A scalable multimodal graph neural network for drug combination response prediction

Area of Science:

• Computational biology
• Cheminformatics
• Machine learning

Background:

• Cancer therapy resistance, particularly to targeted drugs like osimertinib, is a significant clinical challenge.
• Predicting synergistic drug combinations is a key strategy to overcome resistance, but molecular interactions are complex and nonlinear.
• Existing computational models struggle with the heterogeneity of molecular data for accurate synergy prediction.

Purpose of the Study:

• To introduce the Multimodal Molecular Drug Graph Neural Network (MMDGNN), a novel framework for enhanced drug synergy prediction.
• To leverage advanced molecular representation learning by integrating multimodal molecular features.
• To develop a scalable and robust computational tool for precision oncology.

Main Methods:

• MMDGNN integrates molecular fingerprints and SMILES representations using an adaptive graph neural network architecture.
• The model employs heterophily-aware learning to capture substructural dependencies and potential metabolic liabilities.
• Multimodal feature fusion and distributed data parallel training were utilized for improved expressiveness and scalability.

Main Results:

• MMDGNN achieved a Mean Squared Error (MSE) of 16.18, outperforming the MGAE-DC model (17.36).
• The framework obtained a Pearson correlation coefficient of 0.85, surpassing MGAE-DC's 0.84.
• These results represent a 6.8% MSE reduction and a 1.2% correlation improvement, indicating enhanced predictive accuracy and robustness.

Conclusions:

• MMDGNN demonstrates superior performance in predicting drug synergy through multimodal molecular representation learning.
• The scalable and adaptive architecture facilitates integration of diverse molecular data and efficient handling of large datasets.
• MMDGNN offers a promising computational foundation for precision oncology and can be extended to other biomedical applications.