Enhancing Drug Synergy Prediction through Multi-Omics Integration and Attention-Based Graph Neural Networks
Abstract Predicting synergistic drug combinations is an important task in precision oncology, but accurately modeling the heterogeneous molecular characteristics of cancer cell lines and the interactions between drugs remains challenging. In this study, we propose GraOmicSynergy, a graph-based deep learning framework that integrates multi-omics representations of cell lines with molecular graph representations of drug pairs. Cell-line omics data, including gene expression, mutation and copy number aberration, and methylation, are encoded using modality-specific convolutional neural networks, while individual drugs are represented as molecular graphs. An attention-based drug-pair mechanism is employed to model the context-dependent contributions of individual drugs within a combination, and the resulting representations are integrated to predict drug synergy scores. We evaluate GraOmicSynergy using five-fold cross-validation under Mixed-set, Blind-Cellline, and Blind-DrugPair scenarios, with comparisons against DeepSynergy and DeepDDS. The results show that GraOmicSynergy-GE outperforms both baseline methods in the Mixed-set and Blind-Cellline scenarios, achieving lower RMSE and higher Pearson correlation, with statistically supported differences across the evaluated metrics, whereas prediction of completely unseen drug pairs remains substantially more challenging. For the complementary classification task, GraOmicSynergy achieves the highest values for four of six metrics in both the Mixed-set and Blind-Cellline scenarios. Multi-omics analysis indicates that the predictive benefit of additional modalities is configuration- and scenario-dependent, with more evident gains in the Mixed-set setting. An ablation study further shows that attention-based drug-pair fusion provides consistently higher Pearson correlation than concatenation and mean pooling across the evaluated omics configurations. Additional classification and tissue-specific analyses were conducted to characterize model performance across different settings. Overall, the findings support GraOmicSynergy as a context-aware framework for drug synergy prediction, while highlighting the need for independent validation to establish broader generalizability.
Authors
- Duc‐Hau Le (ORCID: https://orcid.org/0000-0002-4951-5916)
- Hoa Dinh Vu
- Giang T.T. Nguyen
Institutions
- Trường Đại học Khoa học và Công nghệ Hà Nội (VN)
- Hanoi University of Science and Technology (VN)
Publication Details
- Journal
- Biology Methods and Protocols
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1093/biomethods/bpag063
- Primary Topic
- Computational Drug Discovery Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00