A scalable multimodal graph neural network for drug combination response prediction.

Background Resistance to targeted cancer therapies, such as osimertinib in EGFR-mutant lung cancer, remains a major obstacle to effective treatment. Predicting synergistic drug combinations offers a promising strategy to overcome such resistance; however, the nonlinear and heterogeneous nature of molecular interactions makes this prediction highly challenging. This study introduces the Multimodal Molecular Drug Graph Neural Network (MMDGNN), a unified framework designed to enhance drug synergy prediction through advanced molecular representation learning. Methods MMDGNN integrates molecular fingerprints and SMILES representations within an adaptive graph neural network architecture capable of heterophily-aware modeling. The model captures substructural dependencies that reflect potential metabolic liabilities and compound synergies. Unlike prior models such as MGAE-DC, MMDGNN fuses multimodal molecular features to improve expressiveness. The framework supports distributed data parallel training for large-scale deployment and was empirically evaluated on four benchmark datasets. Results MMDGNN achieved a mean squared error (MSE) of 16.18, outperforming MGAE-DC (17.36), and obtained a Pearson correlation coefficient of 0.85, compared to 0.84 for MGAE-DC. These correspond to performance gains of 6.8% in MSE reduction and 1.2% in correlation improvement, confirming enhanced predictive accuracy and robustness. Conclusions MMDGNN demonstrates superior capability in learning multimodal molecular representations for drug synergy prediction. Its scalable, adaptive architecture enables integration of diverse molecular modalities and efficient handling of large datasets. While performance may vary in cancer types with limited data and potential off-target effects warrant further validation, MMDGNN provides a promising computational foundation for precision oncology. The framework can be extended to broader biomedical applications requiring multimodal molecular inference.