SynCell: Contextualized Drug Synergy Prediction

Abstract

Drug synergy is profoundly influenced by cellular context, as variations in protein interaction landscapes and pathway activities across cell types reshape how drugs act in combination. Most existing models overlook this heterogeneity, relying on static or bulk-level protein-protein interaction (PPI) networks that ignore cell-specific molecular wiring. The availability of large-scale transcriptomic data now enables the reconstruction of cell-line-resolved interactomes, offering a new foundation for contextualized drug synergy modeling. Here we present SynCell, a Contextualized Drug Synergy framework that integrates drug-protein, protein-protein, and protein-cell line relations within a unified graph architecture. SynCell leverages cell-line-specific PPI networks to embed the molecular context in which drugs act, and employs graph convolutional learning to model how pharmacological effects propagate through cell-specific signaling networks. This formulation treats synergy prediction as a cell-line-contextualized drug-drug interaction problem. Across the large-scale DrugCombDB benchmark, SynCell consistently outperforms state-of-the-art baselines - including DeepSynergy, HypergraphSynergy, HERMES, BAITSAO, DTF, and NHP - particularly in predicting synergies involving unseen drugs or novel cell lines. When benchmarked against these seven methods, SynCell demonstrates substantial gains in generalization and biological interpretability, confirming that contextualizing PPIs with cell-line resolution is indispensable for accurate synergy prediction.

Figures (5)

• Figure 1: SynCell is a contextualized graph framework for drug synergy prediction, leveraging cell-line-specific PPI networks to capture molecular heterogeneity.a, Cellular context shapes drug response. Drug synergy varies across cell lines due to differences in protein activation contexts, where the same drug combination may exhibit synergistic effects in one cell line but not in another. b, Multimodal knowledge graph construction. SynCell integrates data from DrugCombDB, DepMap, STRING, and PrimeKG to build a unified heterogeneous graph. DrugCombDB provides synergy measurements; DepMap summarizes cell-line-specific protein expression profiles; STRING presents global protein–protein interactions; and PrimeKG illustrates biological relationships including drug–target and disease–protein associations. c, Computational framework overview. SynCell operates through three phases: (1) context-aware graph construction derives cell-line-specific PPI subnetworks from the global PPI based on activated genes; (2) multimodal embeddings are generated via SVD for expression data and transformer encoders (ChemBERTa, PubMedBERT) for drug and disease features; (3) a heterogeneous graph neural network integrates hyperedge propagation and adaptive contextual modulation to predict synergy. The model jointly optimizes synergy prediction and organ classification to enhance generalization across unseen biological environments.
• Figure 2: SynCell achieves robust zero-shot generalization to unseen drugs and drug combinations across progressively challenging evaluation settings.a, DrugComb split: generalization to unseen drug combinations while all individual drugs are observed during training. SynCell achieves the highest AUROC (0.795), AUPRC (0.603) and F1 (0.587), outperforming DeepSynergy, HERMES and other baselines. b, DrugSingle split: semi-cold-start scenario where one drug in each test pair is entirely unseen. SynCell maintains superior ranking consistency (AUROC: 0.646) and classification balance (F1: 0.406) despite increased task difficulty (AUPRC: 0.395). c, DrugDouble split: fully cold-start scenario where both drugs in test pairs are absent from training. Although performance drops across all methods, SynCell remains robust (AUROC: 0.602, AUPRC: 0.369, F1: 0.336), demonstrating its ability to infer transferable biological interaction mechanisms beyond memorized drug identities. For all panels, left: ROC curves (FPR vs. TPR); middle: PR curves (Recall vs. Precision); right: bar charts with overlaid scatter points showing mean $\pm$ s.d. across five independent runs. Statistical significance ($p<0.05$) is denoted by asterisks compared to SynCell.
• Figure 3: Cell-line-specific molecular contexts drive generalization more effectively than coarse organ-level taxonomy.a, Data summaries: distribution of drugs per cell line (left), number of cell line types per organ category (middle), and number of unique drugs per organ (right), illustrating heterogeneity in biological coverage across tissue types. b, Random cell-line split performance: SynCell achieves the highest AUROC (0.751), AUPRC (0.516) and F1 (0.509), indicating effective modeling of fine-grained cellular variation when transferred across diverse unseen backgrounds. c, Organ-level performance comparison across eleven tissue categories: radar plots show AUROC (left), AUPRC (middle), and F1 (right). SynCell consistently outperforms all baselines, with notable advantages in epithelial tumor types (e.g., Ovary, Lung, Bowel, Skin), while maintaining robustness in biologically challenging organs (e.g., CNS, Lymphoid, Myeloid), where all methods exhibit performance degradation. Error bars represent 95% confidence intervals across five random splits ($n=5$). Statistical significance ($p<0.001$) is denoted by asterisks.
• Figure 4: SynCell scales robustly to three-drug combinations, capturing context-dependent synergy patterns in structured discrete dose grids.a, Model performance on triple-drug prediction. Left: AUROC comparison showing SynCell achieves the best performance (0.768), outperforming DeepSynergy (0.716), BAITSAO (0.695), NHP (0.692), HypergraphSynergy (0.685), and DTF (0.671), with HERMES substantially lower (0.500). Middle: AUPRC comparison where SynCell again leads (0.803), followed by DeepSynergy (0.788) and BAITSAO (0.756). Right: Precision@TopK curves across varying K, where SynCell consistently maintains the highest precision, demonstrating superior ranking quality in high-confidence regions. b, Experimental validation workflow and representative case studies. Left: drug library and experimental design for three-drug combination screening, with synergy quantified as Gain = (Obs $-$ Exp)/Exp under the Bliss independence model. Right: representative survival gain and apoptosis gain curves for WM-115 and A-375 cell lines across increasing concentrations of the third drug. Predicted synergistic combinations (red) exhibit stronger-than-expected viability reduction and apoptosis induction, whereas predicted non-synergistic combinations (blue) closely follow or underperform the Bliss expectation baseline. These results demonstrate that SynCell captures context-dependent higher-order drug interactions and produces biologically consistent ranking signals aligned with experimental outcomes.
• Figure 5: SynCell uncovers biologically meaningful drug interaction patterns through mechanism-of-action–level analysis and pathway enrichment.a, Number of synergistic drug pairs across organ types, separated into previously reported (blue) and newly predicted (red) interactions. SynCell substantially expands the landscape of candidate synergistic combinations beyond known data, particularly in Ovary, Lung, and Breast tissues. b, Distribution of drugs across mechanism-of-action (MoA) categories, illustrating the diversity and imbalance of pharmacological classes in the dataset. c, Organ-specific MoA combination analysis. For each organ, the top-ranked MoA combinations are shown based on normalized synergy density (number of positive pairs divided by total possible pairs). Blue points denote reported synergies, while red points indicate novel predictions. SynCell prioritizes biologically plausible and context-dependent MoA interactions, with distinct patterns observed across organs. d, Pathway enrichment analysis based on attention-prioritized proteins. Heatmap shows significantly enriched pathways across organ types, with color indicating FDR-adjusted significance. Enriched pathways reveal context-specific biological mechanisms underlying predicted drug synergy.