MoGraphDRP: Multi-omics and graph fusion with bilinear attention for predicting drug sensitivity

Area of Science:

• Computational biology
• Genomics and bioinformatics
• Drug discovery and development

Background:

• Accurate drug response prediction is crucial for personalized cancer medicine.
• Integrating multi-omics data with drug structural features can improve prediction accuracy.

Purpose of the Study:

• To develop a novel deep learning framework (MoGraphDRP) for predicting drug response in cancer cells.
• To integrate multi-omics cellular data with drug molecular graphs and chemical fingerprints.

Main Methods:

• A multi-branch deep learning framework integrating gene expression, mutation, methylation, and pathways.
• Graph Convolutional Networks (GCN) for drug molecular graphs and MLP for chemical fingerprints.
• Multi-head Bilinear Attention for integrating cellular and drug features, refined by an XGBoost ensemble model.

Main Results:

• MoGraphDRP achieved superior accuracy in drug response prediction (PCC=0.9689, RMSE=0.6622, R²=0.9388).
• The model outperformed state-of-the-art methods like BANDRP, DeepCDR, and DeepTTA.
• MoGraphDRP accurately reconstructs missing IC50 values and distinguishes drug sensitivity/resistance.

Conclusions:

• The MoGraphDRP framework offers a powerful and reliable tool for analyzing drug response.
• This approach advances precision medicine by enabling more effective personalized cancer therapies.
• The model provides interpretable insights for preclinical drug design and treatment strategies.