}{{\sum }_{k\in N\left(i\right)}\exp \left({e}_{i,k}\right)}$$\end{document}αi,j=softmaxei,j=expei,j∑k∈Niexpei,kwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a\in {{\mathbb{R}}}^{2{F}^{{\prime} }}$$\end{document}a∈R2F′ is the learnable attention vector, LeakyReLU is a nonlinear activation function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${e}_{i,j}$$\end{document}ei,j represents the importance score transmitted from node j to node i. Softmax normalization is applied to all neighboring nodes of node i to facilitate computation and comparison. After graph attention learning, the drug molecular graph yields a feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{{mol}}\in {{\mathbb{R}}}^{N\times F}$$\end{document}Hmol∈RN×F. N represents the number of nodes in the drug molecular graph, and F represents the length of the node feature vector, which remains consistent across different drugs. Global Maximum Pooling (GMP) is applied to all node features of the drug molecular graph to generate a graph-level representation characterizing the features of the entire drug:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{{graph}}=\mathop{\max }\limits_{i\in N}{h}_{i}^{{\prime} }$$\end{document}hgraph=maxi∈Nhi′where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{i}^{{\prime} }\in {{\mathbb{R}}}^{{F}^{{\prime} }}$$\end{document}hi′∈RF′ is the feature vector of node i, and max represents the element-wise maximum operation across feature dimensions. The same computational process is applied to all screened drugs, ultimately yielding a drug feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{{GAT}}\in {{\mathbb{R}}}^{D\times {F}^{{\prime} }}$$\end{document}HGAT∈RD×F′, where D represents the number of drug and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}^{{\prime} }$$\end{document}F′ represents the output feature vector length. Cell line features based on FCNN We utilized screened gene expression values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x\in {{\mathbb{R}}}^{d}$$\end{document}x∈Rd as initial feature vectors for cell lines. Cell line feature vectors undergo learning through a single-layer fully connected neural network to obtain new cell line features h (Fig. 1C). The computational process of the single-layer fully connected neural network is 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}$$h=\sigma \left({W}^{\left(1\right)}x+{b}^{\left(1\right)}\right)$$\end{document}h=σW1x+b1where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}^{(1)}$$\end{document}W(1) represents the weight matrix, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}^{(1)}$$\end{document}b(1) represents the 1-th bias vector, x represents the initial feature vector of the cell line, and σσ represents the activation function. Single-layer FNN operations are applied to all cell line initial feature vectors, ultimately yielding a cell line feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{{cell}}\in {{\mathbb{R}}}^{N\times {d}^{{\prime} }}$$\end{document}Hcell∈RN×d′, where N represents the number of cell lines and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}^{{\prime} }$$\end{document}d′ represents the length of output features after single-layer fully connected neural network learning. Prediction of drug synergy After processing initial features through the aforementioned methods, we obtained three types of drug features (Fig. 1B) and cell line features (Fig. 1C). These features are fused and concatenated through drug-drug-cell triplet combinations and input into the prediction module. For the binary synergy classification task, the final layer of the model outputs a probability score via a sigmoid activation function to obtain probabilities for predicting drug combination synergy, and the model was optimized using binary cross-entropy loss. For the synergy score prediction task, this module consists of two fully connected layers with LeakyReLU activation functions applied between layers to directly predict continuous Loewe synergy scores, optimized using mean squared error loss. Model training and evaluation The DSimSynergy model was implemented in Python 3.12 and PyTorch 2.8.0, and trained on a platform equipped with a 5090 GPU. On the DrugCombDB dataset, to eliminate sensitivity to drug ordering in drug–drug feature fusion, each drug triplet (Drug A, Drug B, Cell line) was augmented by including its reverse ordering (Drug B, Drug A, Cell line), where the Loewe synergy score remains unchanged upon swapping the drug order. Both orderings were included in the dataset during training, such that the model learns symmetric representations regardless of input order. Data were partitioned into training and test sets at a 9:1 ratio, with the training set further divided into training and validation subsets at a 4:1 ratio for 5-fold cross-validation. The mean values of performance metrics across five training runs were used for result presentation, and the best-performing model was evaluated on two independent external test sets, DrugComb and AZ-DREAM, to assess its generalization capability. To evaluate model prediction performance, we different metrics appropriate for each of the two prediction tasks. For the classification task, we employed the Area Under the ROC Curve (AUROC), Area Under the Precision-Recall Curve (AUPR), accuracy (ACC), and F1-score (F1). For computing ACC and F1, we applied a default decision threshold of 0.5 to the predicted probabilities from the binary classification model. For the synergy score regression task, we used the Pearson Correlation Coefficient (PCC) and Root Mean Square Error (RMSE). Supplementary information Supplementary Information Supplementary Data Supplementary Information Supplementary Data